150 100 50 0
Z
50 100
Figure 720. Continued.
45
40
35
30
15\
case 1
casel18
case 19
0.02 0.04 0.06 0.08
x (m)
0.1 0.12 0.14 0.16
Figure 721. Predicted heat flux profiles for baseline case (case 1), best tradeoff from Pattern 1
(case 19), and best tradeoff from Pattern 2 (case 18).
217
tolerance to blade incidence changes, and have lower stresses than axial designs. Radial turbines
have been used successfully in automotive applications, but are not often used in rocket engines
due to their relatively large size and weight. The size of the compact radial turbines makes them
applicable to rocket engine cycles when a high velocity ratio is involved.
This research will focus on improving turbine efficiency. The radial turbine design must
provide maximum efficiency while keeping the overall weight of the turbine low. This
necessitates a multiobjective optimization. A response surface analysis46 prOVides an efficient
means of tackling the optimization problem.
The research presented here represents the preliminary optimization of a radial turbine
using a simplified 1D radial turbine model adapted from the 1D Meanlinel27 COde utilized by
Papila et al.82 The Meanline code provides performance and geometry predictions based on
selected input conditions. It is an approximate and inexpensive model of the actual processes.
Because it is an approximate model, there is some degree of uncertainty involved, but it can
provide a good starting point in the design process.
Using response surface analysis, an accurate surrogate model was constructed to predict
the radial turbine weight and the efficiency across the selected design space. The surrogate
model was combined with a genetic algorithmbased Pareto front construction and facilitates
global sensitivity evaluations. Because the radial turbine represented a new design, the feasible
design space was initially unknown. Techniques including design constraint boundary
identification and design space reduction were necessary to obtain an accurate response surface
approximation (RSA). The analysis used the optimization framework outlined in Chapter 2. The
framework steps included in the analysis are 1) modeling of the obj ectives using surrogate
models, 2) refining the design space, 3) reducing the problem dimensionality, and 4) handling
where is the estimated mean, r identifies the correlation vector between the set of prediction
points and the points used to construct the model, y a vector of the response at the sampled
points, and 1 denotes an Nsvector of ones. The mean is estimated by
p= 1 R'1) 1 I'R y (210)
and the standard deviation of the response is estimated as
<^r =(lLT 'yp (211)
On the other hand, the estimation variance at nonsampled design points is given by
11'R'r) (2
V(y(x))=o cr rTR 'r+ (2
1 R 11
In this study, the model estimates, estimation variance, and standard deviation of the response
are calculated using the Matlab toolbox DACE.101
2.1.2.3 Radial basis neural networks
Neural networks are another alternative to traditional response surface methods.
Specifically, neural networks are able to fit data with a highly complex and nonlinear response.
Artificial neural networks are made of interconnected local models called neurons. A high
number of neurons can improve the accuracy of the fit, while using fewer neurons improves the
smoothing qualities of the model. Mahajan et al.102 USed neural networks (NN) to optimize a
mechanically aspirated radiation shield for a meteorological temperature sensor to gain
information on the workings of the temperature sensor system and the performance of the
radiation shield. Papila et al.78 used radial basis neural networks to supplement the data used to
construct an RSA in the optimization of a supersonic turbine for rocket propulsion.
Charalambous et al.103 USed neural networks in the prediction of bankruptcy and stressed that the
different designs for testing is termed Design of Experiments (DOE). Designs may be tested
mainly at the design parameter extremes as in centralcomposite designs (CCD) and face
centered cubic designs (FCCD), or they may be tested across the full range of design parameter
values as in multilevel factorial designs46 Or Latin hypercube sampling.60 The design space is
defined by the physical range of parameters, or variables, to be explored. Variable ranges are
often determined based on experience with an existing design, but in many cases, the choice of
variable ranges is an educated guess. The simplest choice is to set simple range limits on the
design variables before proceeding with the optimization, but this is not always possible.
Simpson et al.61 USed a hypercube design space with design variables range selection based on
geometry to prevent infeasible designs in the SBO of an aerospike nozzle. Rodriguez 62
successfully optimized a complex jet configuration using CFD with simple geometrybased
bounds on the design variables. Vaidyanathan et al.63 chose variable ranges based on variations
of an existing design in the CFDbased optimization of a single element rocket inj ector.
Commonly, the design space cannot be represented by a simple boxlike domain. Constraints or
regions of infeasibility may dictate an irregular design. Designs that work best for irregular
design spaces include Latin Hypercube Sampling60 and orthogonal arrays.46 Once the design
space is bounded, the optimization commences. In order to reduce computational expense, it is
very important for a surrogate to be an accurate replacement for an expensive CFD simulation.
The application of the SBO approach in the context of CFDbased optimization with
complex flow models can present a significant challenge, in that the data required to construct
the surrogates is severely limited due to time and computational constraints of the high fidelity
CFD runs. In some cases, the problem of high computational cost in SBO can be addressed by
performing an optimization on a low fidelity model and translating the result to a high fidelity
0002 004 006 008 01 012 014 016 0 002 004 006 008 01 012 014 016
A xem) B (m)
Figure 622. Heat flux and y+ profiles along combustion chamber wall. A) Heat flux profile
along the wall at := 0 showing sharp increase in heat flux corresponding to the
location of velocity vx = 0. B) The y+ values along the near wall cell boundary dip at
stagnation regions.
10'
I I
E50 1001 L. ED 10 150
AQ x (mm) B x (me
Figure~ 63 eprtr n dycnutvt rflsa aiu oain npae=0
Figur 6A) Temperature. B) Eddy conductivity.poie tvrosyloain npaez=0
185
Unsteady approximation
p vx = 0
Ifreattachment point
vx = 0
Absolute percent area difference (compared to p = 2)
10
1 15 2 25 335 4 4.5 5
Figure 210. Absolute percent difference in the area under the Pareto front curves for the
original feasible design space (RS 2) and the refined design space (RS 3) for various
values of p as compared to a traditional least squares loss function (p = 2).
Optimization (EGO) algorithm developed by Jones et al.ll and has been shown to be effective in
searching globally for the optimum. It calculates the amount of improvement that can be
expected at a given location.
ME x)=(L )(x) "f)x (x) x) (x9)x (6
where
the expected value of the minimum function valuel22 (MF4). The expected value is simply the
predicted value at a given location minus the expected improvement. The criterion M~F4 is
similar to M~F3, but the search tends to be more local. Sasena et al.122 Suggests that M~F4 Should
be used only when there is confidence that the optimum region has been found.
MF4 X = V(x) V (x) (237)
The procedure is developed such that it accounts for the presence of multiple obj ectives.
In this way, a single point can be selected to improve the response surface across multiple
obj ectives. The procedure involves applying a DOE with a large number of points across the
design space. The points are then divided into clusters such that the number of clusters equals
the number of desired new points. Clustering the data prior to selection ensures that the data are
spread sufficiently across the design space, reducing the possibility of choosing two points that
are very near each other. For the purpose of multiobj ective optimization, the characteristics for
each obj ective need to be combined before the best points can be selected. The points are
selected based on a simple weighting function. In this study, the data is clustered based on
proximity using the Matlab function kmeans. For a given cluster m, a rank R is given to each
point n
Table 72. Kriging PRESSs error statistics for each design space iteration.
# of o, o,
# Data set pts qmea Lea
1 Original 16 6.93 5.33
2 Pattem 1 12 7.34 4.67
3 Pattern 1 15 4.28 5.90
4 Pattem 2 8 9.89 4.29
5 Pattern 2 12 4.26 6.06
6 Final 27 10.13 5.15
range(qmax) = max(qmax) min(qmax)
range(LC) = max(LC) min(Lc)
a a= do = square root of the process
PRESSms, PRESSms, PRESSms/
4nm Le range(qmx)
3.32 5.63 12.14
5.18 5.08 17.77
6.59 4.62 14.97
1.93 2.24 21.56
4.59 3.78 33.34
4.96 5.01 15.87
variance given in Equation 211
PRESSms/
range(LC)
25.39
34.67
12.26
31.54
20.72
22.59
PRESSms/
std(qmx)
0.41
0.67
0.40
0.65
1.03
0.51
PRESSms/
std(Lc)
1.02
1.04
0.35
1.05
0.66
0.97
~ "
noslip
H2 inlet
'02 pOst tip
slip
Figure 71. Inj ector element subsection. A) Outer row representation of baseline inj ector with
computational subsection shown at top and B) in closeup. C) Computational domain
and boundary conditions on inj ector element face for injector subsection. One
element in the outer element row is simulated. The circumferential spacing w is a
function of the number of inj ector elements in the outer row.
205
functions as selection criteria appeared to result in better accuracy than simply selecting the point
with the minimum function value or centers from each cluster. Using M~F; gave results that were
very similar to using foun. When the maximum error was considered, however, all merit
functions except M~F2 TOSulted in higher maximum errors as the number of points decreased.
Selecting points based on maximizing the probability of improvement resulted in the best
performance given a small number of data points as compared to simply choosing the minimum
function value or cluster centers. The results in the use of the remaining merit functions were
mixed. Maximizing the probability of improvement, maximizing the expected improvement, or
minimizing the expected value of the minimum function value resulted in lower average errors
than selection based on minimizing the statistical lower bound, minimum function value, or
cluster centers. However, simply selecting the cluster centers seemed to limit the maximum
error better than any merit function except maximizing the probability of improvement. In this
analysis, it was possible to reduce the number of points in the final optimization cycle by 94%
while keeping the accuracy of points along the Pareto optimal front within 10% with an average
error of only 3% or less.
3.5 Conclusion
An optimization framework can be used to facilitate the optimization of a wide variety
design problems. The liquidrocket compact radial turbine analysis demonstrated the
applicability of the framework:
* Surrogate Modeling. The radial turbine optimization process began without a clear idea
of the location of the feasible design region. RSAs of output constraints were successfully
used to identify the feasible design space.
* Design Space Refinement. The feasible design space was still too large to accommodate
the construction of an accurate RSA for the prediction of turbine efficiency. A reasonable
design space was defined by eliminating poorly performing areas thus improving RSA
fidelity.
M~.I. =t pt pV d (54)
where pu is the fluid viscosity, Uis the timeaveraged velocity, and u' and v' are the fluctuation
velocities (u' = u U, v' = v v). Equation 54 is related to the full Reynolds momentum
conservation equation. The mixing is thus described as the combination of the momentum
transfer due to viscosity and the momentum transfer by the fluctuation velocity field. The
magnitude of the product of the fluctuation velocities is used to keep the sign of this term
consistent throughout the data points. Only the magnitude of the mixing index is important, so
the absolute value of the timeaveraged quantity represents the mixing index value. The
constraints are given as
B+b<1
0.5
(55)
0.0 < b < 0.5
0.0
The first constraint in Equation 55 is incorporated to keep the frontal height equal to unity so
that the Reynolds number remains constant. The remaining constraints maintain convexity of the
geometry. The constraints are identical to those used in Burman et al.m7
A total of 52 data points were used in the analysis. The design of experiments (DOE)
procedure was used to select the location of the data points that minimize the effect of noise on
the fitted polynomial in a response surface analysis. A modified facecentered composite design
(FCCD) is used to select 27 of the 52 data points. The facecentered composite design usually
uses the corners and center faces of a cube for point selection, but in this case, half of the cube
would violate the first constraint in Equation 55. This means that four points from the FCCD
are infeasible and must be removed from the design. Additional points were added on the
57 Comparison of response surface (top row) and radial basis neural network (bottom
row) prediction contours for total pressure loss coefficient............... ...............5
58 Comparison of response surface (top row) and radial basis neural network (bottom
row) prediction contours for mixing index (including extreme cases) ................... .........154
59 Comparison of response surface (top row) and radial basis neural network (bottom
row) prediction contours for mixing index (excluding extreme cases) ...........................155
510 Variation in obj ective variables with grid refinement ........._._. ......___ ............... 155
511 Difference in predicted mixing index values from response surface (top row) and
radial basis neural network (bottom row) prediction contours constructed with and
without extreme cases ................. ...............156...... ......
512 Comparison of response surface and radial basis neural network prediction contours
for mixing index at B* = 0 and h* = 0 ..........._ ....._.._ ...............15
61 Blanching and cracking of combustion chamber wall due to local heating near
inj ector el ements. ........... ..... ._ ............... 175...
62 Hydrogen flow geometry. ............. ...............176....
63 Hydrogen inlet mesh. .............. ...............176....
64 Zvorticity contours............... ...............17
65 Swirl number at each x location ................. ...............177........... ..
66 Average axial velocity u and average tangential velocity vs with increasing x...............178
67 Reynol ds numb er profile es ........._.._ ...... .___ ............... 178..
68 Nondimensional pressure as a function of x ................. ...............178.............
69 Hydrogen inlet flow profiles ................. ...............179........... ...
610 Combustion chamber crosssectional geometry and thermocouple locations. ................179
611 Estimated wall heat flux using linear steadystate and unsteady approximations. ..........180
612 1D axisymmetric assumption for heat conduction through combustion chamber wall .180
613 Estimated wall temperatures using linear and axisymmetric approximations .................18 1
614 Temperature (K) contours for 2D unsteady heat conduction calculations ................... ..181
615 Experimental heat flux values using unsteady assumptions ................. .....................182
45. Kim, S., Leoviriyakit, K., and Jameson, A., "Aerodynamic Shape and Planform
Optimization of Wings Using a Viscous Reduced Adj oint Gradient Formula," Second
M.I.T. Conference on Computational Fluid and Solid Mechanics at M.I.T, Cambridge,
MA, 2003.
46. Myers, R. H., and Montgomery, D. C., Response Surface M~ethodology, New York: John
Wiley & Sons, Inc., 2002, pp. 308, 393399, 1720.
47. Krige, D., "A Statistical Approach to Some Basic Mine Evaluation Problems on the
Witwatersrand," Journal of the Chemical, Metallurgical and Mining Society of Sonlu
Afiica, Vol. 52, 1951, pp. 119139.
48. Oliver M. and Webster, R., "Kriging: A Method of Interpolation for Geographical
Information System," hIternational Journal of Geographical hIfornzation Systems, Vol. 4,
No. 3, 1990, pp. 313332.
49. Journel, A. and Huijbregts, Ch., M~ining Geostatistics, New York, Academic Press, 1978.
50. Mitchell, T., Machine Learning, McGrawHill, 1997, pp. 81127.
51. Burgee, S., Giunta, A. A., Narducci, R., Watson, L. T., Grossman, B. and Haftka, R. T., "A
Coarse Grained Parallel VariableComplexity Multidisciplinary Optimization Paradigm,"
The hIternational Journal of Supercomputer Applications and H~igh Performance
Computing, Vol. 10, No. 4, 1996, pp. 269299.
52. Queipo N., Haftka R. T., Shyy W., Goel T., and Vaidyanathan R. "SurrogateBased
Analysis and Optimization," Progress in Aerospace Sciences, Vol. 41, 2005, pp. 128.
53. Kontoravdi, C., Asprey, S. P., Pistikopoulos, E. N., and Mantalaris, A. "Application of
Global Sensitivity Analysis to Determine Goals for Design of Experiments An Example
Study on AntibodyProducing Cell Cultures," Biotechnology Progress, Vol. 21, No. 4,
2005, pp. 11281135.
54. Marsden, A., Wang, M., Dennis, J., Moin, P., "Optimal Aeroacoustic Shape Design Using
the Surrogate Management Framework," Optimization and Engineering, Vol. 5, 2004, pp.
235262.
55. Jansson, T., Nilsson, L., Redhe, M., "Using Surrogate Models and Response Surfaces in
Structural Optimization With Application to Crashworthiness Design and Sheet Metal
Forming," Structural M\~ultidisciplinazry Optimization, Vol. 25, 2003, pp. 129140.
56. Joyner, C. and Sabatella, J., "Launch Vehicle Propulsion Optimization Using Response
Surface Methodology," AIAA Paper 19902433, 1990.
57. Shyy, W., Tucker P. K., and Vaidyanathan, R., "Response Surface and Neural Network
Techniques for Rocket Engine Inj ector Optimization," Journal of Propulsion and'Power,
Vol. 18, 2001, pp. 391401.
228
CHAPTER 5
SURROGATE MODELING OF MIXING DYNAMIC S
A bluff bodyinduced flow is used as a model problem to help probe the physics and
surrogate modeling issues related to the mixing dynamics. Understanding of the mixing
characteristics in the wake of the bluff body has important implications to reacting flow, such as
in the shear reacting layer of inj ector flows. The sensitivity of the mixing dynamics to the
trailing edge geometry of the bluff body can be explored using CFDbased surrogate modeling.
Plausible alternative surrogate models can lead to different results in surrogatebased
optimization. The current study demonstrates the ability of using multiple surrogate models to
discover the inadequacy of the CFD model. Since the cost of constructing surrogates is small
compared to the cost of the simulations, using multiple surrogates may offer advantages
compared to the use of a single surrogate. This idea is explored for a complex design space
encountered in a trapezoidal bluff body. Via exploration of local regions within the design
space, it is shown that the design space has small islands where mixing is very effective
compared to the rest of the design space. Both polynomial response surfaces and radial basis
neural networks are used as surrogates, as it is difficult to use a single surrogate model to capture
such local but critical features. The former are more accurate away from the highmixing
regions, while the latter are more accurate near these regions. Thus, surrogate models can
provide benefits in addition to simple model approximation.
5.1 Introduction
The case study focuses on the mixing and total pressure loss characteristics of time
dependent flows over a 2D bluff body. Bluff body devices are often used as flameholding
devices such as in afterburner and ramjet systems. Bluff body devices should have good mixing
capability and low pressure loss across them. Challenges exist in that the bluff body flow is
c, RS (h = 0) c, RS (h = 0 5) c, Rs (h = 1)
2 2 22
O1 0511 1 0' 05 01 0
19 ':2 11 22
O 05 1 0 05 0 OS 1
B B 8
MI RS (h = 0) 750 MI (h = 0 5) re 1 M R (h = 1) 5
70 0L 8 2 1 _O
D 0 5 1 S 0 0 5 1 6 0 0 5 1 0
Mt / N (h = 0) 0Mt / N (h = 0 5) g 1 Mt / N (h = 1)70
D 0 5 1 SO OO O5 1 60 0 0 5 1 0
B B B
row prdcincnorsfrmxn ne (inldn exrm caes at h'(A= 0 (left),(A= 1
0. (cntr) an 1 rgt. omlzd aibe aeson
model. Doing so requires additional analyses to ensure that the lowfidelity model is a
qualitatively similar to the high fidelity model. In other instances, lowfidelity data may be
combined with highfidelity data to reduce the overall number of expensive highfidelity runs.
The lowfidelity data may employ a correction surrogate for improved accuracy, as shown in
Figure 14.
Han et al.64 perfOrmed a surrogatebased optimization on a lowfidelity 2D CFD model
and tested the results on a more expensive 3D model in the optimization of a multiblade fan
and scroll system and verified the 2D model using physical experiments as well as comparisons
to previous studies and the 3D model. Keane65 built an accurate surrogate model of a transonic
wing using highfidelity CFD data supplemented with data from a low fidelity empirical model.
Alexandrov et al.66 USed variation in mesh refinements as variable fidelity models for a wing
design. Knill et al.67 Saved 255 CPU hours by supplementing highfidelity data with a low
fidelity aerodynamic model in the optimization of a high speed civil transport wing. Vitali et
al.68 USed surrogates to correct a lowfidelity model to better approximate a high fidelity model in
the prediction of crack propagation. Thus, the accuracy of the lowfidelity models can be
improved by combining the lowfidelity model with the surrogate correction, and a new
surrogate could be constructed based on the composite model. Venkataraman et al.69 also used a
surrogate correction on a lowfidelity model to aid in the optimization of shell structures for
buckling. Balabanov et al.70 applied a surrogate correction to a lowfidelity model of a high
speed civil transport wing. Haftkan1 and Chang et al.72 Scaled surrogate models with local
sensitivity information to improve model accuracy for analytical and structural optimization
problems.
linear
axisymmetric
500
450
0 0.02 0.04 0.06 0.08
0. 1 0. 12 0. 14 0. 16
Figure 613. Estimated wall temperatures using linear and axisymmetric approximations. Linear
and axisymmetric assumptions give nearly equal wall temperature estimations.
Temperature values at x = Om and x = 0.16m are estimated based on nearest
temperature values.
0 02E
002(
0 015
0 01
0 005
0 005
0 01
0 0 005 0 01 0 015 0 02 0 025 0 03
0 0 005 0 01 0 015 0 02 0 025 0 03
0 02
550
0 ... I i
C 0 0 005 0 01 0 015 0 02 0 025 0 03 D 0 0 005 0 01 0 015 0 02 0 025 0 03
Figure 614. Temperature (K) contours for 2D unsteady heat conduction calculations at x =
0.084 m at A) 1 s, B) 2 s, C) 6.5 s, and D) 7.75 s. Units are given in meters. One
quarter of the combustion chamber cross section is shown. The temperature scale is
consistent across plots.
The design variable bounds were further reduced to match the new design space. The new
design variable ranges are given in Table 33. Quadratic response surfaces were constructed for
the turbine weight, Wrotor, and the turbine totaltostatic efficiency, rls, using the original feasible
design space to screen data points. Points predicted to lie outside of the refined design space
would be omitted. For the refined design space, a third set of data was required. As seen in
Figure 37, using a factorial design tended to leave holes in function space. It was possible that
this could hamper construction of an accurate Pareto front. To prevent this, Latin Hypercube
Sampling was used to help close possible holes for the third data set. The points were efficiently
distributed by maximizing the minimum distance between any points. Of the best points from the
second data set, it only RPM, Tip Flw, and U/C isen varied, while React and Dhex% remained
constant at their lowest values, and AnsqrFrac remained constant at its maximum value among
these points. To ensure that this effect was captured in the third data set, additional points were
added using a 5level factorial design over these three variables. The remaining variables were
held constant according the values observed in the best tradeoff points. Quadratic response
surfaces previously constructed for the turbine weight, Wrotor, and the turbine total to static
efficiency, rls, were used to screen the potential data points. Points predicted to lie outside of the
newly refined design space would be omitted from the analysis. In summary,
1. Only the portion of design space with best performance was reserved to allow for a
concentrated effort on the region of interest and to increase response surface fidelity.
2. Latin Hypercube Sampling (181 / 300 feasible points) was used over all six variables and
was supplemented by a 5level factorial design used over RPM, Tip Flw, and U/C isen
(1 19 / 125 feasible points) to improve resolution among the best tradeoff designs.
3. Points that were predicted to violate constraints or lie outside of region of interest were
omitted.
The combination of the DOEs resulted in a total of 323 feasible design points.
kinetics model with the kE turbulence model and compared the velocity profile to experimental
measurements with good results. Tucker et al.5 stressed the importance of supplying
experimental results to aid in the task of CFD validation.
Shyy et al.29 and Vaidyanathan et al.30 identified several issues inherent in CFD modeling
of inj ectors, including the need for rigorous validation of CFD models and the difficulty of
simulating multielement injector flows due to lengthy computational times. They chose to
simulate single inj ector flow, as this allowed for the analysis of key combustion chamber life and
performance indicators. The life indicators were the maximum temperatures on the oxidizer post
tip, inj ector face, and combustion chamber wall, and the performance indicator was the length of
the combustion zone. These conditions were explored by varying the impinging angle of the fuel
into the oxidizer. The studies were successful in using CFD to specifically access the effects of
small changes in inj ector geometry on combustion performance.
1.2.2 Rocket Engine Turbine Design
Turbines are developed based on basic knowledge of the fuel characteristics, estimated
stresses, and turbine pressure. The turbines are driven using hot exhaust gas from the
combustion process. Turbines were and still are designed by using empirical relations to
calculate a preliminary starting point for the design. The designs rely both on past information
from similar turbine designs and on experimental correlations. Unlike for injector design,
computers have been used in the design process for rocket engine turbines for over 40 years.
Beer 1965 used the aid of computers to estimate the velocity distribution across a turbine blade
to help improve turbine design. Simple 1D codes are routinely used to provide preliminary
information on design temperatures, stresses, overall flow geometry, velocities, and pressures.31
The prediction of turbine performance was later improved by the consideration of aerodynamic
effects that lead to pressure losses.32
3.3.3 Phase 3: Construction of the Pareto Front and Validation of Response Surfaces
As in the second data set, five RSAs each with p = 1...5 were used to fit the data for oro;o
and rl, for the third data set. Pareto fronts were constructed for each RSA set and are shown in
Figure 310. In this case, the Pareto fronts differed by a maximum of only 5%. Because the
difference in the response surfaces for varying values of p is small for the third data set, the
design space was determined to be adequately refined.
Function evaluations from the quadratic response surfaces (p = 2) were used to construct
the Pareto Front shown in Figure 311. Within the Pareto front, a region was identified that
would provide the best value in terms of maximizing efficiency and minimizing weight. This
tradeoff region was selected for the validation of the Pareto front. The results of the subsequent
validation simulations indicated that the response surfaces and corresponding Pareto front were
very accurate. A notable improvement was attained compared to the baseline radial turbine
design. The design selected optimum design had the same weight (Woro; as the baseline case
with approximately 5% improvement in efficiency. The specifications for the optimum design
are given in Table 34.
Within the best tradeoff region, only RPM~ and Tip Flw vary along the Pareto front as seen
in Figure 312. The other variables are constant within the tradeoff region and are set to their
maximum or minimum value. This indicates that increasing the range of one of these variables
might result in an increase in performance. The minimum value of the variable React was chosen
as the only variable range that could reasonably be adjusted. The validation points were
simulated again using a reduced React value. Reducing the minimum value of React from 0.45 to
0.40 increased the maximum efficiency only for oro; > 1. The maximum increase in efficiency
improved from 4.7% to 6.5%, but this increase occurred outside of the preferred tradeoff region.
MOVy
al '
OXlduz
oxidizer pump
oxiize
fuel pump
Figure 11. Rocket engine cycles.
cycle.
A) Expander cycle, B) gas generator cycle, C) preburner
I POTP
Figure 12. SSME thrust chamber component diagram showing lowpressure fuel turbopump
(LPFTP), lowpressure oxidizer turbopump (LPOTP), highpressure fuel turbopump
(HPFTP), highpressure oxidizer turbopump (HPOTP), main fuel valve (MFV), main
oxidizer valve (MOV), combustion chamber valve (CCV), and the main combustion
chamber (MCC). Picture reproduced from [81].
PFTP
HPOTP
HPFTP
GCV
fields are categorized based on the location and size of mean recirculation regions near the bluff
body. These groups will provide a means of reference for describing general trends in the
sample data. Geometries in Group F had the lowest total pressure loss coefficient with an
average value of 1.85. Group B had the highest mixing index with an average value of 655. It
can be predicted that the best designs in terms of high mixing and low pressure drop, and thus
the focus of a possible future design refinement, may lie in the region including and between
Group B and Group F.
5.3.2 Surrogate Model Results
When constructing the response surface models of the two obj ectives, a cubic response
surface model was found to perform better than a simpler quadratic model. The full cubic model
is given by
j) = p, + PB* + PAb* + PAh* + pB*2 + P5B'b* + PAb*2 + 7B h*
+ PAh'b* + PAh*2 10pB*3 1 ~1B*2b* + P42B'b*2 1 P3b*3 1 P4B*2h* (5 6)
+ P,4B'b'h* + P, b*2h* + P,6B'h*2 17Pb'h*2 18Pb'h*2 19Ph*3
where a starred value indicates the variable is normalized between 0 and 1. Insignificant terms
were then removed from some cubic response surfaces, resulting in reduced cubic models. The
models used in this analysis are given in Table 53.
Problems were discovered in the mixing index response surface model fit:
* The response surface has a low R2ad, Value and a high RMS error, so the fit is very poor.
* Roughly half of the coefficients in the response surface are very large (> 1000).
* Three of the large coefficients are positive and four are negative.
The last two symptoms indicate a polynomial with sharp gradients and oscillatory behavior. It
was suspected that these symptoms were due to behavior exhibited by the extreme cases. The
extreme cases were removed in an attempt to improve the response surface fit for the mixing
index. As seen from Table 53, after removal of the extreme cases the response surface fit
Fk~max Fk~~min (238)
where F is the value of the selection criterion (merit function, function value, etc.), F"'min is the
point with the smallest value ofF within the cluster nz, n is a point from the cluster nz, and Fk,max
and Fk,min are the maximum and minimum values F, respectively, within the entire data set for
obj ective k. For M~F; and M~F4, the point in each cluster with the minimum value R is selected.
For MFz and M~F3, the point in each cluster with the maximum value R is selected. The full
procedure for using merit functions to select points is as follows:
1. Construct a response surface using Kriging for each obj ective using an initial sampling.
2. Determine if the design space size should be reduced. Construct a DOE using a large
number of points M~within the new design space.
3. Choose a merit function and calculate the merit function value for each obj ective k at each
point M~in the new DOE using the predicted function value and variance.
4. Calculate R at each point M~ as determined by Equation 23 8.
5. Choose the desired number of points na to be sampled from the new DOE. The number of
points na is selected by the user according to the number of sample points that can be
afforded by the simulation or experiment. The quantity M~ should be much larger than nt.
6. Group the points M~into na number of clusters.
7. Select the point that has the best value of R from each cluster.
The final set of na design points will satisfy three basic criteria:
1. The new points will be evenly spread across the new design space
2. The selected points will be selected such that they are predicted to improve the function
values in each objective
3. The selected points are those in each cluster predicted to best reduce the errors in the
surrogate models for each objective.
Figure 21. Optimization framework flowchart.
69
the available energy for a given fluid is accomplished by increasing the turbine inlet temperature.
To increase the turbine inlet temperature in an expander cycle, a higher heat flux from the thrust
chamber to the cooling fuel is needed. Obtaining this higher heat flux is problematic in several
ways. First, materials and manufacturing development is necessary to produce a thrust chamber
with high heat flux capability. This work is an ongoing area of technology development.
Second, to enable a higher heat exchange, increased surface area and contact time between the
thrust chamber and cooling fuel is needed. These requirements lead to a larger, and heavier,
thrust chamber. In addition, significantly raising the turbine inlet temperature defeats the
expander cycle' s advantage of maintaining a benign turbine environment. Immich et al.126
reviewed methods of enhancing heat transfer to the combustion chamber wall in an expander
cycle. The authors tried three methods of enhancing heat flux to the unburned fuel: 1) Increase
the length of the combustion chamber, 2) increase the combustion chamber wall surface area by
adding ribs, and 3) increase the combustion chamber wall roughness. The authors also
mentioned that in the future they would experimentally investigate the influence of the
distribution of the inj ector element distribution and the effect of the distance of the inj ector
element to the wall on the heat transfer to the wall. However, results of the future analysis are
not available.
The second approach to increasing turbine work is to improve turbine efficiency. If the
turbine inlet temperature is held constant, an increase in turbine work is directly proportional to
efficiency increase. If the required work can be achieved with moderate efficiency, an
improvement in that efficiency can be traded for reduced inlet temperatures, providing better
design environment margins. One way to improve turbine efficiency is to use a radial turbine.
Radial inflow turbines perform better than axial turbines at high velocity ratios, exhibit better
1.1.1 Rocket Engine Cycles
Rocket engines carry both the fuel and the propellant with the vehicle. For chemical
rockets, the propellant is a large part of the total mass of the vehicle. The basic goal in rocket
engine design is to obtain the highest thrust possible with the lowest total vehicle weight.6 As a
general obj ective, it is desirable to have a specific impulse that is as large as possible. Hydrogen
and oxygen are often chosen as the fuel and propellant because they have very high values of
specific impulse. While the goal of maximizing the specific impulse is straightforward, many
rocket engines designed to help achieve the goal are inherently complex and expensive. The
need for better performance drives the complexity.
Pressurefed engines are relatively simple;' two high pressure tanks are connected directly
to the combustion chamber, and valves are used to regulate the flow rate. Pressurefed engines
have low costs and good reliability because they are simple with few components. However,
pressurefed engines require heavy, massive tanks to provide the required pressure. For this
reason, turbopumps were used in other rocket engine cycles to deliver the needed pressure,
allowing for lighter fuel and propellant tanks. Additionally, gas generators or preburners were
required to supply hot gases to the turbines. This led to a need for additional components, where
each component could compromise engine reliability.
In a gasgenerator cycle (Figure 11B) a portion of the fuel is burned before it reaches the
combustion chamber. The exhaust from the gas generator is used to drive a turbine, which is
used to drive the pumps. After the exhaust leaves the turbine, it is dumped.'0 Gas generator
engine cycles provide improved thrust and specific impulse as compared to a pressurefed engine
cycle.ll
The expander cycle (Figure 11A) eliminates the need for a gas generator, thus resulting in
a simpler and more reliable system. The expander cycle is the simplest of the pumpfed engine
 ~
//
///// =
C
Figure 55. Continued.
F'''d' '
*~1:~
rS I 'r
~: '
i~J r
~i''''
~;
/I ,,
Dc.zpl
ul~i
i I
1 ;.
1 ui.s~e
r
.s
I""`
,BT~j
L~~1L
nr I. esf
...~,
~~~'''''~r.. .'
~
1:11
,
j
,' C
c,im
: H1UB
~ IO)..
''
Ofi O~
Figure 56. Flow characterization in design space on the Bb axis. At b = 0, the angle of the
lower bluffbody surface is zero. At B = 0, small angles exist on lower and upper
bluff body surfaces and the sum of the magnitude of these angles is at a maximum.
Along the line B b = 1, the angle of the upper surface is equal to zero.
153
balance in terms of solution run time and solution convergence in terms of grid resolution as
detailed in Goel et al.m7 Multiple grid refinements were investigated to identify the adequate
resolution for this problem. The number of grid points in each grid is given in Table 51. While
the difference in the total number of grid points between Grid 1 and Grid 2 is not large, the grid
density in the near field and wake region for Grid 2 is higher than Grid 1. In Grid 3 and Grid 4,
both the near field grid density and far field grid density are high, making them more refined.
Overall, Grid 1 has the poorest resolution and Grid 4 has the best resolution. Grid 2 was selected
for the analysis as it appeared to offer the best tradeoff in terms of grid resolution and
computational run time. The run time for a single case was approximately eight hours on a 16
CPU cluster with Intel Itanium processors (1.3 GHz) and 16GB of RAM.
5.2.2 Objective Functions and Design of Experiments
The two obj ectives are the total pressure loss coefficient, CD, and the mixing index, MI1..
The total pressure loss coefficient at any time instant is equal to the sum of the pressure and
shear forces on the body divided by the drag force. The total pressure loss coefficient is
averaged between time t and time t* + T. It is assumed that Tis large compared to the
oscillatory time period. The total pressure loss coefficient is given by
CDft x pU ~ xtzc) dS dt (53)
where p is the pressure, zx is the viscous stress tensor, p is the fluid density, and D is the frontal
height of the trapezoidal bluff body.
The measure of mixing efficacy is given by the laminar shear stress plus the unsteady
stress. The mixing index, M.I., is chosen as the integral of the mixing efficacy over the entire
computational domain averaged over a time range T. It is given by
Table 34. Baseline and optimum design comparison.
Objectives Description Baseline Optimum
Wrotor Relative measure of "goodness" for overall 1.147 1.147
weight
Grls Totaltostatic efficiency 85.0% 89.7%
Design Variables Baseline Optimum
Table 32. Response surface fit statistics before (feasible DS) and after (reasonable DS) design
space reduction.
Feasible DS
Reasonable DS
Wrotor grs
0.996 0.995
0.996 0.994
0.0235 0.00170
1.04 0.844
310 310
Wrotor
0.987
0.985
0.094
1.04
224
0.917
0.905
0.020
0.771
224
R2ad,
Root Mean Square Error
Mean of Response
Observations
Table 33. Original and final design variable ranges after constraint application and design space
reduction.
Design
variable
RPM ~
Original ranges
MIN MAX
80,000 150,000
Final ranges
MIN MAX
100000 150,000
Description
Rotational Speed
Percentage of stage pressure drop across
rotor
React
0.45
0.68
0.63
0.65
0.4
0.85
0.45
0.57
U/C isen Isentropic velocity ratio
Tip Flw Ratio of flow parameter to a choked flow
parameter
Dhex% Exit hub diameter as a % of inlet
diameter
0.56 0.63
0.53
0.1 0.4
AnsqrFrac
Used to calculate annulus area (stress
indicator)
0.68
0.85
RPM
122,000
0.55
124,500
0.45
0.63
0.30
0.10
Rotational Speed
Percentage of stage pressure drop across rotor
Isentropic velocity ratio
Ratio of flow parameter to a choked flow
parameter
Exit hub diameter as a % of inlet diameter
Used to calculate annulus area (stress indicator)
React
U/C isen
Tip Flw
Dhex %
AnsqrFrac
0.61
0.25
0.58
0.83
0.85
(i) the dimensionality of the problem, (ii) whether noise is important source of error, (iii) the
number of simulations or experiments that can be afforded, (iv) the type of surrogate used to
model the problem, and (v) the shape of the design space. If noise is the dominant source of
error, DOEs that reduce sensitivity to noise are commonly used. These include central composite
designs (CCD) or facecentered cubic designs (FCCD) for boxshaped domains. Design
optimality designs such as D or Aoptimal83 (Myers and Montgomery 2002, pp. 393 395)
designs are useful for irregular shaped domains and high dimensional domains when minimizing
noise is important. Specifically, these designs can be used to reduce the number of points in an
experimental design for a given accuracy. When noise is not an issue, spacefi11ing designs such
as LatinHypercube Sampling (LHS)60 Or Orthogonal Arrays (OAs)46 are preferred to efficiently
cover the entire design space.
Central composite designs and facecentered cubic designs are intended to minimize the
presence of noise in the response. They are comprised of a single point at the center of the
design space with the remaining points situated along the periphery of the design space, as
shown in Figure 22. The center point detects curvature in the system while all other points are
pushed as far as possible from the center provide noisesmoothing characteristics. The FCCD is
actually a modified version of the CCD with the axial points moved onto the edges of a square.
An FCCD is required when the limits of the design space are strictly set, requiring a square
design space.
The CCD and FCCD are most popular for response surface approximations. They are
commonly used to reduce noise in the response by placing most of the points on the boundary of
the design space. For responses that do not contain noise, such as data obtained using
deterministic computational simulations, a spacefi11ing design may be required for accuracy.
ZH2 ,n, TH2
02,2n, O2,2 "
H2,u vi ou
H20,out, out
Figure 49. Fuel rich hydrogen and oxygen reaction with heat release.
136
6. 1 Introducti on ................. ...............157........... ...
6.2 Experimental Setup............... ...............158
6.3 Upstream Injector Flow Analysis ................ ...............159........... ...
6.3.1 Problem Description ................. ...............159........... ...
6.3.2 Results and Discussion ................. ...............160........... ...
6.3.3 Conclusion ................. ........... ...............162 .....
6.4 Experimental Results and Analysis .............. ...............163....
6.5 Injector Flow Modeling Investigation ................. ...............166..............
6.5.1 CFD Model Setup ........._..... .... ....___ ......_._... ... .........6
6.5.2 CFD Results and Experimental Comparison of Heat Flux ........._._... ............... 168
6.5.3 Heat Transfer Characterization............... ...........16
6.5.4 Species Concentrations............... ............17
6.6 Grid Sensitivity Study ........._.___..... ._ __ ...............171..
6.7 Conclusion ........._.___..... .__ ...............173....
7 MULTIELEMENT INJECTOR FLOW MODELING AND ELEMENT SPACING
EFFECT S............... ...............19
7. 1 Introducti on ................. ...............191........... ...
7.2 Problem SetUp .................... ...............193.
7.3 Feasible Design Space Study ................. ...............195........... ...
7.4 Design Space Refinement ................. ...............199...............
7.5 Conclusion ................ ...............203................
8 CONCLUSIONS .............. ...............219....
8. 1 Radial Turbine Efficiency and Weight Optimization ........................... ...............220
8.2 Bluff Body Mixing Dynamics ............ ............ ...............221.
8.3 SingleElement Inj ector Flow Modeling .....__ ................ ............... 222 ...
8.4 MultiElement Inj ector Flow Modeling. ......... ........____ ......... ..........22
8.5 Future Work............... ...............224.
REFERENCE LIST .............. ...............225....
BIOGRAPHICAL SKETCH .............. ...............238....
FarhangMehr and Azarml21 USed a different selection technique for constructing surrogate
models for crash analyses. The technique also required a standard DOE and Kriging fit to
initialize the response surface. The method then added points based on the predicted
irregularities in the response surface. This method allowed for the response surface to be
improved while requiring potentially fewer points overall. Thus, the design space was
adaptively refined using points that were selected based on the predicted obj ective function,
itself. FarhangMehr and Azarm cited the benefits of their approach in reducing the total number
of points needed for optimizations based on computationally expensive simulations. The
procedure is good for improving surrogate accuracy, but does not take into account the value of
the function, meaning that points may be added in regions of noninterest.
2.4.3 Merit Functions for Data Selection and Reduction
Using the concepts such as design space reduction (Section 2.4.1) along with the benefits
of innovative point selection as in the previous section, merit functions, facilitated by Kriging,
can be smoothly integrated into the multiobjective DSR process. Merit functions are statistical
measures of merit that use information of the function values and model uncertainty in the
surrogate model to indicate the locations where the function values can be improved and
uncertainty reduced. In this research effort, merit functions are used to select data points such
that the accuracy of the surrogate model can be improved using a minimal number of data points.
The merit functions rely on predictions of the function values and function uncertainty and
attempt to balance the effects of the two. Kriging is selected as the surrogate model for the
optimization due to its inherent ability to provide estimates of both the function value and the
prediction variance. Although several merit functions are available, the strengths of a given
merit function over another for the purpose of multiobj ective optimization with an extremely
limited number of function evaluations has not been determined.
6.5
D= 1
Figure 53. Computational domain for trapezoidal bluff body.
>0
0 10 20 30
Figure 54. Computational grid for trapezoidal bluff body.
Figure 55. Bluff body streamlines and vorticity contours. A) Typical instantaneous vorticity
contours, B) typical instantaneous streamlines, and C) timeaveraged streamlines for
flow past a trapezoidal bluff body (B = 0.5, b = 0.25, h = 0.25) at Re = 250.
616 Computational model for singleelement inj ector flow simulation .............. ..............182
617 Velocity contours vx(m/s) and streamlines............... ..............18
618 Temperature (K) contours............... ...............18
619 CFD heat flux values as compared to experimental heat flux approximations. ..............183
620 Wall heat transfer and eddy conductivity contour plots. ............. .....................8
621 Streamlines and temperature contours at plane := 0............... ...............184...
622 Heat flux and y+ profiles along combustion chamber wall. ........._.._.._ ....._.._.. ......185
623 Temperature and eddy conductivity profiles at various y locations on plane := 0.........185
624 Mass fraction contours for select species............... ...............186
625 Mole fractions for all species along combustion chamber centerline (y = 0, z = 0) ........187
626 Select species mole fraction profiles............... ...............18
627 Sample grid and boundary conditions............... ..............18
628 Computational grid along symmetric boundary. ............. ...............189....
629 Wall heat flux and y+ values for select grids ........... ......__ ... ..___.......18
630 Comparison of temperature (K) contours for grids with 23,907, 3 1,184, 72,239, and
103,628 points, top to bottom, respectively ................. ...............190........... ..
71 Injector element subsection............... ..............20
72 Design points selected for design space sensitivity study .............. ......................0
73 Effect of hydrogen mass flow rate on obj ectives ................. ...............206............
74 Oxygen isosurfaces and hydrogen contours. ............. ...............207....
75 Hydrogen contours and streamlines ................. ...............207........... ...
76 Maximum heat flux for a changing radial distance r* ................... ............... 20
77 Heat flux distribution. .............. ...............208....
78 Maximum heat flux as a function of aspect ratio ................. ...............209............
79 Design points in function and design space. .............. ...............209....
710 Merit function (M~Fz) contours for ........._ ....... ...............210.
optimizer to generate Pareto optimal solutions. The description of the algorithm is given as
follows:
1. Randomly initialize a population (designs in the design space) of size npop.
2. Compute objectives and constraints for each design.
3. Rank the population using nondomination criteria. Many individuals can have the same
rank with the best individuals given the designation of rank1. Initialize an archive with all
the nondominated solutions.
4. Compute the crowding distance. This distance finds the relative closeness of a solution to
other solutions in the function space and is used to differentiate between the solutions on
same rank.
5. Employ genetic operatorsselection, crossover, and mutationto create intermediate
population of size npop.
6. Evaluate objectives and constraints for this intermediate population.
7. Combine the two (parent and intermediate) populations, rank them, and compute the
crowding distance.
8. Update the archive:
9. Compare archive solutions with rank1 solutions in the combined population.
10. Remove all dominated solutions from the archive.
11. Add all rank1 solutions in the current population which are nondominated with respect to
the archive.
12. Select a new population npop from the best individuals based on the ranks and the
crowding distances.
13. Go to step 3 and repeat until the termination criteria is reached, which in the current study
is chosen to be the number of generations
2.4 Design Space Refinement Techniques
Given knowledge about a problem, it is desirable to perform CFD simulations with (a)
design parameters set in a region that is known to provide good results in order to refine the
design (exploitation) or (b) to set the design parameters in an unexplored region and use
designs. The resulting Kriging fits including the new data points are shown for Pattern 1 and 2
in Figure 714 and Figure 715, respectively. The response surface for peak heat flux of Pattern
1 in Figure 714 shows smooth contours, indicating a good surrogate fit. The prediction errors of
the Kriging fits echo what can be casually observed from the contour plots. For Pattern 1, the
addition of points slightly decreases the scaled error in qmax, and significantly decreases the
scaled error in the prediction for Le. The prediction of the combustion length of Pattern 2 was
improved by separating the design space into two regions as evidenced by the scaled prediction
error. The fit of the peak heat flux became worse for Pattern 2 after splitting the design space
into two regions, but this may be due to the sparseness of data points in the region.
When the data points were plotted as combustion length versus peak heat flux in Figure
716, it was discovered that all of the newly added points dominated the points in the original
Pareto front. Overall, the heat flux was reduced by up to 20% and the combustion length was
reduced by up to 5% among the eleven new points as compared to the best tradeoff points from
the original data set of 20 points. This indicates the success of using merit functions to select
points for design space refinement. This is a case where a surrogate based on a simple DOE can
be used to positively identify the "bad" regions so that they may be avoided. In this case, the
original surrogates were not accurate in regions away from the original design points,
particularly in the prediction of the combustion length, but they are sufficient to identify regions
that are obviously bad. By ignoring obviously bad regions of the design space, new data points
could be concentrated in the best tradeoff region. The original surrogate model was also
successful in identifying the regions of the design space that may result in function improvement,
as six of the eleven new points lay in the expanded portion of the design space.
NI 11 11
Figure 62. Hydrogen flow geometry.
Figure 63. Hydrogen inlet mesh.
noise resulting in local maxima. Polynomial RSAs were used to smooth the noise. Sturdza et
al.8 used a response surface to optimize the shape of a fuselage to retain laminar flow across a
wing. Clues to the accuracy of RSAs can be found through the simultaneous use of multiple
RSAs. Kurtaran et al.8 used linear, elliptic, and quadratic RSAs to optimize for
crashworthiness. Kurtaran et al. found that the differences between the predictive capabilities of
the various RSAs decreased as the design space was reduced. Hosder et al.89 USed RSAs in
conjunction with techniques to reduce the design space and perform multifidelity analyses in the
multidisciplinary optimization of aircraft. Additional references and their key results are given
in Table 21.
The polynomial RSA assumes that the function of interest f can be represented as a linear
combination of 4 basis functions z, and an error term e. For a typical observation i, a response
can be given in the form of a linear equation as
f(,()= S,,z B e (e,)=O 0 V(e)= cr (21)
where the errors 8, are considered independent with an expected value equal to zero and a
variance equal to o The coefficients a, represent the quantitative relation among basis functions
z,. Monomials are the preferred basis functions.
The relationship between the coefficients B, and the basis functions z, is obtained using N,
sample values of the response f for a set of basis functions z,(') such that the error in the
prediction is minimized in a least squares sense. For N, sample points, the set of equations
specified in Equation 21 can be expressed in matrix form as
f = XP+ e E (E)=0 V (E) = 0 (22)
DSR, the problem of insufficient grid refinement in the CFD model might not have been
discovered.
Understanding the mixing dynamics and the effects of the geometry on mixing is
important in the design of combustion devices such as inj ector flow. In this case, the CFDbased
surrogate model was able to reveal important aspects on the sensitivity of the mixing to the bluff
body geometry. It was found that the mixing could be increased by changing the shape of the
trailing edge of the bluff body. This increased mixing often came at the expense of increasing
drag on the body, however, locations of extreme cases revealed designs that gave favorable
mixing characteristics and low drag.
8.3 SingleElement Injector Flow Modeling
The inj ector flow modeling was conducted as a preliminary exercise to the combustion
chamber flow optimization. Flow was analyzed within the inj ector itself to determine
combustion chamber inlet conditions. Model verification was provided by developing a
computational model of an experimental combustion chamber. Finally, a grid sensitivity
analysis was conducted. It was discovered that the combustion length was more sensitive to the
grid resolution than the wall heat flux. Grid independency is not reached in the present study.
The wall treatment of the turbulence model reduces the sensitivity of wall heat transfer
computations. A grid resolution was selected for the inj ector flow optimization that provided
good tradeoff between heat flux profile accuracy and computational run time.
8.4 MultiElement Injector Flow Modeling
The CFD model sought to approximate the effects near a single inj ector element in the
outer injector row of multielement combustor flow. Geometric design variables were selected
to improve the performance and life of a rocket engine combustion chamber. The spacing
222
area is described in detail. The computational domain and bluff body geometry specifications
are based on the study performed by Burman et al.17
5.2.1 Geometric Description and Computational Domain
The dimensions of the bluff body are given in Figure 52. The variables of interest are B,
b, and h. Altering the variables changes the slant angles of the upper and lower surfaces. The
area, A, of the bluff body is held constant and equal to unity. The frontal height is kept constant
at D= 1. The value of H can be calculated from the aforementioned variables and constants
using the equation
H= 2A+lah b) l(D+B) (51)
The computational domain consists of a trapezoidal bluff body within a rectangular channel.
The flow area is illustrated in Figure 53. The fluid is incompressible, and the flow is laminar
with a Reynolds number of 250 given by
Re = "D(5 2)
where D is the frontal height of the bluff body, v is the kinematic viscosity, and L,. is the
freestream velocity. The upper and lower boundaries have "slip", i.e., zero gradient, boundary
conditions. The left boundary is the inlet, and the right boundary is the outlet.
The 2D, unsteady NavierStokes equations are solved using a CFD code called
STREAM.176 The time dependent calculations were solved using the PISO (Pressure Implicit
with Splitting of Operators) algorithm, and convective terms were calculated using the second
order upwind scheme. Other spatial derivatives are treated with the secondorder central
difference schemes. The grid was constructed using ICEMCFDin7 software. The grid is non
uniform, with a higher density of grid points near the body and in the wake area as shown in
Figure 54. Grids of varying resolution were used to determine the grid that provided the best
to act as a smoothing function. It can be used to approximate a highly nonlinear response with
fewer points than a comparable highorder response surface. Koch et al.91 asserts that using
Kriging may be preferable over RSAs for design problems with a large number of design
variables. Mahadevan et al.92 found that Kriging was more accurate than an RSA and used fewer
function evaluations for engineering reliability estimates.
Kriging has the capability of overcoming the limit of relatively small design space that is
inherent in RSAs, and can be used effectively for larger design spaces. It is gaining popularity in
CFDbased optimization as an alternative to RSAs. Forsberg and Nilsson93 found that Kriging
provided a better approximation as compared to an RSA in the structural optimization of
crashworthiness. However, Jin et al.94 found that Kriging was very accurate, but also very
sensitive to noise as compared to other surrogates in the data fitting of several analytical test
problems. Rijpkema et al.95 found that Kriging was better able to capture local details in an
analytical test function than an RSA, but also cautioned against the unintentional fitting of noise.
Chung and Alonso96 COmpared Kriging to RSAs and found that while both can be accurate,
Kriging is better at fitting functions with several local optima. Simpson et al.97 found that
Kriging and response surface performed comparably in the optimization of an aerospike nozzle.
Kanazaki et al.98 USed Kriging to reduce the computational cost involved in optimizing a three
element airfoil using genetic algorithms by using the Kriging surrogate for function evaluations.
Jouhaud, J. C. et al.99 USed Kriging to adaptively refine the design space in the shape
optimization of an airfoil.
Kriging is represented by a trend perturbed by a "systematic departure." In this study,
ordinary Kriging is used, where the constant trend is represented by the sample mean. Other
types of Kriging include simple Kriging that uses a constant trend of zero, and universal Kriging
qmax (MW/m2)
L (mm)
A4 B
1.1 1.1
1 1 ~s
09 "0.46 ,,
O, *PjiO Q 8
0.4
0 .6 0. 0 00. 0 .81 1.
0.4 0. 0.8 1 1.2
C '
Figure~.4 712 Krgn urgtsadmrtfntincnor o 2dsg onsin Pattern1
Krgn ftfr )qaxadB)L n meit fucinvle o C mxadD
White pont ar aa ons Bakpinsaeponschsnuin eifnto
analsis Cice blc ons r ons hsnfriuain
most of the Pareto front. This indicates that the 20 data points selected by M~Fz actually lie on or
very near the Pareto front. Because Kriging interpolates the data points, the prediction of points
near the 20 data points is very good which leads to the construction of an accurate Pareto front.
It is likely that the prediction of points away from the Pareto front would be very poor. Other
selection criteria were not as successful in predicting points along the Pareto front.
The absolute scaled error at each point is calculated as
elscaled =(34)
ymax ymm
and the absolute combined error is given as
le comlbined Cl scaled (3 5)
The error distribution for data points along the Pareto front for each selection criterion is given in
Figure 315. The use ofM~Fz results in the lowest combined error. Thus, it is demonstrated
qualitatively that selecting points using M~Fz results in the highest accuracy for data sets with a
low number of points.
Due to the random nature of the Matlab function kmeans that is used to generate the
clusters, the cluster sets can be slightly different depending on the starting point of the search,
especially if the data set being clustered is well distributed. For this reason, 100 different cluster
sets were used to reduce effects due to cluster selection. The mean and maximum error
distributions were compared for various numbers of data points and the different selection
criteria and are shown in Figure 316 and Figure 318, respectively. Figure 317 and Figure 319
provide direct comparisons of all selection criteria using the median values from Figure 316 and
Figure 318, respectively. By using the probability of improvement MFz as a selection criterion,
the overall accuracy was maintained to a surprising degree. Except for MF;, using merit
~1
0.3 
0.5 1 1.5 2
t; Region
05S 1 15S 2
0.4
05S 1 15 2
0.4
S0.3
0.
0.2 radeoff
01 region
0.5 1 1.5 2
0.5 1 1.5 2
Figure 312. Variation in design variables along Pareto Front.
Dhex %
3%
AnsqrFrac
5%
React*RPM
2%
Other
53%
Figure 313. Global sensitivity analysis. Effect of design variables on A) Wrotor and B) rls.
When additional points were selected based on large probabilities of improvement given as given
by the merit function M~Fz, improvements in the function values were obtained for every newly
selected variable combination as compared with the original data set. This indicates that the
search for best tradeoff designs can proceed efficiently with careful selection of design points
using merit functions.
Based on a computational model for an inj ector element near the combustion chamber wall
of a multielement inj ector, it was found that an inj ector element that is located far from the
combustion chamber wall does not always result in the best heat transfer characteristics. The
results of the multielement inj ector spacing study suggest that by increasing the spacing
between inj ector elements of the outer row, while reducing the distance of the outer row to the
wall, that the heat transfer results could be better controlled. This configuration helps to direct
heat away from the wall, rather than towards it, and results in an even distribution of heat across
the combustion chamber wall. Sufficiently increasing the distance to the combustion chamber
wall can also result in reduced peak heat. In the future, further exploration into each phenomenon
should provide additional insight.
Table 71. Flow conditions and baseline combustor geometry for parametric evaluation.
Fuel annulus outer diameter (mm) 7.49
Fuel annulus inner diameter (mm) 6.30
Oxidizer post inner diameter (mm) 5.26
Oxidizer post tip recess (mm) 0.43
Combustion chamber diameter (mm) 262
rbasebne (mm) 6.59
Nbasehne 60
Total fuel mass flow rate in outer row (kg/s) 1.986
Total oxidizer mass flow rate in outer row (kg/s) 5.424
Fuel temperature (K) 798.15
Oxidizer temperature (K) 767.59
Chamber pressure (MPa) 5.42
Hz mass percentage of fuel 41.3
HzO mass percentage of fuel 58.7
Oz mass percentage of oxidizer 94.62
HzO mass percentage of oxidizer 5.38
204
as the optimizer would often become stuck at noiseinduced local minima. Also, for simulating
designs with long run times, it may be more time efficient to run simulations in parallel, as this
will allow many simulations to run simultaneously. For these problems, optimization based on
an inexpensive surrogate is a good choice. Surrogatebased optimization (SBO) uses a
simplified, loworder characterization of the design space instead of relying exclusively on
expensive, individually conducted computational simulations or experimental testing. Surrogate
based optimization allows for the determination of an optimum design while exploring multiple
design possibilities, thus providing insight into the workings of the design. A surrogate model
can be used to help revise the problem definition of a design task by providing information on
existing data trends. Furthermore, it can conveniently handle the existence of multiple desirable
design points and offer quantitative assessment of tradeoffs, as well as facilitate global
sensitivity evaluations of the design variables to assess the effect of a change in a design variable
on the system response.52,53 The SBO approach has been shown to be an effective approach for
the design of computationally expensive models such as those found in aerodynamics,54
structures," and propulsion.56 The choice of surrogates may depend on the problem. For
example, Shyy et al.5 compared quadratic and cubic polynomial approximations to a radial basis
neural network approximation in the multiobj ective optimization of a rocket engine inj ector.
Simpson et al.5 suggested using Kriging as an alternative to traditional quadratic response
surface methodology in the optimization of an aerospike nozzle. Dornberger et al.59 COnducted a
multidisciplinary optimization on turbomachinery designs using response surfaces and neural
networks.
To construct the surrogate model, a sufficient number of different designs must be tested to
capture how a system response varies with different design parameters. The process of selecting
Foust et al.146 mOdeled GO2/GH2 COmbustion using an 18 reaction finiterate chemical
model. The chemistry parameters were determined based on the temperature field. The kE
turbulence model was used. Preliminary computations were done to obtain the boundary
conditions upstream of the injector. The final computational model extends slightly upstream of
the injector exit. The model is 2D Cartesian, and the grid was coarse at 101x51. The
computation found good agreement in the species concentration profiles and the velocity
profiles. Using CFD, Foust confirmed the flame holding ability of the oxidizer post tip over
wide ranges of equivalence ratios for a given inj ector.
Cheng et al.15 looked to develop a CFD spray combustion model to help understand the
effects on wall erosion. The Finite Difference NavierStokes (FDNS) solver used a kE model
with wall functions, and the real fluids models were used for the multiphase flow. Heat and
mass transfer between phases was neglected. The velocity and species concentrations were
solved based on a constant pressure assumption, and then the density and temperature ere
determined based on the real fluids model. The pressure was then corrected based on the newly
determined density. Finiterate and equilibrium chemistry models were used. The GO2/GH2
computation used the 4equation equilibrium chemistry model for hydrogenoxygen combustion.
The shear layer growth was well predicted by the model. Good comparisons were made between
the CFD and experiment for velocity and species profiles. Disagreement in the H20 species
profile was attributed to experimental measurement error. It is assumed that the grid was two
dimensional, however, no grid information is provided.
Schley et al.152 COmpared the Penn State University (PSU) code used by Foust et al.146
and the NASA FDNS code used by Cheng et al.15 with a third computational code by called
Aeroshape 3D153 (AS3D). The purpose in the CFD development was to reduce combustor
SSE for objective 1
SSE for objective 2
0.16
0.14
0.12
0.080
0.06
1 2 3 4 5
1 2 3 4 5
Figure 28. Variation in SSE with p for two different responses in the preliminary optimization
of a radial turbine for the unrefined design space (Data Set 2).
 p=3
p=4
p=4
018~
:::~
012~
008
S2 25
4 OB 08 1 12 14
Rotor Wt
18 2 22 24
Rotor Wt
Figure 29. Pareto fronts for RSAs constructed with varying values of p for A) the original
design space and B) the refined design space. The blue dots represent actual
validation data along the Pareto front constructed on the refined design space using
standard least squares regression.
125. BosqueSendra, J., Pescarolo, S., CuadrosRodrigues, L., GarciaCampana, A., and
AlmansaLopez, E., "Optimizing Analytical Methods Using Sequential Response Surface
Methodology. Application to the Pararosaniline Determination of Formaldehyde," Journal
ofAnalytical Chemistry, Vol. 369, 2001, pp. 715718.
126. Immich H. and Caporicci, M., "FESTIP Technology Developments in Liquid Rocket
Propulsion for Reusable Launch Vehicles," AIAA Paper 19963113, 1996.
127. Huber, F., "Turbine Aerodynamic Design Tool Development," presented at the Space
Transportation Fluids Workshop, Marshall Space Flight Center, AL, April, 2001.
128. Sondak, D. L. and Dorney, D. J., "General Equation Set Solver for Compressible and
Incompressible Turbomachinery Flows," AIAA Paper 20034420, July 2003.
129. Venkateswaran, S. and Merkle, C. L., "Analysis of Preconditioning Methods for the Euler
and NavierStokes Equations," Von Karman Institute Lecture Series, March 812, 1999.
130. Baldwin, B. S., and Lomax, H., "Thin Layer Approximation and Algebraic Model for
Separated Turbulent Flow," AIAA Paper 1+78257, Jan. 1978
13 1. NIST Chemistry WebBook [online database], URL: http://webbook.ni st.gov/chemi stry/fluid
[cited 19 July 2006]
132. Dorney, D. J., Sondak, D. L. and Marcu, B., "Application of a RealFluid Turbomachinery
Analysis to Rocket Turbopump Geometries," AIAA Paper 051007, Jan. 2005.
133. Pavli, A. L., "Design and Evaluation of High Performance Rocket Engine Inj ectors for use
with Hydrocarbon Fuels," NASA TM69319, 1979.
134. Calhoon, D., Ito, J., and Kors, D., "Investigation of Gaseous Propellant Combustion and
Associated Inj ector/Chamber Design Guidelines," NASA TM 12 123 4, 1 973.
135. Hutt, J. J. and Cramer, J. M., "Advanced Rocket Injector Development at Marshall Space
Flight Center," AIAA Paper 964266, 1996.
136. Rupe, J. and Jaivin, G., "The Effects of Inj section Mass Flux Distributions and Resonant
Combustion on Local Heat Transfer in a LiquidPropellant Rocket Engine," NASA
Technical Report 32648, 1964.
137. Farhangi, S., Yu, T., Roj as, L., and Sprouse, K., "GasGas Inj ector Technology for Full
Flow Stage Combustion Cycle Application," AIAA Paper 19992757, 1999.
138. Branam, R. and Mayer, W., "Length Scales in Cryogenic Inj section at Supercritical
Pressure," Experiments in Fluids Vol. 33, 2002, pp. 422428.
139. Quentmeyer, R. J. and Roncace, R. A., "HotGasSide Heat Transfer Characteristics of
Subscale, PlugNozzle Rocket Calorimeter Chamber," NASATP33 80, July 1993.
234
Figure 44. Wall burnout in an uncooled combustion chamber.
147.3mm(5W.90" I 1194mm(4 7U")
S268 70mm (10 5017')
~ 285.75mm (111.2517)
~~ 3111.Bam (12271")
~ 336.55mm (13 250")
Figure 45. Test case RCM1 injector. Test rig A) schematic and B) photo.
To test if the response surface fit was better away from the extreme cases, an exploration
of the design space was conducted. Figure 512 plots the response surface and radial basis
neural network predicted values as compared to the actual CFD data from
Table 57 along the line M~.L = f(B* = 0, b*", h* = 0). Of the cases used to construct the
surrogate models, only the extreme cases (B,b,h) = (0,0,0) and (B,b,h) = (0,0,1) fall on this line.
The radial basis neural network predicts both of these points very well, while the response
surface underpredicts the mixing index at b = 1. The variation in the data is very high, with a
sharp drop in the mixing index value over a short distance near the extreme cases and very
smooth behavior away from the extreme cases. Neither surface captures the complete behavior.
In this case, the response surface predicted the interior points very well, as expected, but
performed poorly near the extreme cases. The neural network, on the other hand, better predicts
the trends in the data near the extreme cases.
Areas of large differences within the design space between the two surrogates correspond
to regions where the accuracy in the CFD model is compromised. Particularly, these regions are
concentrated around the extreme designs. These designs were found by Goel et al.m7 to have not
reached grid convergence. The results indicate that the extreme designs may require a greater
degree of grid refinement than other design points.
5.4 Conclusions
Polynomial response surface and radial basis neural networks were used as surrogates for
flow over a trapezoidal bluff body. Both the response surface and the radial basis neural network
approximations adequately predicted the total pressure loss coefficient. However, both
surrogates poorly fit the mixing index and were substantially different from each other. The
large differences in the surrogate approximations served as a warning signal and prompted
further investigation. Inspection of the surrogate models' statistics revealed that the fitting
4,,a~,,,, rwlannIn (ro /r,) (65
For the given combustion chamber diameter, this assumption would result in a wall heat flux
value that is everywhere 28% below the linear heat flux assumption. This is reasonable, as the
assumption of axisymmetric isotherms would result in the lowest possible heat flux given the
temperatures at the thermocouple locations. The temperature at the wall would be given by
T () Inr + To (66)
Taxlsymm ( o)
An unsteady approximation is determined by adding the second term from equation (62) to the
steady state heat flux approximations in Equation 65. Figure 613 shows the estimated wall
temperature profiles based on the experimental thermocouple readings. The estimated wall
temperatures are very similar between the linear and axisymmetric approximations, indicating
that the linear assumption used in the experimental documentation was reasonable in the case of
the wall temperature.
Finally, an approximation of the wall heat flux was conducted using a numerical two
dimensional unsteady heat conduction analysis. Vaidyanathan et al.lso completed a similar
analysis on a different inj ector and provided the conduction analysis results for the inj ector by
Conley et al.22 The crosssections at each thermocouple pair location were analyzed
independently. Because the experiment lasted only seven to eight seconds, it was estimated that
only negligible heat escaped through the outer combustion chamber wall. Therefore, the outer
wall was assumed to be adiabatic. The thermal propagation time scale for the combustion
chamber wall is equal to tp = e2/a = 3.13 s where eis the thickness of the combustion chamber
wall equal to 0.019 m, and a is the thermal diffusivity and is equal to 1.158 x 104 m2/S. The
prediction ability of a neural network is problem dependent. Hiisken et al.104 USed neural
networks in turbine blade optimizations. Shi and Hagiwaralos used neural networks to maximize
the energy dissipation of the crashworthiness of a vehicle component. Pidaparti and Palakall06
compared the prediction capability of neural networks with experimental results in the prediction
of crack propagation in aging aircraft. Chan and Zhu107 demonstrated the ability of neural
networks in modeling highly nonlinear aerodynamic characteristics with many variables.
Brigham and Aquinolos used neural networks to accelerate their optimization search algorithm
by providing inexpensive function evaluations.
Neural networks are made up of a distribution of artificial neurons, whose function is
modeled after biological neurons. A shape function is applied at each neuron with the
combination being a complete function curve. Each neuron is given a weight in the network that
defines that individual neuron's strength. Neural networks are often difficult to train due to the
required specification of several parameters that can affect the fit of the model. Care must also
be taken, as there is sometimes danger in overfitting the data, resulting in poor prediction
capability away from the known data points. Nonetheless, the models are extremely flexible and
can provide an exact data fit, or meet a userdefined error criterion to filter noise to prevent data
overfitting. Neural networks use only local models, and the degree of effective influence of
each local model, or neuron, is a selection made by the user.
Radial basis neural networks are twolayer networks consisting of a radialbasis function
(RBF) and a linear output layer. The radial basis function is given by
f= radbas c xb) (213)
where b is the bias, c is the center vector associated with each neuron, x is the input design
vector, and the function radba~s is the Matlab function name for an RBF, or a local Gaussian
154. Ivancic, B., Mayer, W., Kriille, G., and Briiggemann, D., "Experimental and Numerical
Investigation of Time and Length Scales in LOX/GH2Rocket Combustors, AIAA Paper
19992211, 1999.
155. Lin, J., West, J., Williams, R., Tucker, K., and Chenoweth, J., "CFD Code Validation of
Wall Heat Fluxes for a GO2/GH2 Single Element Combustor," AIAA Paper 20054524,
2005.
156. West, J., Westra, D., Lin, J., and Tucker, P. K., "Accuracy Quantification of the Loci
CHEM Code for Chamber Wall Heat Fluxes in a GO2/GH2 Single Element Inj ector Model
Problem," 3'd Interizational Workshop on Rocket Combustion M~odeling, Paris, France,
2005.
157. Thakur, S. and Wright, J., "Validation of a PressureBased Combustion Simulation Tool
for a Single Element Injector Test Problem, 3'd lzterizational Workshop on Rocket
Combustion M~odeling, Paris, France, March 1315, 2006.
158. Tucker, K., West, J., Williams, R., Lin, J., Rocker, M., Canabal, F., Robles, B., and Garcia,
R., "Using CFD as a Rocket Inj ector Design Tool: Recent Progress at Marshall Space
Flight Center," NASA Marshall Space Flight Center, CASI ID 20050217148, 2005.
159. Meyer, L., Nichols, J., Jones, J. M., and Sabol, W., "Integrated Powerhead Demonstrator
(Booster Hydrogen Oxygen Rocket Engines)," AIAA Paper 19964264, 1996.
160. Beer, J. M. and Chigier, N. A., Combustion Aerodynamnics, Robert E. Krieger Publishing
Company, Malabar, FL, 1983, p. 1.
161. Hill, P. and Peterson, C., M~echanics and 7hernzodynantics ofPropulsion, AddisonWesley,
1992, pp. 124130.
162. Turns, S. R., An Introduction to Combustion, McGraw Hill, 2000, p. 486.
163. Launder, B. and Spalding, D., "The Numerical Computation of Turbulent Flows,"
Computer M~ethods in Applied Mechanics and Engineering, Vol. 3, No. 2, 1974, pp. 269
289.
164. Patel, V., Rodi, W., and Scheuerer, G., "Turbulence Models for NearWall and Low
Reynolds Number Flows: A Review," AIAA Journal, Vol. 23, No. 9, 1985, pp. 13081319.
165. Avva, R., Smith, C., and Singhal, A., "Comparative Study of High and Low Reynolds
Number Versions of kepsilon Models," AIAA Paper 19900246, 1990.
166. Chien, K., "Predictions of Channel and BoundaryLayer Flows with Low Reynolds
Number Turbulence Model," A1AA Journal, Vol. 20, No. 1, 1982, pp. 3338.
167. Pope, S. B., Turbulent Flows, Cambridge University Press, New York, 2000, pp. 442444.
236
problems were due to high variations in the data in localized regions. Removing three extreme
cases greatly improved the fits. Further investigation revealed that the CFD simulations were not
converged, and may have been contributing to the inability of the surrogates to properly
approximate the model.
Simulations of complex flows can sometimes result give unexpected results. Here, three
cases had considerably higher mixing indices than the other cases. This behavior could not be
accurately captured by a single surrogate model alone. The radial basis neural network was
found to better approximate the response near the extreme cases due to its local behavior, while
the response surface provided better prediction of the response away from the extreme cases.
Refinement of the design space near the extreme designs revealed that the differences in the
surrogate models corresponded to the locations of the low accuracy CFD designs. Thus,
surrogate models and design space refinement techniques can be used to identify the presence
and location of inaccurate CFD models.
Table 51. Number of grid points used in various grid resolutions.
Grid # Number of grid points
1 37,320
2 44,193
3 74,808
4 147,528
Table 52. Data statistics in the grid comparison of the CFD data.
CD .
Max 2.21 857
Min 1.79 515
Range (maxmin) 0.42 342
Mean 2.03 617
Standard deviation 0.11 66
unsteady and difficult to predict. There is recirculating flow in the nearwake region that decays
to form a welldeveloped vortex street in the wake region. The instantaneous loss and degree of
mixing changes over time and must be well resolved for accurate solutions.
An earlier investigation of the optimization of a trapezoidal bluff body was conducted by
Burman et al.138 The study uses a relatively coarse mesh (9272 computational cells) that was
selected by performing a grid sensitivity study on a single case based on the drag coefficient. The
effects of grid resolution on the measure of mixing were not investigated, so it is possible that
not all of the flow effects were captured. For higher Reynolds number flow, as observed by
Morton et al.,174 grid resolution can have a marked effect on the prediction of unsteady flows and
can substantially affect the fidelity of the surrogate model.
Surrogate models are used to approximate the effect of bluff body geometry changes on
total pressure loss and mixing effectiveness. The timeaveraged flow field solutions are
compared by looking for common trends and correlations in the flow structures.
5.2 Bluff Body Flow Analysis
In practical considerations, such as the aforementioned afterburner and ramj et combustor,
the flow around many bluff body devices is turbulent. To simplify the analysis, an effective
viscosity is often estimated based on engineering turbulence closures. The effective Reynolds
number is therefore considerably lower than the nominal Reynolds number. For the purpose of
this study, the flow is modeled using a Reynolds number in the range of the effective Reynolds
number without resorting to turbulence models. Furthermore, without either enforcing the wall
function or suppressing high gradient regions, both typically observed in engineering turbulent
flow computations, the fluid flow tends to exhibit more unsteady behavior, rendering the flow
computations more interesting and challenging. The bluff body geometry along with the flow
Tables.131 A detailed description of the code/algorithm development, as well as its application to
several turbine and pump test cases, is presented in Venkateswaran and Merklel29 and Dorney et
al.132
Figure 31 shows static pressure contours (psi) at the midheight of the turbine for the
baseline rotational speed of 122,000 RPM. This figure illustrates the geometry of the turbine, and
the contours indicate that the pressure decrease is nearly evenly divided between the vane and
the rotor. In fact, the reaction was approximately 0.60 for each of the three 3D simulations as
compared to 0.55 for the 1D simulation. Figure 32 contains the predicted totaltostatic
efficiencies from the Meanline and CFD analyses. The CFD results include values with and
without tip clearance. In general, fair agreement is observed between the Meanline and CFD
results. The trends are qualitatively similar, but the Meanline analysis predicts higher
efficiencies. There is approximately a fourpoint difference in the quantitative values. The
quantitative differences in the results are not surprising considering the lack of experimental data
available to anchor the Meanline code. The differences in the predictions with and without tip
clearance decrease with increasing rotational speed. Figure 33 shows the predicted work from
the Meanline and CFD analyses. The trends are again similar between the Meanline and CFD
analyses, but the Meanline values are consistently 5 6% higher than the CFD values.
The similar trends between the 1D Meanline code and 3D CFD analyses indicate that the
optimization can be confidently performed on the 1D Meanline code. It can be expected that for
a given turbine speed the Meanline code will overpredict the totaltostatic efficiency by an
expected degree. The predicted optimum point based on the 1D Meanline code will likely yield
overly optimistic results, but the predicted degree of improvement should translate to the 3D
CFD analysis.
where X is a Ns x No matrix of basis functions, also known as a Gramian design matrix, with the
design variable values as the sampled points. A Gramian design matrix for a quadratic
polynomial in two variables (Ns = 2; Nc = 6) is given by
1 x,, x21 121 11x, 21
1x12 X22 212 X12 22 X2
X ~ ~ =~~ X21 (23)
1 x,, x2; 12 2 2
The vector b of the estimated coefficients, which is an unbiased estimate of the coefficient vector
(3 and has minimum variance, can then be found by
b =(XTX) XTf (24)
At a new set of basis function vector z for design point P, the predicted response and the variance
of the estimation are given by
.,(z)=~n b z ') and V .,(z,) = c2 T XT)z (25)
J=1
2.1.2.2 Kriging
In cases where a given surrogate performs poorly, it is necessary to use a different type of
surrogate model, because the highly nonlinear nature of some processes may not be captured by
surrogates such as RSAs. Kriging is a popular geostatistics technique named after the pioneering
work of D.G. Krige47 and was formally developed by Matheron.90 The Kriging method in its
basic formulation estimates the value of a function or response at some location not sampled as
the sum of two components: the linear model (e.g., polynomial trend) and a systematic departure
representing low (large scale) and high frequency (small scale) variation components,
respectively. Kriging has the added flexibility of being able to either provide an exact data fit or
method in the RSA of a highspeed civil transport model. The approach removes large areas of
the standard boxshaped design leaving more of a simplex or ellipsoid design. Inexpensive, low
fidelity models are used to predict the system response, and the result is used to shape the design
space into a reasonable design space. Points that would lie on the edge of the standard box in a
traditional DOE procedure are moved inward to the boundary of the identified reasonable design
space. By concentrating the data points on the region of interest, the accuracy of the response
surface within that region is improved. The technique was used to separately estimate the
structural bending material weight and the lift and drag coefficients. In these cases, the design
space was reduced at least by half, resulting in improvements in RSA accuracy.
Roux et al.73 employed techniques such as intermediate response surfaces and
identification of the region of interest to improve surrogate accuracy in structural optimization
problems. Linear and quadratic response surfaces were used to fit the data. The design space
was iteratively reduced around the suspected optimum by fitting response surfaces on smaller
and smaller subregions. It was found that surrogate accuracy was improved more for carefully
selected points within a small subregion than for a large number of design points within a large
design space. Both Balabanov et al.70 and Papila et al.82 COmbine multiple obj ectives into a
composite objective function before optimizing. However, this procedure can limit design
selection to one or a few designs, when multiple other desirable designs might exist.
Papila et al.7 pointed out that the inaccuracy in the initial response surface may result in
windowing to a region that does not actually contain the optimum design. Papila et al. argued
that particular attention must be paid to the regions of higher error in an RSA. Papila et al.
explains that because a quadratic RSA is a loworder polynomial, the model can often be
inadequate. This is termed as "bias error," and can obviously be reduced by using higherorder
2.1 Optimization Using Surrogate Models
This step in the framework involves developing alternate models based on a limited
amount of data to analyze and optimize designs. The surrogates provide fast approximations of
the system response making optimization and sensitivity studies possible. Response surface
approximations, neural network techniques, spline, and Kriging are examples of methods used to
generate surrogates for simulations in the optimization of complex flows52 inVOlving applications
such as engine diffusers,74 TOcket inj ectors,63 and supersonic turbines.78,82
The maj or benefit of surrogate models is the ability to quickly obtain any number of
additional function evaluations without resorting to more expensive numerical models. In this
aspect, surrogate models can be used for multiple purposes. Obviously, they are used to model
the design obj ectives, but they can also be used to model the constraints and help identify the
feasible region in design space. Key stages in the construction of surrogate models are shown in
Figure 21.
2.1.1 Design of Experiments
The search space, or design space, is the set of all possible combinations of the design
variables. If all design variables are real, the design space is given as x E R" where N is the
number of design variables. The feasible domain S is the region in design space where all
constraints are satisfied.
For adequate accuracy, the data points used in the surrogate model must be carefully
selected. The proper data selection is facilitated by using Design of Experiments (DOE). One
challenge in design optimization despite the type of surrogate used is called the "curse of
dimensionality." As the number of design variables increase, the number of simulations or
experiments necessary to build a surrogate increases exponentially. This must be taken into
consideration in choosing a DOE. The key issues in the selection of an appropriate DOE include
CHAPTER 4
MODELINTG OF INJECTOR FLOWS
The combustion chamber of a liquid rocket engine is cooled by transferring heat to the
unburned fuel that circulates around the outside of the combustion chamber via a series of tubes
or coolant channels. Rocket inj ectors deliver the fuel and oxidizer to the combustion chamber.
Commonly, the inj ector face is made up of a series of inj ector elements arranged in concentric
rows. The row of inj ectors near the chamber wall can cause considerable local heating that can
reduce the life of the combustion chamber. The local heating near each inj ector element takes the
form of a sinusoidal wall heat flux profile caused by the interactions of the outer row elements.
Often, this effect is not considered during the design process. Reducing the intensity of this local
heating is of prime importance. For example, in the design of the Space Shuttle Main Engine
(SSME), the effect of the local heating in the combustion chamber was not considered during the
design process. After construction, it was discovered that hotspots along the chamber wall
severely reduced the expected chamber life. This resulted in an unforeseen increase in the
reusability operating costs of the engine. Local hotspots can even cause actual burnout of the
chamber wall. By accurately predicting potentially detrimental phenomena in advance, it may be
possible that issues such as wall burnout can be avoided. CFD modeling and validation efforts, in
conjunction with the experimental data, can assist in the understanding of combustor flow
dynamics, eventually leading the way to efficient CFDbased design.
This chapter outlines current and proposed methodologies in the CFDbased optimization
of liquid rocket engine components. First, a summary is presented of past and present inj ector
analysis techniques. Then, the basic governing equations are presented that are applicable in the
simulation of turbulent reacting flow. Finally, a simplified onedimensional analysis is
021( H20
,o2
.O
0.8
c*
.9 0.6
O 0.4
E
0.2
0
1 00
1Q
O
E 1U"
1U'
1U"
x (mm)
B x (mm)
Figure 625. Mole fractions for all species along combustion chamber centerline (y = 0, z = 0).
A) uniform and B) log scaling.
187
 Actual
MF 1
MF 2
MF 3
MF 4
fmin
centers
MF3
S0.14
0.12 E
I I I I I I
04 06 08 1 12
Wroo
14 16 18
Figure 314. Data points predicted by validated Pareto front compared with the predicted values
using six Kriging models based on 20 selected data points. The validated Pareto front
is labeled "Actual."
Wrtr
 (
 I
combined
Absolute Error distribution
i
j 
Absolute Error Distribut on
S015
01
111 015
01
O5
Figure 315. Absolute error distribution for points along Pareto front using 20 selected data
points each where the points were selected using (1) MF;, (2) M~F2, (3) M~F3, (4) M~F4,
(5) fou,, and (6) cluster centers.
multiobj ective optimization problems, the use of evolutionary algorithms (EAs) is a natural
choice to get many Pareto optimal solutions in a single simulation due to its population based
approach and ability to converge to global optimal solutions.
A feasible design x l) dominates another feasible design X(2) (denoted by x l) < X(2)), if both
of the following conditions are true:
1. The design x(1) is no worse than x(2) in all objectives, i.e., ~f, (x )< fJ x(2') for all j 
1,2,..., M obj ectives.
x /I~X(2) > Vj eMf(L~) x ; fX(2~) OrT Vj et h x )I <:f (X(2))(1
2. The design x l) is strictly better than X(2) in at least one obj ective, or ~f, x ) < ~f, x(2) foT
at least one je {1t, 2,...,M} ).
x(1 < X(2)> Aj E IfJ((x) )< fJ(X(2)) (232)
If two designs are compared, then the designs are said to be nondominated with respect to
each other if neither design dominates the other. A design x e S where S is the set of all
feasible designs, is said to be nondominated with respect to a set A c S if Bae A: a < x
Such designs in function space are called nondominated solutions. All the designs x (x e S)
which are nondominated with respect to any other design in set S, comprise the Pareto optimal
set. The function space representation of the Pareto optimal set is the Pareto optimal front.
When there are two obj ectives, the Pareto optimal front is a curve, when there are three
obj ectives, the Pareto optimal front is represented by a surface and if there are more than three
obj ectives, it is represented by a hypersurface.
In this shtdy, an elitist nondominated sorting genetic algorithm NSGAIIll3 and a parallel
archiving strategy to overcome the Pareto drift problem114 are used as the multiobj ective
the combustion length, while case 12 gives arguably the worst characteristics due to its high
maximum heat flux and long combustion length. Figure 73 also shows the maximum heat flux
and combustion length as a function of the hydrogen mass flow rate. It is apparent that the
maximum heat flux is not a function of the mass flow rate alone, but that there must be other
influencing factors. There does seem to be an overall decrease in the maximum heat flux with the
mass flow rate, as well as an overall increase in combustion length with increasing mass flow
rate, which is a function of N*. No obvious relationship was seen between the maximum heat
flux or combustion length and the crosssectional area. This indicates that, a smaller cross
sectional area does not automatically translate into a longer combustion length, as one might
expect. This was also demonstrated by the grid sensitivity study.
The CFD results for four cases, representing the baseline case along with cases
demonstrating good or poor heat flux characteristics, are shown in Figure 74. It can be seen that
the flame is flattened somewhat based on the shape of the computational domain. Another maj or
observation is of the location of the recirculation region. Instead of hot gases being directed at
the wall, as in the baseline case and case 12, the recirculation region of case 2 and case 6 is
located between inj ector elements, and the streamlines are largely parallel to the combustion
chamber wall. This led to an observation that all of the CFD solutions resulted in one of two
scenarios. Eleven of the cases showed a sharp peak in the wall heat flux near the inj ector face
with the downstream heat flux considerably lower than the peak heat flux. This is termed
"Pattern 1," and corresponds to cases that have hot gases directed at the wall as in Figure 75A.
The remaining five cases had a relatively low heat flux near the combustion chamber wall that
grew gradually to a slight peak further downstream. This group is termed "Pattern 2," and
corresponds to the cases that have a recirculation zone that would exist between inj ector
demonstrated to help define the relationship between flow parameters and heat transfer due to
reacting flow.
4.1 Literature Review
Inj ector research has been ongoing for over 40 years. Inj ector inlet flow has a significant
influence on combustor performance. The understanding of the inj ector characteristics is critical
in determining the nature of the flow within the combustion chamber. In particular, the inlet
flow geometry and inj ector outlet diameter have large influences on flow in the combustion
chamber.26 However, at times, the research has raised more questions than answers. Work must
still be done to explain certain elements of the flow within the combustion chamber. A general
diagram of inj ector flow is given in Figure 41.
Historically, inj ectors have been designed using experimental techniquesl33 and empirical
calculations. A design was built and tested, and then improvements were made based on the
results. For example, Calhoon et al.134 extensively reviewed standard techniques for inj ector
design, including the analysis of coldfire and hotfire testing to study general injector
characteristics. The results of these tests were used along with a number of additional multi
element coldfire tests to design the full inj ector. The full injector was then fabricated and tested
for performance and heat flux characteristics and combustion stability. However, the
experimental design techniques were insufficient to predict for many conditions that could
reduce inj ector or combustion chamber life.
4.1.1 SingleElement Injectors
A significant portion of inj ector research has been conducted using experiments consisting
of a single inj ector element. The single element analysis is often used as a starting point in
modeling full combustor flow. A test firing of a single element inj ector is shown in Figure 42.
Hutt and Cramerl35 found that if it is assumed that all of the inj ectors in the core are identical,
O 2007 Yolanda Mack
168. Tennekes, H. and Lumley, J., A First Course in Turbulence, MIT Press, Cambridge, MA,
1972, p. 156.
169. Libby, P. A. and Williams, F. A., Turbulent Reacting Flows, Academic Press, New York,
1994.
170. Thakur, J., Wright, J., and Shyy, W., "A PressureBased Algorithm for Reacting Flows
with Finite Rate Chemistry," AIAA Paper 20042334, 2004.
171. Weske, D. R. and Sturov, G. YE., "Experimental Study of Turbulent Swirled Flows in a
Cylindrical Tube," Fluid Mechanics Soviet Research, Vol. 3, No. 1, 1974.
172. Yajnik, K. S. and Subbaiah, M. V., "Experiments on Swirling Turbulent Flows. Part 1.
Similarity in swirling flows," Journal ofFluid2\~echa~nics, Vol. 60, No. 4, 1973, pp. 665
687.
173. Talbot, L., "Laminar Swirling Pipe Flow," Journal ofApplied2\~echa~nics, Vol 21., No. 1,
1954, pp. 17.
174. Morton, S. A., Steenman, M. B., Cummings, R. M., and Forsythe, J. R., "DES Grid
Resolution Issues for Vortical Flows on a Delta Wing and an F18C," AIAA Paper 2003
1103, 2003.
175. Burman, J., Papila, N., Shyy, W., and Gebart, B. R., "Assessment of Response Surface
Based Optimization Techniques for Unsteady Flow Around Bluff Bodies," AIAA Paper
20025596, 2002
176. Thakur, S. and Wright J., "STREAM A Computational Fluid Dynamics and Heat
Transfer Code NavierStokes Solver: Theory & Applications," Streamline Numerics,
Gainesville, Florida, 2002.
177. ICEM CFD version 5.0, ICEM CFD Engineering, 2004.
178. Goel, T., Mack, Y., Shyy, W., Haftka, R. T., Queipo, N.V., "Numerical and Surrogate
Model Uncertainty Assessment for Bluff Body Facilitated Mixing," AIAA Paper 2005
0125, 2005.
179. Incropera, F. P. and DeWitt, D. P., Fundamentalsd~~~dd~~~ddd~~ of Heat and Ma~ss Transfer, John Wiley
& Sons, Inc., New York, 1996, pp. 863.
180. Vaidyanathan, A. and Segal, C., "Quantitative Measurements of OH Concentration in
GH2/GO2 combustion," AIAA Paper 20075591, 2007.
181. Wang, B. P., "A New Method for Dual Response Surface Optimization," AIAA Paper
19964041, 1996.
237
Figure 617. Velocity contours v,(m/s) and streamlines. The inj ector is located in the lower
right corner of the combustion chamber. The reattachment point is at x = 77 mm.
10
0 20 40 60 80 100 120 140 160
_I
~
Ylh,_
 ~
: II
I~' I~.r~
F j~i ~i.~~
i 'I
Figure 618. Temperature (K) contours. Injector center is at x,y :
millimeters .
(0,0). Distances are given in
* D linear
axisymmetric
*unsteady adiabatic
CFD
r
r \
*
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
x (m)
Figure 619. CFD heat flux values as compared to experimental heat flux approximations.
I i
LIST OF FIGURES
Figure page
11 Rocket engine cycles............... ...............33.
12 SSME thrust chamber component diagram .............. ...............33....
13 SSME component reliability data ................. ...............34...............
14 Surrogatebased optimization using multifidelity data ................. ................. ...._.34
21 Optimization framework flowchart. ........_................. ...............69. ....
22 DOEs for noisereducing surrogate models. .............. ...............70....
23 Latin Hypercube Sampling. ............. ...............70.....
24 Design space windowing .............. ...............71....
25 Smart point selection............... ...............7
26 Depiction of the merit function rank assignment for a given cluster. .............. .... ..........._72
27 The effect of varying values of p on the loss function shape.............__ ........._ ......72
28 Variation in SSE with p for two different responses ........... _...... __ ........_......73
29 Pareto fronts for RSAs constructed with varying values of p............... ..................7
210 Ab solute percent difference in the area under the Pareto front curves ................... ...........74
31 Midheight static pressure (psi) contours at 122,000 rpm. ............. ......................9
32 Predicted Meanline and CFD totaltostatic efficiencies. ................. .................9
33 Predicted Meanline and CFD turbine work. ............. ...............95.....
34 Feasible region and location of three constraints ..........._._ ......_.._ .. ......_.._......9
35 Constraint surface for Cx2 Utip = 0.2............... ...............96..
36 Constraint surfaces for B1 = 0 and B1 = 40. ................ ............... ......... ....... ..96
37 Region of interest in function space ................. ...................... ..................97
38 Error between RSA and actual data point............... ...............97.
39 Pareto fronts for p = 1 through 5 for second data set............... ...............98..
MI I
b
.
*
,e
III
Figure 22. DOEs for noisereducing surrogate models. A) Central composite design and B)
face centered cubic design for two design variables. The extreme points are selected
around a circle for CCD and a square for FCCD.
Figure 23. Latin Hypercube Sampling. A) LHS with holes in the design space. B) Orthogonal
array LHS can help address this issue by filling the design space more evenly.
surrogatebased DSR using merit functions for point selection was thus successful in providing
optimal design using a small number of design points.
8.5 Future Work
This study provides a starting point for CFDbased design of rocket engine components.
In particular, this study provides a new method of attacking complex rocket engine component
design. Future studies may include
1. Full 3D simulations of the selected optimum radial turbine design and a comparison with
the 1D Meanline code results to determine if the efficiency gains are comparable over the
baseline case.
2. A threedimensional blade shape optimization of the selected optimum radial turbine
design.
3. Adding additional design points near the selected designs of the multielement inj ector
analysis. Doing so would provide a more accurate representation of the design space near
the selected designs.
4. Improving the accuracy of the multielement inj ector model by including the second row
of inj ector elements in the CFD model. In this case, the best designs as selected in the
current study can provide a starting point for the additional analysis. The effect of the
inj ector interactions can be quantified, and the applicability of the slip boundary and
symmetric computational domain can be determined.
5. Systematic experimental investigation of the effects of the wall distance and/or the inter
element inj ector spacing on peak heat flux and combustion efficiency for a multielement
injector. Experimental studies on the selected best, or similar, multielement injector
designs and a comparison with the baseline case would provide experimental confirmation
of the results of the simple study.
224
For a good fit, R~ should be close to 1.
2.1.3.3 Prediction error sum of squares
When there are an insufficient number of data points available to test the RSA, the
prediction error sum of squares (PRESS) statistic is used to estimate the performance of the
surrogate. A residual is obtained by fitting a surrogate over the design space after dropping one
design point from the training set. The value predicted by the surrogate at that point is then
compared with the expected value. PRESS is given by
PRESS=~Jf f/ (220)
and PRESSms is given by
PRESS '=1 (221)
where j* is the value predicted by the RSA for the ith point which is excluded while generating
the RSA. If the PRESSms value is close to cra, this indicates that the RSA performs well.
2.2 Dimensionality Reduction Using Global Sensitivity Analysis
At the beginning of a design optimization, several variables are chosen as design variables
with the assumption that they are important to the optimization. However, having large numbers
of design variables can greatly increase the cost of the optimization. Section 2. 1.2 introduced the
concept of the "curse of dimensionality," meaning that as the number of design variables
increase, the number of data points required to obtain a good approximation of the response can
increase exponentially. It is of great benefit, therefore, to simplify the design problem by
identifying variables that are unimportant and removing them from the analysis. The most
efficient way of doing this is to perform a sensitivity analysis. A global sensitivity analysis can
adequately filling a hypercube design space region with points becomes cost prohibitive.
Balabanov et al.70 noted that to sample only at the vertices of a boxshaped design requires 2n
points, where n is the number of design variables. In contrast, a simplex sampled at the vertices
would require only n + 1 points. In particular, for n = 25, a box domain has a volume 1010 times
larger than a corresponding ellipsoid. Several techniques are used to improve the optimization
process and surrogate accuracy by reducing the domain over which the function is approximated.
Madsen et al.74 refined the design space to an ellipsoid shape in the shape optimization of a
diffuser to prevent the unnecessary CFD analysis of physically infeasible designs. In the
optimization of a superdetonative ram accelerator, Jeon et al.76 found that a surrogate model that
was a poor fit due to sharp curvature in the response and was able to achieve a better result by
performing a transformation on the design space, hence, smoothing out the curvature. Papila et
al.7 explored the possibility of reducing errors in regions the design space by constructing a
reduced design space on a small region of interest. Balabanov et al.70 eliminated suspected
infeasible design regions from the design space using a lowfidelity analysis prior to optimizing
using a highfidelity CFD analysis in the design of a high speed civil transport wing. Papila et
al.7 refined the design space by expanding the variable bounds to account for optimal designs
that lay on the edge of the design space in the shape optimization of a supersonic turbine.
1.4 Summary
Rocket engine component design is a complex process that can be made more
straightforward using CFDbased optimization techniques. The performance and efficiency of
key components can be improved using the SBO methods. The application of the SBO approach
to CFD problems can be challenging due to time and computational constraints. Tools including
lowfidelity analyses and design space refinement can help reduce computational expense. In
2.4.4 Method of Alternative Loss Functions
By modifying the least squares loss function, it is possible to fit several similar, but slightly
different polynomial response surfaces to the same set of data. For adequately refined design
spaces, it can be shown that the difference among these polynomials is negligible. For design
spaces that require some type of refinement, large differences can exist. This property of using
the differences in different surrogates highlight potential problems in the surrogate model or in
the data is not new." However, the proposed method has an advantage in that surrogates of the
same type can be compared requiring no change in the data set. This may enable a quantitative
measurement in the degree of refinement needed in a problem.
For a polynomial response surface the prediction at any point i is given as
9,= [b,xx, (239)
where k is the number of terms in the polynomial response surface and x,, are values from the
matrix X from Equation 23. For example, at any point i, a quadratic response surface with two
variables can be given as
f = bo + b, x, + b2x 12 b3 2 + b4 1 2 + b, x2 (240)
or simplified to
f = b + b,x, +b2x 2+b3x 3+b4x 4+b,x, (241)
The notation used for these and the following equations are adapted from Myers and
Montgomery.46 To determine the coefficients b,, the loss function to be minimized can be
defined as
L =~ fE, 72 _ 9] 72 bo C bx, (242)
=1 =1 =1 ]=1
Table 43. Reduced reaction mechanisms for hydrogenoxygen combustion.
Reaction A ((m3/gmOl)n1/s) B EA (kJ/gmol)
H2 + 02 ++ OH + OH 1.70 x10'0 0.00 2.41 x104
H2 +OH ++H20 +H 2.19x1010 0.00 2.59x103
OH +OH ++OH +H 6.02x109 0.00 5.50x102
H + O ++ OH +H 1.80 x107 1 00 4.48x103
Table 42. Selected CFD and numerical studies for shear coaxial iniectors.
Author Propel
lants
P T(K) Summary/
(MPa) parameters
compared
37.9 94.4/298, Developed
118/ 154 threephase
COmbustion
model
1.3 297 Velocity and
species
COmpanison
Grid spec. Turbulence
model
Accuracy
Liang et
al.
(1986)27
Foust et
al.
(1996)146
Cheng et
al.
(1997) 5
Ivancic
et al.
(1999)154
Lin et al.
(2005)'"'
West et
al.
(2006)156
LOX/GH2
Eddy Unknown. Cited inadequate
viscosity experimental knowledge.
kE Good species matching with
inhouse experiment.
Velocity matching degrades
downstream.
k E Good species matclung to
experiment by Foust et al.
Poor H20 species
matching.
k E Fair matching of radial OH
distribution to inhouse
experiment.
Menter Good peak heat flux
BSL prediction. Overpredicts
downstream heat flux. Low
grid dependence.
GO2/
GH 2
2D
Cartesian
101x51
Axisym.
151x81
Axisym.
1.5 x10'
elements
Axisym.
61,243
117,648
points
Axisym
structured/
hybrid
50,000 
1,000,000
points
LOX/GH2,
GO2/GH2
LOX/GH2
GO2/GH2
GO2/GH2
3.1, 117/309,
1.3 290/298
Velocity and
species
comparison
6.0 127/125 Time/length
scale
inVCStigation
5.2 767/798 Heat flux
comparison
(RCM1)
5.2 767/798 Heat flux
comparison
(RCM1)
Wilcox k
co, Menter
BSL and
SST
Good peak heat flux
prediction when adequate
grid resolution in high
gradient areas. Menter's
SST model provides
superior peak heat flux
prediction.
Thakur GO2/GH2
and
Wright
(2006)'5
5.2 767/798 Heat flux
comparison
(RCM1)
Axisym.
26,000
104,000
points
Menter Good peak heat flux
SST prediction. Overpredicts
downstream heat flux.
8.37x103
0.0
5.94x104
0.0
0.0
H+0 + + t OH + O
H + O+M ++OH +M
O+O +M +O2 +M
H+H +M ++ H2 + M
OH + H + M ++ H20 + M
1.22x1014 0.91
1.00x1010 0.00
2.55x1012 1.00
5.00x109 0.00
8.40 x10 5
2.00
qmax (MW/m2)
L (mm)
1~ 0' a494
12lr 1 4
P II. rr 1 440 495
1 ** 0 4
0 496
".~ 046
Z p 0 497
g, 10 47
0~ .1498
0 49 48 0. i 0 499
0 6 9 ~~0 5 0.0.2 0. .4 0.5 0.6 0.8 
0.2 0. a.4 0.5 0.6 PY0.8
C D
Figure 713. Kriging surrogates and merit function contours for 12 design points in Pattern 2.
Kriging fit for A) qmax and B) Lc. Merit function values for C) qmax and D) Lc. White
points are data points. Black points are points chosen using merit function analysis.
Circled black points are points chosen for simulation. Points in Pattern 1 region are
ignored.
212
8.1 Radial Turbine Efficiency and Weight Optimization
The compact radial turbine preliminary optimization identified several factors that can
improve its efficiency. The lowcost surrogate model analysis revealed that reducing the radius
of the turbine while increasing blade size has a positive impact on efficiency without impacting
the turbine's weight. The most efficient designs have higher rotational speeds and isentropic
velocity ratios. A grid sensitivity analysis revealed that the rotational speed and the velocity
ratio accounted for 97% of the variability in the rotor weight, while the turbine efficiency was
sensitive to all variables included in the analysis.
The applicability of DSR techniques was successfully demonstrated in the preliminary
optimization of a radial turbine optimization.
1. In the case where the feasible design space was unknown, response surfaces proved
invaluable in determining design constraints.
2. After the feasible design space was determined, information from a Pareto front was used
to refine the design space to improve surrogate accuracy without compromising the
performance of either objective.
For the radial turbine optimization, the infeasible points were not thrown away. Instead,
they served a useful purpose by providing the location of the bounds of the feasible design space.
For a sixvariable design space, surrogate models were necessary in extracting these boundsa
task which would have been very difficult, otherwise. In addition, the study addressed the
problem of refining the design space in the presence of multiple obj ectives. To prevent missing
the region of interest in one obj ective or the other when the design space was refined, Pareto
fronts were used to simultaneously locate the regions of interest in both obj ectives.
Finally, the radial turbine case was analyzed again using merit functions to aid the
analysis. It was discovered that the use of merit functions could improve the efficiency of the
optimization process by improving point selection during design space refinement. Using merit
220
TABLE OF CONTENTS
page
ACKNOWLEDGMENTS .............. ...............4.....
LIST OF TABLES ................ ...............8............ ....
LIST OF FIGURES .............. ...............10....
AB S TRAC T ........._. ............ ..............._ 15...
CHAPTER
1 INTRODUCTION ................. ...............17.......... ......
1.1 Thrust Chamber Characteristics .............. ...............18....
1.1.1 Rocket Engine Cycles............... ...............19.
1.1.2 Engine Reliability Issues .............. ...............20....
1.2 Current and Past Design Practices ................. ...............21........... ..
1.2. 1 Rocket Engine Inj ector Design ................. ...............22.............
1.2.2 Rocket Engine Turbine Design .............. ...............24....
1.3 CFDBased Optimization .............. .. .......... ......... ... ........2
1.3.1 Optimization Techniques for Computationally Expensive Simulations ................25
1.3.2 Design Space Refinement in CFDBased Optimization Using Surrogate
Model s............... ...............29.
1.4 Summary ................. ...............3.. 0......... ....
1.5 Goal and S cope ................. ...............3.. 1............
2 OPTIMIZ ATION FRAMEWORK ................. ...............3.. 5......... ....
2.1 Optimization Using Surrogate Models .............. ...............36....
2.1.1 Design of Experiments .............. .. ....... ...............3
2. 1.2 Surrogate Model Identifieation and Fitting .............. ...............38....
2. 1.2. 1 Polynomial response surface approximation............... .............3
2. 1.2.2 Kriging .............. ..... ... ..............4
2. 1.2.3 Radial basis neural networks ................. ...............44........... ..
2.1.3 Surrogate Model Accuracy Measures .............. ...............47....
2.1.3.1 Root mean square error ................. ...............47...
2. 1.3.2 Coefficient of multiple determination ................. ................ ......... .48
2. 1.3.3 Prediction error sum of squares ................. ........... ... ............ ...4
2.2 Dimensionality Reduction Using Global Sensitivity Analysis............... ................4
2.3 MultiObj ective Optimization Using the Pareto Optimal Front ................. ................. .52
2.4 Design Space Refinement Techniques .............. ........ .............5
2.4. 1 Design Space Reduction for Surrogate Improvement ................. ............. ......56
2.4.2 Smart Point Selection for Second Phase in Design Space Refinement..................59
2.4.3 Merit Functions for Data Selection and Reduction .............. .....................6
2.4.4 Method of Alternative Loss Functions ................ ...............65...............
Heating effects on the combustion chamber wall due to the arrangement of inj ector
elements is of prime importance. Inj ector placement can result in high local heating on the
combustion chamber wall. Figure 44 shows the effect of local heating that resulted in the
burnout of an uncooled combustion chamber. Rupe and Jaivinl36 found a positive correlation
between the temperature profile along the wall and the placement of inj ector elements. Farhangi
et al.137 inVCStigated a gasgas inj ector and measured heat flux to the combustion chamber wall
and injector face. It was found that the mixing of the propellants controlled the rate of reaction
and heat release. Farhangi et al. suggested that the injector element pattern could be arranged in
a way that moved heating away from the injector face by delaying the mixing of the propellants.
4.1.3 Combustion Chamber Effects and Considerations
Several studies have suggested that pressure has little influence on the combustion
dynamics. Branam and Mayerl38 Studied turbulent length scales by injecting cryogenic nitrogen
through a single inj ector to help in understanding cryogenic rocket propellant mixing efficiency.
It was found that changes in pressure and injection velocity had very little effect on flow
dynamics. Similarly, Calhoon et al. 134 found that the pressure effects on chamber heat flux were
very small. Quentmeyer and Roncacel39 found no change in the heat flux data with a change in
pressure when calculating the heat flux to the wall of a plug flow calorimeter chamber. In
addition, Wanhainen et al.140 found that changing the chamber pressure had no effect on
combustion stability. While, Mayer et al.141 found that pressure is relatively constant throughout
the combustion chamber, while Moonl42 asserts that even the very small pressure gradients that
exist near the inj ector exist can significantly alter velocity profiles and mixing in the combustion
chamber. The relatively constant pressure indicates that density gradients in the combustion
chamber must be due to temperature, rather than pressure. Thus, the flow can be considered
incompressible. In addition to pressure independence, Preclik et al.143 experimentally measured
3200.0
;/ . 1300.
Figure 31. Midheight static pressure (psi) contours at 122,000 rpm.
0.922
0.99
0.880
0.86"
0.84r
0.822
0.8e
0.78e
0.76s
0.74r,
100
 II!!!! c
110 120 130
103 P
Figure 32. Predicted Meanline and CFD totaltostatic efficiencies.
Unsteady approximation
unsteady linear
unsteady axisymmetric
2D unsteady adiabatic
C\I
E
Z 2
r
0 0.02 0.04 0.06
0.08 0.1
x(m)
0.12 0. 14
0.16
Figure 615. Experimental heat flux values using unsteady
mn meters.
p= constant
outlesymmetry
A
assumptions. The x values are given
H2 inlet
S02 inlet
"K77~""1
.r
~d~,
i'
Figure 616. Computational model for singleelement injector flow simulation. A) Combustion
chamber boundary conditions, B) grid closeup, and C) representative combustion
chamber section.
182
heat tansfe
Ol OXygenrich core
recirculation
heat transfer
recirculation
I
n/ shear layer
reattachmentpm
homogenous products
shear layer
Figure 42. Flame from gaseous hydrogen gaseous oxygen single element shear coaxial
inj ector.
Figure 41. Coaxial injector and combustion chamber flow zones.
V1totax= V+Vz,z (229)
where V, is the partial variance of the obj ective with respect to x, and Vz,z is the measure of the
obj ective variance that is dependent on interactions between x, and Z. Similarly, the partial
variance for Z can be defined as Vz. Therefore, the total obj ective variability can be written as
V = V, + Vz + y6,z (230)
Sobolll2 prOposed a variancebased nonparametric approach to estimate the global sensitivity
for any combination of design variables using Monte Carlo methods that is also amenable.
While Sobol used Monte Carlo simulations to conduct the global sensitivity analysis, the
expressions given above can be easily computed analytically once the RSA is available. In the
present study, the above referenced expressions are evaluated analytically using polynomial
RSAs of the objective functions. No accommodations are made for irregularshaped domains, so
the analytical treatment works best in a wellrefined design space where the assumption of a box
like domain will not introduce significant errors. Using a polynomial RSA as the function f(x),
the approximation can be decomposed as in Equation 222 and the sensitivity indices can be
obtained.
2.3 MultiObjective Optimization Using the Pareto Optimal Front
After developing a computationally inexpensive way of evaluating different designs, the
final step is to perform the actual optimization. In the case of a single objective, this requires a
simple search of design space for the minimum value of the obj ective. For two or more
objectives, additional treatment is needed. Highly correlated objectives can be combined into a
single objective function. When the objectives are conflicting in nature, there may be an infinite
number of possible solutions that will provide possible good combinations of obj ectives. These
solutions are known as Pareto optimal solutions. While there are numerous methods of solving
then the measure of energy release from a single element' s flow Hield is a good approximation of
the efficiency of the entire core. In particular, the mixing characteristics of a single inj ector can
be used to determine the element pattern and spacing needed for good mixing efficiency.
Singleelement inj ectors are often used in inj ector experimentation to approximate the
mixing effects of the full inj ector. Calhoon et al.134 tested a multiple element inj ector chamber
such that the radial spacing of the elements could be varied. In cold flow testing, it was found
that mixing increased dramatically when multiple elements were used. However, it was
determined that the element spacing only had a slight influence on the mixing efficiency. Yet, as
the distance between the elements increased, the mixing efficiency increased. This was due to
large recirculation regions that brought flow from the wellmixed far region to the near Hield.
When recirculation effects are compensated for, Calhoon et al. found that there is no significant
effect of multiple element interactions compared to single element mixing. This, in theory,
indicates that a satisfactory combustion analysis can be done using only a single injector.
4.1.2 MultiElement Injectors
A multielement inj ector face is made up of an array of inj ector elements, and usually
contains from seven to hundreds of injector elements as shown in Figure 43. Small changes in
the design of the inj ector and the pattern of elements on the inj ector face can significantly alter
the performance of the combustor. Elements must be arranged to maximize mixing and ensure
even fuel and oxidizer distribution. Gill26 found that the element diameters and diameter ratios
largely influence mixing in the combustion chamber, and that small diameters lead to overall
better performance. The type of elements need not be consistent across the entire inj ector face.
The outer elements must be chosen to help provide some wall cooling in the combustion
chamber. Gill suggests that using a coaxial type inj ector for the outer row of inj ectors provides
an ideal nearwall environment.
CHAPTER 1
INTRODUCTION
The Space Shuttle Main Engine (SSME) was developed as a reusable launch engine.
However, even with noticeable success, the complexity of the SSME drove up the maintenance
costs of the engine due to the harsh operating environment among the engine components. These
additional maintenance costs were unforeseen in the original design and analysis efforts.l For
this reason, NASA is striving to design a simpler, more reliable engine.2,3 Today, there are more
computational, diagnostic and experimental tools available to assist in streamlining the design
process than when the SSME was designed. For example, flow and combustion processes can
now be modeled using advanced computational fluid dynamics (CFD) tools.4,5 Using CFD tools
can help narrow down the design of an engine component before a prototype is built and tested.
Often, designs are based on past experience, intuition, and empirical calculations. Without a
comprehensive and methodical process, the designs may not be satisfactorily refined without
tedious and inefficient trialanderror exercises. Optimization tools can be used to guide the
selection of design parameters and highlight the predicted best designs and tradeoff designs. In
this study, CFD and optimization tools are used to develop further capabilities in the design and
analysis of a radial turbine and an inj ector: critical liquid rocket engine components.
Current design practices for inj ector and turbine designs include past experience, intuition,
and empirical calculations. It is difficult to fully analyze, design, and optimize the combustion
chamber using experimental information alone. Because of measurement difficulties in
combustion chambers due to high temperatures and pressures, much is still too poorly
understood of the flow dynamics to accurately predict the response of future inj ector designs.
Using CFD can be used as a tool to help analyze and understand the flow dynamics within the
combustion chamber to fill in knowledge gaps left by the inj ector experiments. This research
Twall,hnear qlerS + 63
In the experimental documentation, the steady and unsteady heat fluxes are calculated using
Equations 61 and 62, respectively, and are shown in Figure 611.
Equations 61 through 63 assume that the heat flow is onedimensional.179 For a square
duct, the onedimensional linear assumption may not be correct. The actual equation would need
to be determined through a numerical solution of the twodimensional heat conduction equation.
The correct heat flux equation would change the experimentally documented values of heat flux
at the wall. However, because at each cross section only two temperature measurements are
made, there is not enough data to do a full twodimensional analysis of heat transfer through the
combustion chamber wall. Additional temperature measurements taken within the wall would be
required for an accurate estimation of wall heat flux. Therefore, a onedimensional axisymmetric
approximation was included in addition to the onedimensional linear approximation.
If it is assumed that the isotherms within the combustion chamber wall are axisymmetric, a
second onedimensional assumption can be conducted. The geometry of the isotherms is given
in Figure 612. This is reasonable due to the approximate axisymmetry of the temperature
contours within the combustion chamber wall, and the high thermal conductivity of the copper.
For 1D axisymmetric heat conduction, the steady state heat flux approximation is given by
kAT
q",,~s (64)
where rwan is the distance from the center of the combustion chamber to the combustion chamber
wall, and r, and ro are the distances from the center of the combustion chamber to the inner and
outer thermocouples, respectively, as shown in Figure 612. Using this equation, the heat flux at
the wall centerline can be determined based on the 1D heat conduction equation, giving
~~
: ~C~i I

III
~ 1; Ii~i ...i I
I I Ir
Figure 74. Oxygen isosurfaces and hydrogen contours. A) Baseline case (case 1). B) Worst
overall case (case 12). C) Case with the lowest peak heat flux (case 2). D) Best
overall case (case 6). The combustion chamber wall is at the top of each figure.
Approximately half of the designs showed solutions similar to A) and B), while the
remainder was similar to C) and D). Solutions are mirrored across the :plane.
25 30 35 40
ii
Figure 75. Hydrogen contours and streamlines. A) Case 12 and B) Case 6. The wall is located at
the top of each plot. Case 6 shows the persistence of a gaseous hydrogen layer near
the combustion chamber wall.
207
4.1.4 Review of Select CFD Modeling and Validation Studies
The maj ority of CFD validation efforts to date have concentrated on single element coaxial
injectors. Work has focused on LOX/GH2 Or GO2/GH2 COmbustion. Most efforts focused on
matching species or velocity profies. Only recently have efforts been in place to work towards
using CFD as a combustor design tool. These recent efforts focus on accurately modeling the
heat characteristics of combustors.
Liang et al.27 Sought to improve the effects of multiphase modeling that included liquid,
gas, and liquid droplets. A 2D axisymmetric model was used along with a chemical model
consisting of a 9equation kinetic model along with 4 equilibrium equations. The equilibrium
equations were included to help anchor the flame during the computation by allowing for
instantaneous reactions. In this case, equilibrium reactions were necessary due to the coarseness
of the grid used. The turbulence model used is the eddy viscosity model. In this case, no effort
is made to couple combustion with the turbulence equations. For the simulation of a gasgas
inj ector, a mixture fraction of unity was used for the propellants, and the combustion chamber
was originally filled with oxygen. An artificial ignition region was placed near the oxygen post
tip. The computation was run for a physical equivalent of 10ms. An attempt was made to use
the 2D model to simulate LOX/GH2 multielement inj ector flow by changing the walls from a
noslip to a slip boundary condition. Local peak temperature is around 2000K, and the average
temperature is around 1500K. Liang et al. mentioned that, based on the CFD results, atomization
might be a ratecontrolling factor, but there was insufficient experimental data to confirm this
fact. Liang et al. notes that grid resolution has a significant effect on flame ignition and flame
steadiness, but no grid sensitivity study is conducted to explore the effects. Temperature,
velocity, and mass fraction contours were obtained, but no comparison was made to
experimental results, so the accuracy of the simulation is unknown.
whereJ; is the actual value and j; is the predicted value. For Ns design points, the root mean
square (rms) error a is given by
r 1= (217)
SN,
A small rms error indicates a good fit. The rms error is a standard measure of fit that may be
used to compare the performances of surrogates of different types.
The adjusted rms error oa is used to adjust the rms error based on the number of parameters
in a polynomial RSA and is given by
cr (218)
a (N q)
where Ns is the number of data points, and q is the number of terms in the polynomial
approximation. A polynomial with a large number of terms is penalized and deemed less
desirable for a polynomial approximation with fewer terms for the same number of data points.
For a good fit, oa should be small compared to the data value ranges.
For Kriging, the standard error measurement is given by the estimated standard deviation
of the response given by Equation 211i. The square root of the estimated standard deviation of
the response can be compared to the rms error of polynomial RSAs.
2.1.3.2 Coefficient of multiple determination
The, adj uLuse coeficientII of mIultle determIIinaionLV R)4d defines the prediction capability of
the polynomial RSA as
R2 = 1 %cr=1a ( 1)\2 Where f= ,= (219)
Author Application Method highlighted Key results
9
Chapter 7, CFDbased optimization techniques are applied to improve the design process of a
multielement rocket inj ector. In this case, the optimization techniques must be applied in the
context of a CFDbased design process where a limited number of function evaluations are
available. Chapter 8 concludes the effort.
Table 11. Summary of CFDbased design optimization applications with highlighted surrogate
based optimization techniques.
Shyy et al.
Simpson et al."0'
Dornberger et
al.59
Rodriguez62
Vaidyanathan et
al.63
Han et al.64
Keane65
Alexandrov et
al.66
Knill et al.6
Balabanov et
al. '
Madsen et al. 4
Jeon et al. 6
Papila et al. 7
Balabanov et
al. 0
Papila et al. 8
Rocket engine
injector
Aerospike nozzle
Turbomachinery
designs
Jet configuration
Rocket engine
injector
Multiblade
fan/scroll
system
Transonic wing
Wing design
High speed civil
transport wing
High speed civil
transport wing
Diffuser
Ram accelerator
Supersonic turbine
Civil transport
wing
Supersonic turbine
Surrogate comparison
Surrogate comparison,
design space
selection
Surrogate comparison
Design space selection
Design space selection
Lowfidelity
optimization
Multifidelity
optimization
Multifidelity
Optimization
Multifidelity
optimization
Multifidelity analysis
with surrogate
correction
Design space
refinement
Design space
refinement
Design space
refinement
design space
refinement
Polynomial approximation compared to neural
network surrogate model.
Kriging model compared to quadratic response
surface. Simple design variable ranges
selected using geometry.
Multidisciplinary optimization using response
surfaces and neural networks.
Optimization using simple geometrybased
bounds on design variables
Design variables chosen from variations of
existing design
Optimized lowfidelity 2D model to predict for
highfidelity 3D model
Supplemented highfidelity CFD data with low
fidelity empirical model
Varied mesh density for variable fidelity models
Saved 255 CPU hours using lowfidelity
supplemental data
Used coarse and fine finite element models as
low and highfidelity models
Removed infeasible design space regions to
prevent unnecessary CFD analyses
Performed transformation on design space for
better response surface fit
Refine design space to improve regions with
high error
Used lowfidelity model to refine design space
before optimizing using highfidelity model
Design space expansion to explore optimal
designs at edge of design space
Design space
refinement
MF(Lo)
MF(qmax)
1.2 E aT 1. 0 4975
0 45
1. 1.1 CtI0 498
1 0 46
0 4985
S0. 0 47 'z 0.
.0 499
0 48 0 5
5 0 49 0 e a 5 049
0 6 0 5 0.6 g 0 5
0.2 0.4 6 0.11.2 0.2 0.4 (6 0.e 1 1.2
A B
Figure 710. Merit function (M~F2) COntours for A) qmax and B) Lc. Dark blue indicates regions
of lowest uncertainty. White points are data points. Black points are points chosen
using merit function analysis. Points circled by dashed lines indicate best tradeoff
points. Circled black points are new points chosen for simulation.
Pp (MW/m2) ~Le (mm) C
50 100
*0 1 2
1 1 O45 1 1~
1 ~40 1 0 9
z 35 z 9
25 8
.~~~~0 _, ; I. J .. : 0 811 28
A 'B
Figure 711. Kriging surrogates based on initial 16 design points. A) qmax and B) Lc. White
points are data points. Black points are points chosen using merit function analysis.
Points circled by dashed lines indicate best tradeoff points. Circled black points are
new points chosen for simulation.
210
velocity change and the streamline reattachment points are clearly apparent in the dips in the y+
values shown in Figure 622B. The effect of the velocity change is echoed in the temperature
values near the wall (y+ ~ 2.5) as seen in Figure 623A.
The mixing action of the shear layer transports heat from the flame to the wall. The point
of maximum heat flux occurs at 63 mm. This corresponds to the point of maximum eddy
conductivity that also occurs at 63 mm from the inj ector face, shown in Figure 623B. Similar to
the experimental results, the peak wall temperature does not occur in the same location as the
maximum heat flux. On the other hand, the maximum temperature along the wall occurs at 72
mm, but is essentially constant between 64 mm and 79 mm. This is immediately upstream of the
reattachment point which is 80 mm from the inj ector face. The disconnection between the peak
heat flux and the peak wall temperature can be related to the ratio of eddy viscosity to laminar
viscosity. The wall heat flux profile corresponds to the eddy viscosity, and similarly the thermal
conductivity. The temperature profile, on the other hand corresponds to the laminar, or molecular
viscosity.
6.5.4 Species Concentrations
From the temperature profile given in Figure 618 at y = 0, it can be seen that much of the
reaction occurs before 72 mm. This is also evident in the mass fraction contours in Figure 624.
The reaction proceeds beyond that at a much slower rate as the system moves toward
equilibrium. The species concentration through the flame center is shown in Figure 625.
Figure 625 shows the diffusion of hydrogen towards the centerline between 0 and 22mm. The
mole fraction of H2 inCreaSes at y = 0 due to the consumption of Oz. The y = 2 profile is located
just beyond the H2 inlet, so the increase in H2 reflects the spreading of the H2 Stream. Because
the reaction is fuel rich, the hydrogen is not completely consumed. There is almost the almost
complete disappearance of gaseous oxygen by approximately 53mm. This is also consistent with
w (N = Rtan(73)
r()= tN 2%)
where R is the distance from the center of the injector to the center of an outer element inj ector
element and D is the inj ector element diameter.
The baseline inj ector element geometry is based on the experimental test case RCM1 from
Section 4. 1.4, while the overall inj ector and element spacing is based on the Integrated
Powerhead Demonstratorl59 (IPD) main inj ector. For this inj ector, the maximum number or
inj ector elements N;;; allowable in the outer row is 101. The flow conditions and inj ector
geometry for the baseline inj ector are given in Table 71. The inlet flow represents incompletely
burned flow from upstream preburners, so both the fuel and oxidizer contain some water as a
constituent.
7.3 Feasible Design Space Study
The design points based on r* and N* are shown in Figure 72. The first seven design
points were selected based on a preliminary design sensitivity study. Point 16 lies at the center of
the design space. Latin Hypercube Sampling was then used to select the seven remaining design
points within the variable ranges.
Figure 73 shows the values of the obj ectives at each design point. It can be seen that the
mean heat flux across the combustion chamber wall stays relatively constant across all of the
design points. The maximum heat flux, however, varies depending on the combustor geometry.
The design points have an average mean heat flux of 8.2 MW/m2 and an average maximum heat
flux of 37.3 MW/m2. In general, as the maximum heat flux becomes lower, the combustion
length is longer, indicating slower combustion. For design purposes, there would need to be
some balance between the maximum heat flux and combustion length. From Figure 73B, it can
be seen that case 6 shows good balance between minimizing the peak heat flux and minimizing
If the amount of swirl generated in the pipe is small, it can be assumed that there will be
negligible swirl at the exit of the hydrogen inlet.
4.3 Simplified Analysis of GO2/GH2 COmbusting Flow
As shown earlier, several factors are involved in inj ector and injection pattern design that
are of interest in this research effort:
* A single inj ector can be analyzed as an approximation to full injector analysis
* The outer injector elements must be chose such that cooling is enhanced at the combustion
chamber wall
* Inj ector placement can result in high local heating
* Pressure effects on combustion chamber wall heat flux are very small
* Wall heat fluxes are largely independent of the mixture ratio
* A sharp increase in heat flux occurs consistently at a certain distance away from the
inj ector face
* Analysis of the flow upstream of the inj section may be necessary to determine the effect on
combusting flow
A preliminary combustion analysis can give some insight into the nature of H2/02 reacting flow.
In particular, the relationship between temperature, mass flow rate, pressure, heat flow rate, and
equivalence ratio can be explored using a simplified version of the combustion chamber.
A system can be considered wherein hydrogen and oxygen are inj ected such that the molar
flow rate of hydrogen exceeds that of oxygen (Figure 49). The reaction is given as
20H2 + 02 4 2H2 + 2(#1)H2 (445)
and the following assumptions are made:
* Constant pressure
* Complete combustion with no dissociation
* Fuelrich combustion ($> 1)
* Steadystate
* All combustion products exit at the same temperature
tllc. I!!!!!!c
E
100 110 120
103 RPM
130 140 150
Figure 33. Predicted Meanline and CFD turbine work.
Rse/Rs,, > 0.85
0.95
0.9 lAN2 < 850
0.85
0.8
0.7 5
07
0.65
0 6
0,55
.5~ 0 55
U/C isen
Infeasible
Tip Spd
> 2500
06 063 ().65
Figure 34. Feasible region and location of three constraints. Three of five constraints are
automatically satisfied by the range reduction of two design variables.
140. Wanhainen, J. P., Feiler, C. E., and Morgan, C. J., "Effect of Chamber Pressure, Flow Per
Element, and Contraction Ratio on AcousticMode Instability in HydrogenOxygen
Rockets," NASA TN D4733, NASA, Washington, D.C., August 1968.
141. Mayer, W., Telaar, J., Branam, R., Schneider, G., and Hussong, J., "Raman Measurements
of Cryogenic Inj section at Supercritical Pressure," Heat and Ma~ss Transfer, Vol. 39, pp.
709 719.
142. Moon, L. F., "Pressure and Velocity in a Developing Coaxial Jet," A1AA Journal, Vol. 14,
No. 1, pp. 4349.
143. Preclik, D., Wiedmann, D., Oechslein, W., and Kretschmer, J., "Cryogenic Rocket
Calorimeter Chamber Experiments and Heat Transfer Simulations," AIAA Paper 1998
3440, 1998.
144. Conley, C.A., "HighPressure GO2/GH2 COmbustion Chamber Dynamics," MS thesis,
University of Florida, 2006.
145. Reed, B. D., "Small Hydrogen/Oxygen Rocket Flowfield Behavior from Heat Flux
Measurements," AIAA Paper 19932162, 1993.
146. Foust, M. J., Deshpande, M., Pal, S., Ni, T., Merkle, C. L., and Santoro, R. J.,
"Experimental and Analytical Characterization of a Shear Coaxial Combusting GO2/GH2
Flowfield," AIAA Paper 19960646, 1996.
147. Morren, S. H., Myers, R. M., Benko, S. E., Arrington, L. A., and Reed, B. R., "A
Laboratory Model of a Hydrogen/Oxygen Engine for Combustion and Nozzle Studies,"
NASA TM106281, 1993.
148. de Groot, W. A., McGuire, T. J., and Schneider, S. J., "Qualitative Flow Visualization of a
110N Hydrogen/Oxygen Laboratory Model Thruster," AIAA Paper 19972847, 1997.
149. Strakey, P., Talley, D., Tseng, L., and Miner, K., "The Effects of LOX Post Biasing on
SSME Inj ector Wall Compatibility," AIAA Paper 19992888, 1999.
150. Smith, T., Klem, M., Breisacher, K., Farhangi, S., and Sutton, R., "Experimental
Evaluation of a Subscale Gaseous Hydrogen/Gaseous Oxygen Coaxial Rocket Injector,"
NASA TM 2002211982, 2002.
151. Cheng, G. C., Anderson, P. G., and Farmer, R. C., "Development of CFD Model for
Simulating Gas/Liquid Inj ectors in Rocket Engine Design, AIAA Paper 19973228, 1997.
152. Schley, C., Hagemann, G., Tucker, K. P., Venkateswaran, S., and Merkle, C. L.,
"Comparison of Computational Codes for Modeling HydrogenOxygen Inj ectors," AIAA
Paper 19973302, 1997.
153. Schley, C.A., Hagemann, G., Golovitchev, V., "Comparison of High Pressure H2/02
Rocket Model Engine Reference Simulations," AIAA Paper 19952429, 1995.
235
Thermocouples
I I wall
 Steadystate approximation
Unsteady approximation
C\I
E
~ 2.5
r

0 0.02 0.04 0.06 0.08
x(m)
0. 1 0. 12 0.14 0. 16
Figure 611. Estimated wall heat flux using linear steadystate and unsteady approximations.
Figure 612. 1D axisymmetric assumption for heat conduction through combustion chamber
wall.
180
qmax (MW/m2)
12
11
z
09
08
04 05 06 07 08 09 1 11 12 04 05 06 07 08 09 1 11 12
A B
Figure 714. Kriging fits for all 15 design points from Pattern 1. A) qmax and B) Lc.
qmax (MW/m2) Le (mm)
12
1 1
z 9a
32
30
28
26z
12 220
A B
Figure 715. Kriging fits for all 12 design points from Pattern 2. A) qmax and B) Lc.
213
L (mm)
O

PAGE 1
1 CFDBASED SURROGATE MODELING OF LIQUID ROCKET ENGINE COMPONENTS VIA DESIGN SPACE REFINEMENT AND SENSITIVITY ASSESSMENT By YOLANDA MACK A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2007
PAGE 2
2 2007 Yolanda Mack
PAGE 3
3 To my son, Trevor.
PAGE 4
4 ACKNOWLEDGMENTS I would like to thank the peopl e in my life who have supported me during my graduate studies. In particular, I would like to thank my advisor, Dr. Wei Shyy, for his support and for pushing me to strive for excellence. I also thank Dr. Raphael Haftka for his guidance through the years on my research. I thank the member s of my supervisory committee members, Dr. Corin Segal, Dr. William Lear, Dr. Andreas Hase lbacher, and Dr. Don Slinn, for reviewing my work and offering suggestions during my studies. I thank Dr. Nestor Queipo for inviting me to explore new methods and new ideas that have been invaluable in my work. I thank Dr. Siddharth Thakur and Dr. Jeffrey Wright for their troubleshooting assistance over the years. I thank Mr. Kevin Tucker and the others that I ha ve collaborated with at NASA Marshall Space Flight Center along with the Institute for Future Space Transport under the Constellation University Institute Project for providing the motivations for my work as well as financial assistance. I would also like to thank Zonta International and the South East Alliance for Graduate Education and the Professoriate for their financial support and recognition. I thank Antoin Baker for his love and suppor t through any difficulty as we pursued our graduate degrees. I would like to thank my father, Calvin Mack, for his gentle guidance, and my mother, Jacqueline Mack, to whom I express my deepest gratitude for he r unconditional love and assistance over the years. I thank my sisters, Brandee and Cailah Mack, and other members of my family for their undying support and for forever believing in me.
PAGE 5
5 TABLE OF CONTENTS page ACKNOWLEDGMENTS...............................................................................................................4 LIST OF TABLES................................................................................................................. ..........8 LIST OF FIGURES................................................................................................................ .......10 ABSTRACT....................................................................................................................... ............15 CHAPTER 1 INTRODUCTION................................................................................................................. .17 1.1 Thrust Chamber Characteristics.......................................................................................18 1.1.1 Rocket Engine Cycles.............................................................................................19 1.1.2 Engine Reliability Issues........................................................................................20 1.2 Current and Past Design Practices....................................................................................21 1.2.1 Rocket Engine Injector Design...............................................................................22 1.2.2 Rocket Engine Turbine Design..............................................................................24 1.3 CFDBased Optimization.................................................................................................25 1.3.1 Optimization Techniques for Comput ationally Expensive Simulations................25 1.3.2 Design Space Refinement in CFDBa sed Optimization Using Surrogate Models......................................................................................................................... .29 1.4 Summary.................................................................................................................... .......30 1.5 Goal and Scope............................................................................................................. ....31 2 OPTIMIZATION FRAMEWORK.........................................................................................35 2.1 Optimization Using Surrogate Models.............................................................................36 2.1.1 Design of Experiments...........................................................................................36 2.1.2 Surrogate Model Identif ication and Fitting............................................................38 2.1.2.1 Polynomial response surface approximation................................................39 2.1.2.2 Kriging.........................................................................................................41 2.1.2.3 Radial basis neural networks........................................................................44 2.1.3 Surrogate Model Accuracy Measures....................................................................47 2.1.3.1 Root mean square error................................................................................47 2.1.3.2 Coefficient of multiple determination..........................................................48 2.1.3.3 Prediction error sum of squares....................................................................49 2.2 Dimensionality Reduction Using Global Sensitivity Analysis.........................................49 2.3 MultiObjective Optimization Using the Pareto Optimal Front.......................................52 2.4 Design Space Refinement Techniques.............................................................................54 2.4.1 Design Space Reduction for Surrogate Improvement............................................56 2.4.2 Smart Point Selection for Second Phase in Design Space Refinement..................59 2.4.3 Merit Functions for Data Selection and Reduction................................................61 2.4.4 Method of Alternative Loss Functions...................................................................65
PAGE 6
6 3 RADIAL TURBINE OPTIMIZATION.................................................................................75 3.1 Introduction............................................................................................................... ........76 3.2 Problem Description........................................................................................................ .79 3.2.1 Verification Study..................................................................................................79 3.2.2 Optimization Procedure..........................................................................................81 3.3 Results and Discussion..................................................................................................... 81 3.3.1 Phase 1: Initial Design of Experime nts and Construction of Constraint Surrogates....................................................................................................................8 1 3.3.2 Phase 2: Design Space Refinement........................................................................84 3.3.3 Phase 3: Construction of the Pareto Front and Validation of Response Surfaces....................................................................................................................... .86 3.3.4 Phase 4: Global Sensitivity Analysis and Dimensionality Reduction Check.........87 3.4 Merit Function Analysis...................................................................................................8 8 3.4.1 Data Point Selection and Analysis.........................................................................88 3.4.2 Merit Function Comparison Results.......................................................................89 3.5 Conclusion................................................................................................................. .......91 4 MODELING OF INJECTOR FLOWS................................................................................105 4.1 Literature Review.......................................................................................................... .106 4.1.1 SingleElement Injectors......................................................................................106 4.1.2 MultiElement Injectors........................................................................................107 4.1.3 Combustion Chamber Effects and Considerations...............................................108 4.1.4 Review of Select CFD Mode ling and Validation Studies....................................112 4.2 Turbulent Combustion Model.........................................................................................117 4.2.1 Reacting Flow Equations......................................................................................117 4.2.2 Turbulent Flow Modeling.....................................................................................120 4.2.3 Chemical Kinetics................................................................................................124 4.2.4 Generation and Decay of Swirl............................................................................125 4.3 Simplified Analysis of GO2/GH2 Combusting Flow......................................................127 5 SURROGATE MODELING OF MIXING DYNAMICS....................................................137 5.1 Introduction............................................................................................................... ......137 5.2 Bluff Body Fl ow Analysis..............................................................................................138 5.2.1 Geometric Description and Computational Domain............................................139 5.2.2 Objective Functions and Design of Experiments.................................................140 5.3 Results and Discussion...................................................................................................14 2 5.3.1 CFD Solution Analysis.........................................................................................142 5.3.2 Surrogate Model Results......................................................................................143 5.3.3 Analysis of Extreme Designs...............................................................................145 5.3.4 Design Space Exploration....................................................................................147 5.4 Conclusions................................................................................................................ .....148 6 INJECTOR FLOW MODELING.........................................................................................157
PAGE 7
7 6.1 Introduction............................................................................................................... ......157 6.2 Experimental Setup......................................................................................................... 158 6.3 Upstream Injector Flow Analysis...................................................................................159 6.3.1 Problem Description.............................................................................................159 6.3.2 Results and Discussion.........................................................................................160 6.3.3 Conclusion............................................................................................................162 6.4 Experimental Results and Analysis................................................................................163 6.5 Injector Flow Modeling Investigation............................................................................166 6.5.1 CFD Model Setup.................................................................................................167 6.5.2 CFD Results and Experimental Comparison of Heat Flux..................................168 6.5.3 Heat Transfer Characterization.............................................................................169 6.5.4 Species Concentrations.........................................................................................170 6.6 Grid Sensitiv ity Study..................................................................................................... 171 6.7 Conclusion................................................................................................................. .....173 7 MULTIELEMENT INJECTOR FLOW MODELING AND ELEMENT SPACING EFFECTS........................................................................................................................ ......191 7.1 Introduction............................................................................................................... ......191 7.2 Problem SetUp............................................................................................................. .193 7.3 Feasible Design Space Study..........................................................................................195 7.4 Design Space Refinement...............................................................................................199 7.5 Conclusion................................................................................................................. .....203 8 CONCLUSIONS.................................................................................................................. 219 8.1 Radial Turbine Efficiency and Weight Optimization.....................................................220 8.2 Bluff Body Mixing Dynamics........................................................................................221 8.3 SingleElement Injector Flow Modeling........................................................................222 8.4 MultiElement Injector Flow Modeling..........................................................................222 8.5 Future Work................................................................................................................ ....224 REFERENCE LIST................................................................................................................. ....225 BIOGRAPHICAL SKETCH.......................................................................................................238
PAGE 8
8 LIST OF TABLES Table page 11 Summary of CFDbased desi gn optimization applications with highlighted surrogatebased optimization techniques...........................................................................................32 21 Summary of DOE and surr ogate modeling references......................................................67 22 Design space refinement (DSR) techniques with their applications and key results.........68 31 Variable names and descriptions.......................................................................................92 32 Response surface fit statistics before (f easible DS) and after (reasonable DS) design space reduction................................................................................................................ ...93 33 Original and final design variable ranges after cons traint application and design space reduction................................................................................................................ ...93 34 Baseline and optimum design comparison........................................................................93 41 Selected injector experimental studies.............................................................................128 42 Selected CFD and numerical stud ies for shear coaxial injectors.....................................129 43 Reduced reaction mechanisms for hydrogenoxygen combustion..................................129 51 Number of grid points used in various grid resolutions...................................................149 52 Data statistics in the grid comparison of the CFD data...................................................149 53 Comparison of cubic response surface coe fficients and response surface statistics........150 54 Comparison of radialbasis neural network parameters and statistics.............................150 55 RMS error comparison for response surf ace and radial basis neural network................150 56 Total pressure loss coefficient and mixi ng index for extreme and two regular designs for multiple grids............................................................................................................. .151 57 Total pressure loss coefficient and mi xing index for designs in the immediate vicinity of Case 1 using Grid 3........................................................................................151 61 Flow regime description..................................................................................................17 4 62 Effect of grid resolution on wa ll heat flux and combustion length..................................175 63 Flow conditions............................................................................................................ ....175
PAGE 9
9 71 Flow conditions and baseline combustor geometry for parametric evaluation...............204 72 Kriging PRESSrms error statistics for each design space iteration...................................205
PAGE 10
10 LIST OF FIGURES Figure page 11 Rocket engine cycles....................................................................................................... ...33 12 SSME thrust chamber component diagram.......................................................................33 13 SSME component reliability data......................................................................................34 14 Surrogatebased optimizati on using multifidelity data.....................................................34 21 Optimization framework flowchart....................................................................................69 22 DOEs for noisereducing surrogate models.......................................................................70 23 Latin Hypercube Sampling................................................................................................70 24 Design space windowing...................................................................................................71 25 Smart point selection...................................................................................................... ....71 26 Depiction of the merit function rank assignment for a given cluster.................................72 27 The effect of varying values of p on the loss function shape.............................................72 28 Variation in SSE with p for two different responses.........................................................73 29 Pareto fronts for RSAs constr ucted with varying values of p ............................................73 210 Absolute percent difference in th e area under the Pareto front curves..............................74 31 Midheight static pressure (psi) contours at 122,000 rpm.................................................94 32 Predicted Meanline and CFD to taltostatic efficiencies...................................................94 33 Predicted Meanline and CFD turbine work.......................................................................95 34 Feasible region and location of three constraints...............................................................95 35 Constraint surface for Cx2/Utip = 0.2................................................................................96 36 Constraint surfaces for 1 = 0 and 1 = 40.........................................................................96 37 Region of interest in function space...................................................................................97 38 Error between RSA and actual data point..........................................................................97 39 Pareto fronts for p = 1 through 5 for second data set.........................................................98
PAGE 11
11 310 Pareto fronts for p = 1 through 5 for third data set............................................................99 311 Pareto Front with validation data.......................................................................................99 312 Variation in design vari ables along Pareto Front.............................................................100 313 Global sensitivity analysis............................................................................................... 100 314 Data points predicted by validated Pareto front compared with the predicted values using six Kriging models based on 20 selected data points.............................................101 315 Absolute error distribution for points along Pareto front.................................................101 316 Average mean error distribution over 100 clusters..........................................................102 317 Median mean error over 100 clusters...............................................................................103 318 Average maximum error di stribution over 100 clusters..................................................103 319 Median maximum error over 100 clusters.......................................................................104 41 Coaxial injector and com bustion chamber flow zones....................................................130 42 Flame from gaseous hydrogen Â– gase ous oxygen single element shear coaxial injector....................................................................................................................... ......130 43 Multielement injectors.................................................................................................... 131 44 Wall burnout in an uncooled combustion chamber.........................................................132 45 Test case RCM1 injector................................................................................................132 46 Temperature contours for a single element injector........................................................133 47 CFD heat flux results as compared to RCM1 experimental test case............................134 48 Multielement injector simulations..................................................................................135 49 Fuel rich hydrogen and oxygen reaction with heat release..............................................136 51 Modified FCCD.............................................................................................................. .151 52 Bluff body geometry........................................................................................................ 151 53 Computational domain fo r trapezoidal bluff body...........................................................152 54 Computational grid for trapezoidal bluff body................................................................152 55 Bluff body streamlines and vorticity contours.................................................................152
PAGE 12
12 57 Comparison of response surface (top row) and radial basis neural network (bottom row) prediction contours for to tal pressure loss coefficient.............................................154 58 Comparison of response surface (top row) and radial basis neural network (bottom row) prediction contours for mixi ng index (including extreme cases)............................154 59 Comparison of response surface (top row) and radial basis neural network (bottom row) prediction contours for mixi ng index (excluding extreme cases)...........................155 510 Variation in objective vari ables with grid refinement.....................................................155 511 Difference in predicted mixing index va lues from response surface (top row) and radial basis neural network (bottom row) prediction cont ours constructed with and without extreme cases......................................................................................................156 512 Comparison of response surface and radial basis neural network prediction contours for mixing index at B = 0 and h = 0.............................................................................156 61 Blanching and cracking of combustion chamber wall due to local heating near injector elements.............................................................................................................. 175 62 Hydrogen flow geometry.................................................................................................176 63 Hydrogen inlet mesh........................................................................................................ 176 64 Z vorticity contours..........................................................................................................17 7 65 Swirl number at each x location.......................................................................................177 66 Average axial velocity u and average tangential velocity v with increasing x ...............178 67 Reynolds number profiles................................................................................................178 68 Nondimensional pressure as a function of x ...................................................................178 69 Hydrogen inlet flow profiles............................................................................................179 610 Combustion chamber crosssectional geometry and thermocouple locations.................179 611 Estimated wall heat flux using linear steadystate and unsteady approximations...........180 612 1D axisymmetric assumption for heat conduction through combustion chamber wall.180 613 Estimated wall temperatures using li near and axisymmetric approximations.................181 614 Temperature (K) contours for 2D unsteady heat conduction calculations.....................181 615 Experimental heat flux valu es using unsteady assumptions............................................182
PAGE 13
13 616 Computational model for singleel ement injector flow simulation.................................182 617 Velocity contours vx(m/s) and streamlines.......................................................................183 618 Temperature (K) contours................................................................................................18 3 619 CFD heat flux values as compared to experimental heat flux approximations...............183 620 Wall heat transfer and e ddy conductivity contour plots..................................................184 621 Streamlines and temperature contours at plane z = 0.......................................................184 622 Heat flux and y+ profiles along combustion chamber wall.............................................185 623 Temperature and eddy conductivity profiles at various y locations on plane z = 0.........185 624 Mass fraction contours for select species.........................................................................186 625 Mole fractions for all species along combustion chamber centerline ( y = 0, z = 0)........187 626 Select species mo le fraction profiles................................................................................188 627 Sample grid a nd boundary conditions..............................................................................189 628 Computational grid along symmetric boundary..............................................................189 629 Wall heat flux and y+ values for select grids..................................................................189 630 Comparison of temperature (K) co ntours for grids with 23,907, 31,184, 72,239, and 103,628 points, top to bottom, respectively.....................................................................190 71 Injector element subsection..............................................................................................20 5 72 Design points selected for design space sensitivity study...............................................206 73 Effect of hydrogen mass flow rate on objectives.............................................................206 74 Oxygen isosurfaces and hydrogen contours...................................................................207 75 Hydrogen contours and streamlines.................................................................................207 76 Maximum heat flux for a changing radial distance r *.....................................................208 77 Heat flux distribution..................................................................................................... ..208 78 Maximum heat flux as a function of aspect ratio.............................................................209 79 Design points in function and design space.....................................................................209 710 Merit function ( MF2) contours for...................................................................................210
PAGE 14
14 711 Kriging surrogates based on initial 16 design points.......................................................210 712 Kriging surrogates and merit function c ontours for 12 design points in Pattern 1..........211 713 Kriging surrogates and merit function c ontours for 12 design points in Pattern 2..........212 714 Kriging fits for all 15 design points from Pattern 1.........................................................213 715 Kriging fits for all 12 design points from Pattern 2.........................................................213 716 Pareto front based on original 16 data points (dotted line) an d with newly added points (solid line)............................................................................................................ .214 717 Approximate division in the design space between the two pattern s based on A) peak heat flux...................................................................................................................... .....215 718 Variation in flow streamlines a nd hydrogen contours in design space............................215 719 Location of best tradeoff points in each pattern group in design space.........................216 720 Injector spacing for select ed best tradeoff design point.................................................216 721 Predicted heat flux profiles for baseline case (case 1), best tradeoff from Pattern 1 (case 19), and best tradeo ff from Pattern 2 (case 18).....................................................217 722 Heat flux profiles for different grid re solutions. The coarser grid was used to construct the surrogate model..........................................................................................218 723 Peak heat flux and combustion length as a function of hydrogen mass flow rate...........218
PAGE 15
15 Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy CFDBASED SURROGATE MODELING OF LIQUID ROCKET ENGINE COMPONENTS VIA DESIGN SPACE REFINEMENT AND SENSITIVITY ASSESSMENT By Yolanda Mack August 2007 Chair: Wei Shyy Cochair: Raphael Haftka Major: Aerospace Engineering Computational fluid dynamics (CFD) can be us ed to improve the design and optimization of rocket engine components that traditi onally rely on empirical calculations and limited experimentation. CFD baseddesign optimizatio n can be made computationally affordable through the use of surrogate modeling which can th en facilitate additional parameter sensitivity assessments. The present study investigates su rrogatebased adaptive design space refinement (DSR) using estimates of surrogate uncertainty to probe the CFD analyses and to perform sensitivity assessments for complex fluid phys ics associated with liquid rocket engine components. Three studies were conducted. First, a surrogatebased pr eliminary design optimization was conducted to improve the efficiency of a comp act radial turbine for an expander cycle rocket engine while maintaining low weight. Design space refinement was used to identify function constraints and to obtain a high accuracy surrogate model in the region of interest. A merit function formulation for multiobjective design point selection reduced the number of design points by an order of magnitude while mainta ining good surrogate accuracy among the best tradeoff points. Second, bluff bodyinduced flow was investigated to identify the physics and
PAGE 16
16 surrogate modeling issues relate d to the flowÂ’s mixing dynamics. Multiple surrogates and DSR were instrumental in identifying designs for which the CFD model was deficient and to help to pinpoint the nature of the defi ciency. Next, a threedimensional computational model was developed to explore the wa ll heat transfer of a GO2/GH2 shear coaxial single element injector. The interactions between turbulen t recirculating flow structures chemical kinetics, and heat transfer are highlighted. Finally, a simplified computati onal model of multielement in jector flows was constructed to explore the sensitivity of wa ll heating and improve combustion e fficiency to injector element spacing. Design space refinement using surrogate models and a multiobjective merit function formulation facilitated an efficient framework to investigate the multiple and competing objectives. The analysis suggests that by adjus ting the multielement injector spacing, the flow structures can be modified, resu lting in a better balance betw een wall heat flux and combustion length.
PAGE 17
17 CHAPTER 1 INTRODUCTION The Space Shuttle Main Engine (SSME) was de veloped as a reusable launch engine. However, even with noticeable success, the co mplexity of the SSME drove up the maintenance costs of the engine due to the harsh operating environment among the engine components. These additional maintenance costs were unforeseen in the original design and analysis efforts.1 For this reason, NASA is striving to design a simpler, more reliable engine.2,3 Today, there are more computational, diagnostic and experimental tool s available to assist in streamlining the design process than when the SSME was designed. For example, flow and combustion processes can now be modeled using a dvanced computational fl uid dynamics (CFD) tools.4,5 Using CFD tools can help narrow down the design of an engine co mponent before a prototype is built and tested. Often, designs are based on past experience, in tuition, and empirical calc ulations. Without a comprehensive and methodical process, the desi gns may not be satisfact orily refined without tedious and inefficient trialanderror exercises. Optimization t ools can be used to guide the selection of design parameters a nd highlight the predicted best de signs and tradeoff designs. In this study, CFD and optimization tools are used to develop further capabilities in the design and analysis of a radial turbine and an injector : critical liquid rocket engine components. Current design practices for inje ctor and turbine designs incl ude past experience, intuition, and empirical calculations. It is difficult to fully analyze, design, and optimize the combustion chamber using experimental information alone Because of measur ement difficulties in combustion chambers due to high temperatures and pressures, much is still too poorly understood of the flow dynamics to accurately pred ict the response of futu re injector designs. Using CFD can be used as a tool to help an alyze and understand the flow dynamics within the combustion chamber to fill in knowledge gaps left by the injector experiments. This research
PAGE 18
18 effort will use CFD as an analysis and design tool for rocket engine components to improve component design and provide information on the system processes that may be difficult to obtain experimentally. CFD valid ation efforts are ongoing to fine tune CFD codes and improve their predictive ability. To use CFD effectivel y as a design tool for ro cket engine component design, an efficient design methodology must be used. A CFDbased optimization methodology can be used to analyze, design, and optimi ze combustion chambers by using the increased knowledge provided by CFD simulation to pred ict better designs. Through successive refinement of the design process, or design sp ace refinement (DSR), the design optimization methodology is streamlined and improved to make it suitable for complex CFDbased design. In summary, this research effort seeks to improve the design efficiency of complex rocket engine components using improved CFDbased optimiza tion and design space refinement techniques. This chapter presents a general introduction to the issues that motivate the work, including background information of rocket engine cycles a nd individual component reliability. Past and current design practices for sele ct rocket engine components ar e described and their limitations are noted. Next, CFDbased optimization and DS R are introduced as methods for improving upon traditional experimentallybas ed analysis and design. Fina lly, a roadmap of the various chapters is presented. 1.1 Thrust Chamber Characteristics A rocket engine thrust chamber is comprised of a main injector with hundreds of injector elements, a combustion chamber, and a nozzle. Supporting elements include turbopumps and possibly a preburner or gas generator. In addi tion to the fuel tanks, pi ping, and valve system, a rocket engine is complex, and all components must be designed fo r maximum efficiency.
PAGE 19
19 1.1.1 Rocket Engine Cycles Rocket engines carry both the fuel and the propellant with the vehicle. For chemical rockets, the propellant is a large part of the total mass of the vehi cle. The basic goal in rocket engine design is to obtain the highest thrust possible with the lowest total vehicle weight.6 As a general objective, it is desirable to have a specific impulse that is as larg e as possible. Hydrogen and oxygen are often chosen as the fuel and propellant because they have very high values of specific impulse. While the goal of maximizing the specific impulse is straightforward, many rocket engines designed to help achieve the goa l are inherently complex and expensive. The need for better performance drives the complexity. Pressurefed engines are relatively simple;7 two high pressure tanks are connected directly to the combustion chamber, and valv es are used to regulate the fl ow rate. Pressurefed engines have low cost8 and good reliability9 because they are simple with few components. However, pressurefed engines require heavy, massive tanks to provide the required pressure. For this reason, turbopumps were used in other rocket en gine cycles to deliver the needed pressure, allowing for lighter fuel and propellant tanks. Ad ditionally, gas generators or preburners were required to supply hot gases to the turbines. This led to a need for additional components, where each component could compromise engine reliability. In a gasgenerator cycle (Figur e 11B) a portion of the fuel is burned before it reaches the combustion chamber. The exhaust from the gas ge nerator is used to drive a turbine, which is used to drive the pumps. After the e xhaust leaves the turbine, it is dumped.10 Gas generator engine cycles provide improved th rust and specific impulse as compared to a pressurefed engine cycle.11. The expander cycle (Figure 11A) eliminates the need for a gas generator, thus resulting in a simpler and more reliable syst em. The expander cycle is the simplest of the pumpfed engine
PAGE 20
20 cycles,12 and it offers a high specific impulse.13 Current expander cycle rocket engines in use are Pratt and WhitneyÂ’s RL60 and the Ariane 5 ESCB. In the expander cycle, the unburned fuel is preheated by being passed through tubes used to cool the combustion chamber. This heated gas is passed through the turbine used to drive the fuel and oxidizer pumps The fuel is then sent to the combustion chamber where it is combusted. Us ing preheated fuel instead of hot combustion exhaust leads to a low turbine failure rate.14 One disadvantage to the expander cycle is that the engine is limited by the amount of power that the turbine can deliver. In this case, power output can be increased by improving the turbine drive, or increasing the amount of heat transferred to the unburned fuel used to drive the turbine.15 In the Space Shuttle main engine (SSME) cycle shown in Figure 12, a portion of the fuel and oxygen is used to power the turbopumps. The SSME has two separate gas generators; the exhaust of one powers the turbopump that deliver s fuel, while the other powers the turbopump that delivers oxygen. Unlike the gas generator cy cle, the fuelrich exhaust is not dumped. Instead, the remaining fuel and oxygen also en ter the combustion chamber and are combusted. For this reason, the gas generators in the SSME are referred to as preburners. A simple preburner cycle is shown in Figure 11C. Using preburners maximized the specific impulse of the engine.16 Thus, the SSME has very high performance, but is also highly complex. For this reason, it is very expensiv e to operate and maintain. 1.1.2 Engine Reliability Issues One of NASAÂ’s goals in reusable launch vehi cles is to improve engine reliability. After each shuttle flight, the engines we re removed from the Space Shuttle and inspected. Parts of the engine were replaced as nece ssary. Improving engine reliabilit y and engine component life can help drive down the costs of reus able launch vehicles. Jue and Kuck1 examined the reliability of the various Space Shuttle Main Engine (SSME) components by determining the probability for
PAGE 21
21 catastrophic failure of various components as shown in Figure 13. The components with the lowest reliabilities are the highpressure fuel turbopump (HPFT P), the highpressure oxidizer turbopump (HPOTP), and the large throat main combustion chamber (LTMCC). The turbopumps and the combustion chamber are com ponents common to all rocket engine cycles (with the exception of a pressure fed cycle in which turbopum ps are absent). Improving these components can therefore cause large improvement s in overall engine re liability. Using an expander cycle, for example, can improve th e turbopump reliability by providing a milder turbine environment, thus improving the life of the turbine blades. Improving the injector configuration can increase the reliability of the main combustion chamber (MCC) by helping to alleviate local hot spots within the co mbustion chamber and prevent wall burnout. To help address the issue of engine reliability, NASA is gath ering data for the combustion chamber design of the next generation of reus able launch vehicles. Specifically, NASA is overseeing tests on several gasgas H2/O 2 rocket injectors by Marshall et al.17 Particularly, the injector tests are being run to supply data for the purpose of CFD validation. In this way, direct comparisons can be made between the experiment al and computational da ta. Also, by steering the experiments, NASA can ensure that any specifically needed da ta is extracted in sufficient amounts. Efforts are also in place to improve th e life and efficiency of additional thrust chamber components, such as the turbin e which drives the fuel pump. 1.2 Current and Past Design Practices Injectors of the past and even of today are designed primarily thr ough the use of empirical models.18 Injectors are also frequently designed based on the results of simple smallscale tests, which are then scaled to a fullscale model.19 Many injectors are also designed based on cold (nonreacting) flow tests.20 These cold flow tests are us ed primarily to test the mixing capabilities of an injector desi gn. Then correlations were de veloped to apply them to hot
PAGE 22
22 (reacting) flow tests.21 Various design parameters are test ed experimentally by making the thrust chamber apparatus of modular design. In this way, the injector can be analyzed by switching out injectors, changing combustion ch amber length, and changing nozzl es in addition to varying the chamber pressure, mass flow rates, etc.17,22 The prospects of CFDbased design have been explored only recently. This is due to mode ling difficulties as a result of the presence of turbulence in combination with reacting flow and possibly including multiphase and supersonic flow. CFD validation efforts are still ongoing. The design of turbines for rocket engine appl ications is more advanced than injector design. This is because turbines are designed straightforwardly for its aerodynamic and stress properties; the turbines for rocket engines have no added complications due to reacting and/or multiphase flow. Turbines are now routinely designed with the aid of CFD, allowing for sophisticated flow prediction and designs. 1.2.1 Rocket Engine Injector Design Osborn et al.23 reported on early water cooled co mbustion chambers that experienced significant wall erosion near th e injector. This erosion wa s common, and trial and error experimentation often resulted in the quick de struction of much of the thrust chamber. Peterson et al.24 constructed an injector with 137 el ements based on tests from a smaller injector with only 37 elements. It was noted th at due to improved coolin g in the larger scale injector, the number of injector elements, and hence, injector spacing, could be increased. This in turn meant that the mass flow rate per elemen t was increased, requiring an increase in the size of the injector orifices to regul ate pressure. Separate thrust ch ambers were built to distinguish differences in performance for cool ed and uncooled thrust chambers. The early optimization efforts of an injector element may have consisted of a dimensional analysis of geometric design variables and flow properties, as in the Apollo Service Propulsion
PAGE 23
23 System Injector Orifice Study Program.25 In this study, a dimensional analysis along with BuckinghamÂ’s pi theory reduced the number of design variables from 14 to 8, which still made testing all possible combinations an intractable problem using e xperimental techniques alone. Thus, an experimental design analyses was applied to reduce the number of tests. Limited by the use of experiments only, six criti cal design variables were chosen and were tested at two levels per variable. After 32 cases had been built and tested, the sensitivity of the response to the design variables required a reevaluation of the optim ization process. However, the results of the analysis eventually yielded information about th e relationship between the flow behavior and the chosen geometry parameters. Curve fits were then developed based on the optimization results for later use. Gill26 described the movement from injector de signs based on histor ical data and the injector designerÂ’s experience and intuition to co mputeraided designs. These early analyses included onedimensional approximations such as equilibrium reactions. These equilibrium approximations were based on the assumption of Â“fas tchemistry,Â” and were used to characterize the combustion field to aid the design process. GillÂ’s research looked at the effects of the injector geometry and injection pattern on the co mbustion process. Performance of the injector had to be maintained while increasing the life of the injector by preventing burning through or erosion of the combustor components. Most experimental techniques were insufficient to predict for many conditions that could reduce injector or combustion chamber life. For this reason, researchers now look to employ CFD techniques to aid in injector design. For example, Liang et al.27 developed semiempirical calculations to aid in the computer modeling of liquid rocket combustion chambers. Sindir and Lynch28 simulated a single element GO2/GH2 injector using an 8species, 18 reaction chemical
PAGE 24
24 kinetics model with the kturbulence model and compared the velocity profile to experimental measurements with good results. Tucker et al.5 stressed the importance of supplying experimental results to aid in the task of CFD validation. Shyy et al.29 and Vaidyanathan et al.30 identified several issues inherent in CFD modeling of injectors, including the n eed for rigorous validation of CF D models and the difficulty of simulating multielement injector flows due to lengthy computational times. They chose to simulate single injector flow, as this allowed fo r the analysis of key co mbustion chamber life and performance indicators. The lif e indicators were the maximum temperatures on the oxidizer post tip, injector face, and combustion chamber wall, and the performance indicator was the length of the combustion zone. These conditions were expl ored by varying the impinging angle of the fuel into the oxidizer. The studies were successful in using CFD to specifica lly access the effects of small changes in injector geometry on combustion performance. 1.2.2 Rocket Engine Turbine Design Turbines are developed based on basic knowledge of the fuel characteristics, estimated stresses, and turbine pressure. The turbines are driven using hot exhaust gas from the combustion process. Turbines were and sti ll are designed by using em pirical relations to calculate a preliminary starting point for the desi gn. The designs rely both on past information from similar turbine designs and on experiment al correlations. Unlike for injector design, computers have been used in the design process fo r rocket engine turbines for over 40 years. Beer 1965 used the aid of computers to estimate the velocity distributi on across a turbine blade to help improve turbine design. Simple 1D co des are routinely used to provide preliminary information on design temperatures, stresses, overa ll flow geometry, velocities, and pressures.31 The prediction of turbine perf ormance was later improved by th e consideration of aerodynamic effects that lead to pressure losses.32
PAGE 25
25 Current design methodologies now combine past design practices with CFD analyses. Experimental correlations and simple computer programs are used to determine the preliminary design. CFD can then be used to finetune or analyze the design.33 In particular, the CFD analyses are used to shape the blade such that problems such as large pressure losses or unsteadiness are minimized.34 1.3 CFDBased Optimization The designs of many aerospace components are m oving from the experimental realm to the computational realm. Some products can now be completely designed on computers before they are built and tested, including aircraft such as the Boeing 777300,35 Bombardier Learjet 45,36 and Falcon 7X.37 Using CFD in coordination with optim ization techniques provides a means of effectively handling complex design problem s such as rocket engine components. 1.3.1 Optimization Techniques for Com putationally Expensive Simulations Scientific optimization techni ques have evolved substantiall y and can now be effectively used in a practical environment. When appropriately combined with computer simulations, the entire design process can be streamlined, reduci ng design cost and effort. Popular optimization methods include gradientbased methods,3843 adjoint methods,44,45 and surrogate modelbased optimization methods such as res ponse surface approximations (RSA),46 Kriging models,4749 or neural network models.50 Gradientbased methods rely on a step by step search for an optimum design using the method of steepest descent on th e objective function accord ing to a convergence criterion. Adjoint methods require formulations th at must be integrated into the computational simulation of the physical laws. In both gradie ntbased and adjoint methods, the optimization is serially linked with the computer simulation. At times, gradientbased optimization is not feasible. For example, Burgee et al.51 found that an extremely noisy objective function did no t allow the use of grad ientbased optimization,
PAGE 26
26 as the optimizer would often become stuck at noiseinduced local minima. Also, for simulating designs with long run times, it may be more time effi cient to run simulations in parallel, as this will allow many simulations to run simultaneously. For these problems, optimization based on an inexpensive surrogate is a good choice. Surrogatebased optimiz ation (SBO) uses a simplified, loworder characterization of the design space instead of relying exclusively on expensive, individually conducted co mputational simulations or expe rimental testing. Surrogatebased optimization allows for the determination of an optimum design while exploring multiple design possibilities, thus providing insight into the workings of the design. A surrogate model can be used to help revise the problem defin ition of a design task by providing information on existing data trends. Furthermore, it can conveni ently handle the existence of multiple desirable design points and offer quantitative assessment of tradeoffs, as well as facilitate global sensitivity evaluations of the desi gn variables to assess the effect of a change in a design variable on the system response.52,53 The SBO approach has been shown to be an effective approach for the design of computationally expensive m odels such as those found in aerodynamics,54 structures,55 and propulsion.56 The choice of surrogates may depend on the problem. For example, Shyy et al.57 compared quadratic and cubic polynomi al approximations to a radial basis neural network approximation in the multiobjective optimization of a rocket engine injector. Simpson et al.58 suggested using Kriging as an altern ative to traditiona l quadratic response surface methodology in the optimization of an aerospike nozzle. Dornberger et al.59 conducted a multidisciplinary optimization on turbomachinery designs using response surfaces and neural networks. To construct the surrogate model, a sufficient nu mber of different designs must be tested to capture how a system response varies with differe nt design parameters. Th e process of selecting
PAGE 27
27 different designs for testing is termed Design of Experiments (DOE). Designs may be tested mainly at the design parameter extremes as in centralcomposite designs (CCD) and facecentered cubic designs (FCCD), or they may be tested across the full range of design parameter values as in multilevel factorial designs46 or Latin hypercube sampling.60 The design space is defined by the physical range of parameters, or va riables, to be explored Variable ranges are often determined based on experience with an existing design, but in many cases, the choice of variable ranges is an educated guess. The simp lest choice is to set simple range limits on the design variables before proceeding with the opti mization, but this is not always possible. Simpson et al.61 used a hypercube design space with desi gn variables range selection based on geometry to prevent infeasible designs in the SBO of an aerospike nozzle. Rodriguez 62 successfully optimized a complex jet configura tion using CFD with simple geometrybased bounds on the design variables. Vaidyanathan et al.63 chose variable ranges based on variations of an existing design in the CFDbased optimizat ion of a single elemen t rocket injector. Commonly, the design space cannot be represented by a simple boxlike domain. Constraints or regions of infeasibility may dict ate an irregular design. Designs that work best for irregular design spaces include Latin Hypercube Sampling60 and orthogonal arrays.46 Once the design space is bounded, the optimization commences. In or der to reduce computational expense, it is very important for a surrogate to be an accura te replacement for an expensive CFD simulation. The application of the SBO approach in th e context of CFDbased optimization with complex flow models can present a significant cha llenge, in that the data required to construct the surrogates is severely limited due to time and computational constraints of the high fidelity CFD runs. In some cases, the problem of high computational cost in SBO can be addressed by performing an optimization on a low fidelity model and translating the result to a high fidelity
PAGE 28
28 model. Doing so requires additional analyses to ensure that the lowfidelity model is a qualitatively similar to the high fidelity model. In other instances, lowfidelity data may be combined with highfidelity data to reduce the overall number of expensive highfidelity runs. The lowfidelity data may employ a correction surrogate for improved accuracy, as shown in Figure 14. Han et al.64 performed a surrogatebased optimization on a lowfidelity 2D CFD model and tested the results on a more expensive 3D model in the optimization of a multiblade fan and scroll system and verified the 2D model us ing physical experiments as well as comparisons to previous studies and the 3D model. Keane65 built an accurate surrogate model of a transonic wing using highfidelity CFD data supplemented with data from a low fidelity empirical model. Alexandrov et al.66 used variation in mesh refinements as variable fidelity models for a wing design. Knill et al.67 saved 255 CPU hours by supplementing highfidelity data with a lowfidelity aerodynamic model in the optimization of a high speed civil transport wing. Vitali et al.68 used surrogates to correct a lowfidelity model to better approximate a high fidelity model in the prediction of crack propagation. Thus, the accuracy of the lowfi delity models can be improved by combining the lowfidelity mode l with the surrogate correction, and a new surrogate could be constructed based on th e composite model. Venkataraman et al.69 also used a surrogate correction on a lowfideli ty model to aid in the optimization of shell structures for buckling. Balabanov et al.70 applied a surrogate co rrection to a lowfide lity model of a high speed civil transport wing. Haftka71 and Chang et al.72 scaled surrogate models with local sensitivity information to improve model accuracy for analytical and structural optimization problems.
PAGE 29
29 With continuing progress in computational si mulations, computationalbased optimization has proven to be a useful tool in reducing the design process duration and expense. However, because of the complexity of many design problems such as is involved with liquid rocket engine components, CFDbased optimization can be a daun ting task. In some cases, a surrogate model is fit to the points and the predictive ca pability for the unknown cases is found to be unacceptably low. One solution is to try an alte rnative surrogate modelÂ—perhaps one with more flexibilityÂ—and approximate the response again. Another remedy is to refine the design space based on the imperfect but useful insight gained in the process. By reducing the domain size, design space refinement often naturally leads to the improvement of a surrogate model. 1.3.2 Design Space Refinement in CFDBased Optimization Using Surrogate Models The design space is the region in which objec tive functions are sampled in search of optimal designs, where the region is bounded by defi ned constraints. Design space refinement (DSR) is the process of narrowing down the sear ch by excluding regions as potential sites for optimal designs because (a) they obviously violat e the constraints, or (b) the objective function values in the region are poor. In the optimiza tion of a high speed civil transport wing, Balabanov et al.70 discovered that 83% of the points in th eir original design space violated geometric constraints, while many of the remaining point s were simply unreasonable. By reducing the design space, they eliminated their unreasonable designs and improved the accuracy of the surrogate model. Roux et al.73 found that the accuracy of a polyno mial RSA is sensitive to the size of the design space for structural optimi zation problems. They recommended the use of various measures to find a small Â“reasonable design space.Â” It is often advantageous to re fine the design space to meet th e needs of the design problem. For example, design constraints can prohibit the use of a simple hypercube design space.74,75 In other instances the fitted surrogate is simply inaccurate. For a highdimensional design space,
PAGE 30
30 adequately filling a hypercube design space regi on with points becomes cost prohibitive. Balabanov et al.70 noted that to sample only at the vert ices of a boxshaped design requires 2 n points, where n is the number of design variables. In contrast, a simplex sampled at the vertices would require only n + 1 points. In particular, for n = 25, a box domain has a volume 1010 times larger than a corresponding ellipsoid. Several techniques are used to improve the optimization process and surrogate accuracy by reducing the domain over which the function is approximated. Madsen et al.74 refined the design space to an ellipsoid shape in the shape optimization of a diffuser to prevent the unnecessa ry CFD analysis of physically infeasible designs. In the optimization of a superdetonative ram accelerator, Jeon et al.76 found that a surrogate model that was a poor fit due to sharp curvature in the resp onse and was able to ach ieve a better result by performing a transformation on the de sign space, hence, smoothing out the curvature. Papila et al.77 explored the possibility of reducing errors in regions the design space by constructing a reduced design space on a small region of interest. Balabanov et al.70 eliminated suspected infeasible design regions from the design space usi ng a lowfidelity analysis prior to optimizing using a highfidelity CFD analysis in the design of a high speed ci vil transport wing. Papila et al.78 refined the design space by expanding the va riable bounds to account for optimal designs that lay on the edge of the design space in th e shape optimization of a supersonic turbine. 1.4 Summary Rocket engine component design is a co mplex process that can be made more straightforward using CFDbased optimization tec hniques. The performance and efficiency of key components can be improved using the SBO me thods. The application of the SBO approach to CFD problems can be challenging due to time a nd computational constraints. Tools including lowfidelity analyses and design space refinement can help redu ce computational expense. In
PAGE 31
31 particular, design space refineme nt techniques can be employed to facilitate a more efficient optimization process by reducing th e number of necessa ry computationally expensive CFD runs. 1.5 Goal and Scope The research goal is to investigate ad aptive DSR using measures based on model uncertainty from varying surrogate models and focusing on regions of interest to improve the efficiency of the CFDbased design of rocket engine components. Chapter 2 reviews the proposed DSR techniques as well as the optimiza tion framework that is used to complete the optimizations. This includes the use of new op timization tools to accelerate the optimization process. In Chapter 3, DSR techniques are applie d in the lowfidelity optimization of a radial turbine for liquid rocket engines. The preliminary design of a ra dial turbine is a logical next step, because 1) it is a new design where th e optimal design space size and location is unknown, and 2) a 1D solver is used to provide function evaluations allowi ng for the collection of a large number of data points. The DSR techniques are used to dete rmine both an applicable design space size by excluding infeasible regions a nd the region of interest by excluding poorly performing design points. The optimization framework technique s are used to complete the optimization process. The optimization techni ques then combined with CFD with a final objective of improving the rocket engine injector design process. The governing equations for combusting flow and a background on rocket engi ne injectors are given in Chapter 4. The optimization methods are demonstrated using an analysis of a trapezoidal bluff body to improve combustor mixing characteristics in Chapter 5. Th e bluff body analysis is used to show that DSR can identify regions of poor simulation accu racy. Chapter 6 presents the CFD modeling effort for a single element injector. The CFD resu lts are compared to an experimental work for accuracy. Then, the flow dynamics leading to wall heat transfer are analyzed, providing information that would be difficult to obtain using an experimental analysis alone. Finally, in
PAGE 32
32 Chapter 7, CFDbased optimization techniques are applied to improve the design process of a multielement rocket injector. In this case, th e optimization techniques must be applied in the context of a CFDbased design process where a limited number of function evaluations are available. Chapter 8 concludes the effort. Table 11. Summary of CFDbased design optimiza tion applications with highlighted surrogatebased optimization techniques. Author Application Method highlighted Key results Shyy et al.79 Rocket engine injector Surrogate comparison Polynomial approximation compared to neural network surrogate model. Simpson et al.80 Aerospike nozzle Surrogate comparison, design space selection Kriging model compared to quadratic response surface. Simple design variable ranges selected using geometry. Dornberger et al.59 Turbomachinery designs Surrogate comparison Multidisciplinary optimization using response surfaces and neural networks. Rodriguez62 Jet configuration Design space selection Optimization using simple geometrybased bounds on design variables Vaidyanathan et al.63 Rocket engine injector Design space selection Design variab les chosen from variations of existing design Han et al.64 Multiblade fan/scroll system Lowfidelity optimization Optimized lowfidelity 2D model to predict for highfidelity 3D model Keane65 Transonic wing Multifidelity optimization Supplemented highfidelity CFD data with lowfidelity empirical model Alexandrov et al.66 Wing design Multifidelity optimization Varied mesh density for variable fidelity models Knill et al.67 High speed civil transport wing Multifidelity optimization Saved 255 CPU hours using lowfidelity supplemental data Balabanov et al.70 High speed civil transport wing Multifidelity analysis with surrogate correction Used coarse and fine finite element models as lowand highfidelity models Madsen et al.74 Diffuser Design space refinement Removed infeasible design space regions to prevent unnecessary CFD analyses Jeon et al.76 Ram accelerator Design space refinement Performed transformation on design space for better response surface fit Papila et al.77 Supersonic turbine Design space refinement Refine design space to improve regions with high error Balabanov et al.70 Civil transport wing design space refinement Used lowfidelity model to refine design space before optimizing using highfidelity model Papila et al.78 Supersonic turbine Design space refinement Design space expansion to explore optimal designs at edge of design space
PAGE 33
33 A fuel p um p oxidizer p um p oxidizer fue l Thrust chamber turbine B fuel p um p oxidizer p um p o xidizer fue l Thrust chamber turbine gas generator exhaust C fuel p um p oxidizer p um p oxidizer fue l Thrust chamber turbine preburner Figure 11. Rocket engine cycles A) Expander cycle, B) gas ge nerator cycle, C) preburner cycle. Figure 12. SSME thrust chamber component di agram showing lowpressure fuel turbopump (LPFTP), lowpressure oxidi zer turbopump (LPOTP), highpressure fuel turbopump (HPFTP), highpressure oxidize r turbopump (HPOTP), main fuel valve (MFV), main oxidizer valve (MOV), combustion chamber valve (CCV), and the main combustion chamber (MCC). Picture reproduced from [81].
PAGE 34
34 Figure 13. SSME component reliability data. Chart reproduced from [1]. Figure 14. Surrogatebased optimization using mu ltifidelity data. A multifidelity analysis may require a surrogate model for correc ting the lowfidelity model to better approximate the highfidelity model. refine design space construct surrogate Verify optimum location with highfidelity model Lowfidelity model w/ correction Surrogate correction Lowfidelity model Assess surrogate Highfidelity model
PAGE 35
35 CHAPTER 2 OPTIMIZATION FRAMEWORK The optimization framework steps include 1) modeling of the objectiv es using surrogate models, 2) refining the design space, 3) reduc ing the problem dimensionality, and 4) handling multiple objectives with the aids of Pareto front and a global sensitivity evaluation method. Several procedures may require moving back and fo rth between steps. Fi gure 21 illustrates the process used to develop optimal designs fo r the problems involving multiple and possibly conflicting objectives. The steps of the framew ork are detailed in the following sections. The first step in any optimization problem is to identify the performance criteria, the design variables and their allowable ranges, and the design constraints. This critical step requires expertise about the physical process. A multiob jective optimization problem is formulated as (),:1,;:1,ji M inimizewherefjMxiN FxFx Subject to ()0,:1, ()0,:1,p kwherecpP wherehkK CxC HxH Once the problem is defined, the designs are evaluated through experiments or numerical simulations. For numerical simulations, the type of numerical model and design evaluation used varies with the goals of the study. For a simple preliminary design optimization, the use of an inexpensive 1D solver may be su fficient. However, for the final detailed design, more complex solvers may be needed.82,78 The choice of a model has an important bearing on the computational expense of evaluating designs. When obtaining many design points is timeprohibitive, it is often more prudent to use an inexpensive surrogate model in place of the expensive numerical model.
PAGE 36
36 2.1 Optimization Using Surrogate Models This step in the framework involves deve loping alternate models based on a limited amount of data to analyze and optimize designs. The surrogates provide fast approximations of the system response making optimization and se nsitivity studies possible. Response surface approximations, neural network tec hniques, spline, and Kriging are examples of methods used to generate surrogates for simulations in the optimization of complex flows52 involving applications such as engine diffusers,74 rocket injectors,63 and supersonic turbines.78,82 The major benefit of surrogate models is th e ability to quickly obtain any number of additional function evaluations with out resorting to more expensive numerical models. In this aspect, surrogate models can be used for multiple purposes. Obviously, they are used to model the design objectives, but they can also be used to model the constraint s and help identify the feasible region in design space. Key stages in the construction of surrogate models are shown in Figure 21. 2.1.1 Design of Experiments The search space, or design space, is the set of all possible combinations of the design variables. If all design variables are real, the design space is given asN xR where N is the number of design variables. The feasible domain S is the region in design space where all constraints are satisfied. For adequate accuracy, the data points used in the surrogate mode l must be carefully selected. The proper data selec tion is facilitated by using Desi gn of Experiments (DOE). One challenge in design optimization despite the type of surrogate used is called the Â“curse of dimensionality.Â” As the number of design variables increase, the number of simulations or experiments necessary to build a surrogate incr eases exponentially. This must be taken into consideration in choosing a DOE. The key issues in the selection of an appropriate DOE include
PAGE 37
37 ( i ) the dimensionality of the problem, ( ii ) whether noise is important source of error, ( iii ) the number of simulations or experi ments that can be afforded, ( iv ) the type of surrogate used to model the problem, and ( v ) the shape of the design space. If noise is the dominant source of error, DOEs that reduce sensitivity to noise ar e commonly used. These include central composite designs (CCD) or facecentered cubic designs (FCCD) for boxshaped domains. Design optimality designs such as Dor Aoptimal83 (Myers and Montgomery 2002, pp. 393 Â– 395) designs are useful for irregular shaped domains and high dimens ional domains when minimizing noise is important. Specificall y, these designs can be used to re duce the number of points in an experimental design for a given accuracy. When noi se is not an issue, sp acefilling designs such as LatinHypercube Sampling (LHS)60 or Orthogonal Arrays (OAs)46 are preferred to efficiently cover the entire design space. Central composite designs and facecentered cubic designs ar e intended to minimize the presence of noise in the response. They are co mprised of a single point at the center of the design space with the remaining points situated along the periphery of the design space, as shown in Figure 22. The center po int detects curvature in the system while all other points are pushed as far as possible from the center provid e noisesmoothing character istics. The FCCD is actually a modified version of th e CCD with the axial points move d onto the edges of a square. An FCCD is required when the limits of the de sign space are strictly set, requiring a square design space. The CCD and FCCD are most popular for re sponse surface approximations. They are commonly used to reduce noise in the response by placing most of the po ints on the boundary of the design space. For responses that do not c ontain noise, such as data obtained using deterministic computational simulations, a spacefilling design may be required for accuracy.
PAGE 38
38 While CCD and FCCD only contain points at th e center and edges of the design space, spacefilling designs distribute points throughout the design space. The most common type of spacefilling design is LatinHypercube Sampling (LHS). The LHS technique consists of dividing the design space N times along each variable axis creating Nk subsections where N is the number of data points and k is the number of design variables. A total of N data points are randomly placed such that only one point exists in each subsect ion interval. This procedure can possibly leave holes in the design space as shown in Figur e 23A. A criterion to maximize the minimum distance between points can help spread points more evenly and prevent bunching. Orthogonal array LHS84 is another method of improving the distri bution of data points within the design space. It insures a quasiuniform distributi on of points across the design space. The design space is uniformly divided into larger subsectio ns and LHS is applied while ensuring one point lies in each large subsection (Figure 23B). 2.1.2 Surrogate Model Identification and Fitting There are many types of surrogate models to choose from. There are parametric models that include polynomial response surfaces approxim ations (RSA) and Kriging models, and there are nonparametric models such as radial ba sis neural networks ( RBNN). The parametric approaches assume the global functional form of the relationship between the response variable and the design variables is known, while the nonparametric ones us e different types of simple, local models in different regions of the data to build up an overall model. Response surface approximations assume that the data is noisy a nd thus it fits the data with a simple, smooth global function to minimize the rootmeansquare (rms) error. Kriging and RBNN use functions localized to the data points. Or dinary Kriging, the type of Krigi ng used in this study, uses all of the data points for that purpose and interpolates over all of them, while RBNN uses a subset of the points and minimizes the rms error. The subs et is determined based on the desired minimum
PAGE 39
39 size of the error. RSA requires a solution of linear equations, as does Kriging for a given correlation function. In Krig ing, the process of finding th e correlation function is an optimization problem that is done either with cr ossvalidation or minimum likelihood estimates. On the other hand, RBNNs let the user experi ment with one parame ter that defines the correlation (SPREAD). However, the process of minimizing the rms is an optimization problem. 2.1.2.1 Polynomial response surface approximation Response surface approximations (RSA) we re developed for experimental design processes. They consist of a polynomial curve fit to a set of da ta points. Usually the polynomial is a 2ndorder model, but higher order models may also be used. A 2ndorder model requires a minimum of 122 kk design points, where k is the number of design variables. Thus, a 2ndorder model with three variables woul d require at least 10 points. A 3rdorder model with three variables would require twice that amount. A 2ndorder model provides good tradeoff in terms of accuracy and computational expense. Fo r RSAs to be effective in design optimization, the sampled points must be near the optimum desi gn. In this case, it is often necessary to perform a screening exercise on th e data to ensure that the de sign space is selected in the appropriate region. Once the pr oper design region has been narrowed down, the region can be populated with data points using a DOE such as CCD or FCCD. The popular polynomial RSAs have been used extensively in design optimization for numerous applications using varying DOEs with varying re sults. RaisRohani and Singh85 used polynomial RSAs to increase the structural effici ency of composite stru ctures while maximizing reliability. They found that the accuracy and e fficiency of RSAs can vary depending on the choice of DOE. Polynomial RSAs have been su ccessful in smoothing re sponses that contain noise. In the aerodynamic design of a subsonic wing, Sevant et al.86 encountered numerical
PAGE 40
40 noise resulting in local maxima. Polynomial RSAs were used to smooth the noise. Sturdza et al.87 used a response surface to optim ize the shape of a fuselage to retain laminar flow across a wing. Clues to the accuracy of RSAs can be found through the simultaneous use of multiple RSAs. Kurtaran et al.88 used linear, elliptic, and quadratic RSAs to optimize for crashworthiness. Kurtaran et al. found that the differences between the pr edictive capabi lities of the various RSAs decreased as the desi gn space was reduced. Hosder et al.89 used RSAs in conjunction with techniques to reduce the desi gn space and perform multifidelity analyses in the multidisciplinary optimization of aircraft. Additi onal references and their key results are given in Table 21. The polynomial RSA assumes that the function of interest f can be represented as a linear combination of Nc basis functions zj and an error term For a typical observation i a response can be given in the form of a linear equation as ()2 1()0 zcN i ijjiii jfzEV (21) where the errors i are considered independent with an expected value equal to zero and a variance equal to 2. The coefficients j represent the quantitative re lation among basis functions zj. Monomials are the preferred basis functions. The relationship between the coefficients j and the basis functions zj is obtained using Ns sample values of the response fi for a set of basis functions zj ( i ) such that the error in the prediction is minimized in a least squares sense. For Ns sample points, the set of equations specified in Equation 21 can be expressed in matrix form as 20 fX E VI (22)
PAGE 41
41 where X is a s cNN matrix of basis functions, also known as a Gramian design matrix, with the design variable values as the sampled points. A Gramian design matrix for a quadratic polynomial in two variables ( Ns = 2; Nc = 6) is given by 22 112111112121 22 122212122222 22 121122 22 1211221 1 1 1ssssssiiiiii NNNNNNxxxxxx xxxxxx x xxxxx xxxxxx X (23) The vector b of the estimated coefficients, which is an unbiased estimate of the coefficient vector and has minimum variance, can then be found by 1 TTbXXXf (24) At a new set of basis function vector z for design point P, the predic ted response and the variance of the estimation are given by 1 ()2 1ÂˆÂˆ ()()TzzzXXzcN iT jj PP jfbzandVf (25) 2.1.2.2 Kriging In cases where a given surrogate performs poorly, it is necessary to use a different type of surrogate model, because the high ly nonlinear nature of some pr ocesses may not be captured by surrogates such as RSAs. Kriging is a popular ge ostatistics technique named after the pioneering work of D.G. Krige47 and was formally developed by Matheron.90 The Kriging method in its basic formulation estimates the value of a functi on or response at some location not sampled as the sum of two components: the linear model (e .g., polynomial trend) and a systematic departure representing low (large scale) and high fr equency (small scale) variation components, respectively. Kriging has the added flexibility of being able to eith er provide an exact data fit or
PAGE 42
42 to act as a smoothing function. It can be used to approximate a highly nonlinea r response with fewer points than a comparable highord er response surface. Koch et al.91 asserts that using Kriging may be preferable over RSAs for de sign problems with a large number of design variables. Mahadevan et al.92 found that Kriging was more accura te than an RSA and used fewer function evaluations for engineering reliability estimates. Kriging has the capability of overcoming the limit of relatively small design space that is inherent in RSAs, and can be used effectively fo r larger design spaces. It is gaining popularity in CFDbased optimization as an alterna tive to RSAs. Forsberg and Nilsson93 found that Kriging provided a better approximation as compared to an RSA in the structural optimization of crashworthiness. However, Jin et al.94 found that Kriging was very accurate, but also very sensitive to noise as compared to other surrogate s in the data fitting of several analytical test problems. Rijpkema et al.95 found that Kriging was better able to capture local details in an analytical test function than an RSA, but also cautioned against the unintentional fitt ing of noise. Chung and Alonso96 compared Kriging to RSAs and found that while both can be accurate, Kriging is better at fitting functions with several local optima. Simpson et al.97 found that Kriging and response surface performed comparably in the optimization of an aerospike nozzle. Kanazaki et al.98 used Kriging to reduce the computati onal cost involved in optimizing a threeelement airfoil using genetic al gorithms by using the Kriging surr ogate for function evaluations. Jouhaud, J. Â–C. et al.99 used Kriging to adaptively refine the design space in the shape optimization of an airfoil. Kriging is represented by a trend perturbed by a Â“systematic departure.Â” In this study, ordinary Kriging is used, where the constant trend is represented by the sample mean. Other types of Kriging include simple Kriging that uses a constant trend of zer o, and universal Kriging
PAGE 43
43 uses a linear trend. The systematic departure component represents the fluctuations around the trend with the basic assumption being that thes e are correlated, and the correlation is a function of distance between the locations under consideration. More precise ly, it is represented by a zero mean, secondorder, stationary process (mean and variance constant with a correlation depending on a distance) as described by a correlation mode l. Hence, these models suggest estimating deterministic functions as ()()cov((),())()()00,ijyEij xxxx (26) where is the mean of the response at sampled design points, and is the error with zero expected value and with a correlation structur e that is a function of a generalized distance between the sample data points. ()()2cov(),(),ijijR xxxx (27) In this study, a Gaussi an correlation structure100 is used 2 1,expijij kkk kvNR xxxx (28) where R is the correlation matr ix among the sample points, Nv denotes the number of dimensions in the set of design variables x identifies the standard devia tion of the response at sampled design points, and k is a parameter which is a measure of the degree of correlation among the data along the kth direction. Specifically, the parameters and are estimated using a set of N samples ( x y ) such that a likelihood function is maximized.100 Given a probability distribution and the corresponding parameters, the likelihood f unction is a measure of the probability of the sample data being drawn from it. The model estimates at nonsampled points are 1ÂˆÂˆÂˆ ()Ty rR y 1 (29)
PAGE 44
44 where Âˆ is the estimated mean, r identifies the correlation vect or between the set of prediction points and the points used to construct the model, y a vector of the response at the sampled points, and 1 denotes an Nsvector of ones. The mean is estimated by 1 11ÂˆTT 1R11R y (210) and the standard deviation of the response is estimated as 1 2ÂˆÂˆ ÂˆT sN y 1R y 1 (211) On the other hand, the estimation variance at nonsampled design points is given by 1 21 11 (())1T T TR VyR R 1r xrr 11 (212) In this study, the model estimates, estimation va riance, and standard deviation of the response are calculated using the Matlab toolbox DACE.101 2.1.2.3 Radial basis neural networks Neural networks are anothe r alternative to traditiona l response surface methods. Specifically, neural networks are able to fit da ta with a highly complex and nonlinear response. Artificial neural networks are made of inte rconnected local models called neurons. A high number of neurons can improve the accuracy of the fit, while using fewer neurons improves the smoothing qualities of the model. Mahajan et al.102 used neural networks (NN) to optimize a mechanically aspirated radiation shield for a meteorological temperat ure sensor to gain information on the workings of the temperatur e sensor system and the performance of the radiation shield. Papila et al.78 used radial basis neur al networks to supplement the data used to construct an RSA in the optimization of a supersonic turbine for rocket propulsion. Charalambous et al.103 used neural networks in the predic tion of bankruptcy and stressed that the
PAGE 45
45 prediction ability of a neural network is problem dependent. Hsken et al.104 used neural networks in turbine blade op timizations. Shi and Hagiwara105 used neural networks to maximize the energy dissipation of the crashworthiness of a vehicle component. Pidaparti and Palakal106 compared the prediction capability of neural networ ks with experimental results in the prediction of crack propagation in agi ng aircraft. Chan and Zhu107 demonstrated the ability of neural networks in modeling highly non linear aerodynamic characteristi cs with many variables. Brigham and Aquino108 used neural networks to accelerat e their optimization search algorithm by providing inexpensive function evaluations. Neural networks are made up of a distributi on of artificial neurons, whose function is modeled after biological neurons. A shape f unction is applied at each neuron with the combination being a complete function curve. Each neuron is given a weight in the network that defines that individual neuronÂ’s strength. Neural networks are often diffi cult to train due to the required specification of several parameters that ca n affect the fit of the model. Care must also be taken, as there is sometimes danger in ove rfitting the data, resul ting in poor prediction capability away from the known data points. None theless, the models are extremely flexible and can provide an exact data fit, or meet a userdefin ed error criterion to filter noise to prevent data overfitting. Neural networks us e only local models, and the de gree of effective influence of each local model, or neuron, is a selection made by the user. Radial basis neural networks are twolayer networks consistin g of a radialbasis function (RBF) and a linear output layer. The radial basis function is given by bfradbas cx (213) where b is the bias, c is the center vector asso ciated with each neuron, x is the input design vector, and the function radbas is the Matlab function name for an RBF, or a local Gaussian
PAGE 46
46 function with centers at individual locations from the matrix x. The radial basis function is given by 2expradbasnn (214) where n is given by the term in parenthesis in Equation 213. Based on these equations, the function prediction is given by 2 1Âˆ expN ii iywb cx (215) where wi is the weight vector for each neuron i, and N is the number of neurons. The Gaussian RBF decreases monotonically with the distance from its center. The bias b is set to 0.8326/SPREAD where SPREAD is the spread constant, a user defined value that specifies to the radius of influence for each neuron where the ra dius is equal to SPREAD/2. Specifically, the radius of influence is the dist ance at which the influence reaches a certain small value. If SPREAD is too small, then prediction is poor in regions that are not near the position of a neuron. If SPREAD is too large, then the sensi tivity of the neurons will be small. Because RBNNs are based on the combination of multiple local approximations, a spacefilling DOE is necessary. If large spaces exist in the DOE, it may be difficult to ove rlap information from nearby points. The capability of the RBNN to f ill the information gaps in the design space is based on the size SPREAD. If SPREAD is too sm all, the RBNN surrogate may not provide any predictive ability in some regi ons within the large space. For this region, there must be coordination between the spacing of data point s within the design sp ace and the size of SPREAD. Increasing the number of neurons will improve the accuracy of the surrogate. Neurons are added to the network one by one until the sum of the squares of the errors (SSE) is reduced enough to reach a specified error goa l. If the error goal is set to zero, neurons will be added until
PAGE 47
47 the network exactly predicts the input data. Howe ver, this can lead to overfitting of the data which may result in poor prediction between data point s. On the other hand, if error goal is large, the network will be undertrained and predictions even on data points will be poor. For this reason, an error goal is judicious ly selected to provide early st opping of the network training to prevent overfitting and increase the overall prediction accuracy. The purpose of the user is obviously to de termine good values of SPREAD and the error goal, GOAL. This can be done by reserving a sma ll number of data points as test data. The remaining data is designated as the training data. First, the network is trained using the training data and the Matlab function newrb. Then, the network is simulated at the test points and the error is calculated between the actual and predicte d response. The process is repeated for a small number of values with selected range of SPR EAD and a small set of values for GOAL. The values of SPREAD and GOAL that produce the smalle st errors in the test data are identified, then the ranges of SPREAD and GOAL are reduced and new values are selected. The process iteration continues until the SPR EAD and GOAL combination is identified that minimizes the error in the test data. The idea l combination can usually be identified within three iterations. In this way, the training data is used to train the neural network, and th e test data is used to tune it. 2.1.3 Surrogate Model Accuracy Measures Once the surrogate models are available, it is imperative to establish the predictive capabilities of the surrogate mode l away from the available data. Several measures of predictive capability are given below. 2.1.3.1 Root mean square error The error i at any design point i is given by Âˆiii f f (216)
PAGE 48
48 where fi is the actual value and Âˆi f is the predicted value. For Ns design points, the root mean square (rms) error is given by 2 1 i i ssN N (217) A small rms error indicates a good fit. The rms e rror is a standard measur e of fit that may be used to compare the performances of surrogates of different types. The adjusted rms error a is used to adjust the rms error based on the number of parameters in a polynomial RSA and is given by 2 1sN i i sa N q (218) where Ns is the number of data points, and q is the number of terms in the polynomial approximation. A polynomial with a large num ber of terms is penalized and deemed less desirable for a polynomial approximation with fewer terms for the same number of data points. For a good fit, a should be small compared to the data value ranges. For Kriging, the standard error measurement is given by the estimated standard deviation of the response given by Equation 211. The squa re root of the estimated standard deviation of the response can be compared to the rms error of polynomial RSAs. 2.1.3.2 Coefficient of multiple determination The adjusted coefficient of multiple determination 2 adj R defines the prediction capability of the polynomial RSA as 2 1 2 1211;wheres sN as i i N s i iadjN f f N ffR (219)
PAGE 49
49 For a good fit, 2 adj R should be close to 1. 2.1.3.3 Prediction error sum of squares When there are an insufficient number of data points available to test the RSA, the prediction error sum of squares (PRESS) statisti c is used to estimate the performance of the surrogate. A residual is obtained by fitting a su rrogate over the design sp ace after dropping one design point from the training set. The value pr edicted by the surrogate at that point is then compared with the expected value. PRESS is given by 2 1ÂˆPRESSsN ii i f f (220) and PRESSrms is given by 2* 1Âˆ PRESSsrmsN ii i s f f N (221) where *Âˆi f is the value predicted by the RSA for the ith point which is excluded while generating the RSA. If the PRESSrms value is close to a, this indicates that the RSA performs well. 2.2 Dimensionality Reduction Using Global Sensitivity Analysis At the beginning of a design optimization, seve ral variables are chosen as design variables with the assumption that they are important to the optimization. However, having large numbers of design variables can greatly increase the cost of the optimization. Section 2.1.2 introduced the concept of the Â“curse of dimensionality,Â” meaning that as the number of design variables increase, the number of data points required to obtain a good approximation of the response can increase exponentially. It is of great benef it, therefore, to simplify the design problem by identifying variables that are unimportant and removing them fr om the analysis. The most efficient way of doing this is to perform a sensiti vity analysis. A global sensitivity analysis can
PAGE 50
50 provide essential information on the sensitivity of a design object ive to individua l variables and variable interactions. By rem oving the variables that have ne gligible influence on the design objective, the dimensionality of the problem can be reduced.109,110,111 Global sensitivity analyses enable the study of the behavior of differe nt design variables. This information can be used to identify the variables which are the least important and thereby can reduce the number of vari ables. A surrogate model f (x) of a square integrable objective as a function of a vector of in dependent input variables x ( 0,11,i x iN ) is assumed and modeled as uniformly distributed random variables. The surrogate model can be decomposed as the sum of functions of increas ing dimensionality as 01212,,,,xiiijijNN iij f ffxfxxfxxx (222) where0d1 x0x f f. If the following condition 11 ... 00siikfdx (223) is imposed for k = i1, Â…, is, then the decomposition describe d in Equation 222 is unique. In the context of a global sensitivity analysis, the total variance denoted as V ( f ) can be shown to be equal to 1 11()n iijN iijNVfVVV (224) where 2 0VfEff and each of the terms in Equation 224 represent the partial variance of the independent variables ( Vi) or set of variables to the total variance. This provides an indication of their relative importance. The partial variances can be calculated using the following expressions:
PAGE 51
51 ([]) ([,]) ([,,])ii ijijij ijkijjijikjkijkVVEfx VVEfxxVV VVEfxxxVVVVVV (225) and so on, where V and E denote variance and the expected value respectively. Note that 1 0iii E fxfdx and 1 2 0([])iiiVEfxfdx Now the sensitivity i ndices can be computed corresponding to the independent variables and set of variables. For example, the first and second order sensitivity indi ces can be computed as ()()ij i iijV V SS VfVf (226) Under the independent model inputs assumption, the sum of all the sensitivity indices is equal to one. The first order sensitivity index for a given va riable represents the main effect of the variable but it does not take into account the effect of interaction of the va riables. The total contribution of a variable on the total variance is given as the sum of all the interactions and the main effect of the variable. The total sensit ivity index of a variable is then defined as ,,,... ()iijijk jjijjikki total iVVV S Vf (227) To calculate the total sensitivity of any design variable xi, the design variable set is divided into two complementary subsets of xi and Z ,1,;j Z xjNji The purpose of using these subsets is to isolate the influence of xi from the influence of the remaining design variables included in Z The total sensitivity index for xi is then defined as total totali iV S Vf (228)
PAGE 52
52 ,total Z iiiVVV (229) where Vi is the partial variance of the objective with respect to xi and Vi,Z is the measure of the objective variance that is depe ndent on interactions between xi and Z. Similarly, the partial variance for Z can be defined as VZ. Therefore, the total objective variability can be written as Z ZiiVVVV (230) Sobol112 proposed a variancebased nonparametric a pproach to estimate the global sensitivity for any combination of design variables using M onte Carlo methods that is also amenable. While Sobol used Monte Carlo simulations to conduct the global sensitivity analysis, the expressions given above can be eas ily computed analytically once th e RSA is available. In the present study, the above referenced expression s are evaluated analytic ally using polynomial RSAs of the objective functions. No accommodations are made for irregularshaped domains, so the analytical treatment works best in a wellrefined design space where the assumption of a boxlike domain will not introduce si gnificant errors. Using a polynomial RSA as the function f(x), the approximation can be decomposed as in Equation 222 and the sensitivity indices can be obtained. 2.3 MultiObjective Optimization Using the Pareto Optimal Front After developing a computationally inexpensiv e way of evaluating di fferent designs, the final step is to perform the actual optimization. In the case of a single objective, this requires a simple search of design space for the minimum value of the objective. For two or more objectives, additional treatment is needed. Highl y correlated objectives can be combined into a single objective function. When the objectives are conflicting in nature, there may be an infinite number of possible solutions th at will provide possi ble good combinations of objectives. These solutions are known as Pareto optimal solutions. While there are numerous methods of solving
PAGE 53
53 multiobjective optimization problems, the use of evolutionary algorithms (EAs) is a natural choice to get many Pareto optimal solutions in a single simulation due to its population based approach and ability to converg e to global optimal solutions. A feasible design x(1) dominates another feasible design x(2) (denoted by x(1) < x(2)), if both of the following conditions are true: 1. The design x(1) is no worse than x(2) in all objectives, i.e., (1)(2) jjffxx for all j = 1,2,Â…, M objectives. (1) x (2)(1)(2)(1)(2) jjjjjMfforjMff xxxxx (231) 2. The design x(1) is strictly better than x(2) in at least one objective, or (1)(2) jjffxxfor at least one {1,2,...,} j M (1)(2)(1)(2)()()jjjMff xxxx (232) If two designs are compared, then the designs are said to be nondominated with respect to each other if neither design dominates the other. A design xS, where S is the set of all feasible designs, is said to be nondominated with respect to a set AS, if : aAax. Such designs in function space are called nondominated solutions. All the designs x ( xS) which are nondominated with respect to any othe r design in set S, comprise the Pareto optimal set. The function space representation of the Pa reto optimal set is the Pareto optimal front. When there are two objectives, the Pareto optim al front is a curve, when there are three objectives, the Pareto optimal front is represente d by a surface and if there are more than three objectives, it is represented by a hypersurface. In this study, an elitist nondominat ed sorting genetic algorithm NSGAII113 and a parallel archiving strategy to overcome the Pareto drift problem114 are used as the multiobjective
PAGE 54
54 optimizer to generate Pareto optimal solutions. The description of the algorithm is given as follows: 1. Randomly initialize a popul ation (designs in the design space) of size npop. 2. Compute objectives and constraints for each design. 3. Rank the population using nondom ination criteria. Many indi viduals can have the same rank with the best individua ls given the designation of rank1. Initialize an archive with all the nondominated solutions. 4. Compute the crowding distance. This distance finds the relative closeness of a solution to other solutions in the function space and is us ed to differentiate between the solutions on same rank. 5. Employ genetic operatorsÂ—selection, crossover, and mutationÂ—to create intermediate population of size npop. 6. Evaluate objectives and constraint s for this intermediate population. 7. Combine the two (parent and intermediate ) populations, rank them, and compute the crowding distance. 8. Update the archive: 9. Compare archive solutions with rank1 solutions in the combined population. 10. Remove all dominated solutions from the archive. 11. Add all rank1 solutions in the current population whic h are nondominated with respect to the archive. 12. Select a new population npop from the best individuals based on the ranks and the crowding distances. 13. Go to step 3 and repeat until the termination criteria is reached, whic h in the current study is chosen to be the number of generations 2.4 Design Space Refinement Techniques Given knowledge about a problem, it is desira ble to perform CFD si mulations with (a) design parameters set in a region that is known to provide good results in order to refine the design (exploitation) or (b) to set the design parameters in an unexplored region and use
PAGE 55
55 additional simulations to decide whether thes e designs have potentia l (exploration). The approaches of exploitation versus exploration are the basis of most global search algorithms. In the context of design space refinement, a mo re modest goal is set by using exploitation to identify the bad regions of a design space: A Â“reasonable design spaceÂ” approach uses simple analysis models or inexpensive constraints to crop areas of the design space that give blatantly poor results. A surrogate based on a small number of points can be used in the identification of Â“badÂ” regions. In this case, the surrogate may not be accurate enough to identify the good regions, but is sufficien t to identify regions th at are obviously bad. Exploration is needed to identify poorly repr esented areas of the design space by using error indicators to identify regions where the surrogate model fits poorly. E xploration is necessary even in regions with apparent moderate performa nce, as the improved surrogate may reveal that the performance in the region is better than expe cted. The design space refinement process using the ideas of exploitation and exploration shoul d provide more accurate surrogate models and reduce the simulation of poorly performing designs. Design space refinement often leads to irre gular domains. Most DOE techniques are intended for boxlike domains. For noisedominat ed problems and RSAs alphabet designs (Aoptimal, Doptimal, Goptimal, etc.) are available.46 For problems dominated by bias errors, space filling designs such as LHS are normally used, generating them first in an enclosing box and throwing out point s outside the domain. Surrogate models have been shown to be an effective means of reducing the number of computationally and time intensive function evalua tions required in CFDbased optimization. However, to use surrogates effectively, it must fi rst be ensured that the surrogate models are high accuracy representations of the system response. An inaccurate surrogate model can prevent further analysis and can cast doubt on the validity of the data. Accuracy can suffer if the choice
PAGE 56
56 of DOE was inappropriate for the ty pe of surrogate chosen or if the design space was too large. The advantages of design space refinement offe r promising ways to deal with an initial inaccurate surrogate model. Current techniques often include forms of design space reduction or windowing to obtain a reasonable design space. 2.4.1 Design Space Reduction for Surrogate Improvement Reducing the design space by reduc ing the ranges of the desi gn variables often leads to better accuracy in the surrogate model. This is often due to the fact that the response in the reduced design space has less curvature. This f eature makes the response easier to approximate. Papila et al.78 reduced the ranges of the design variables to perform a preliminary optimization on a supersonic turbine for the rocket engines of a reusable launch vehicle using RSAs. The goal of the study was to maximize the turbine effici ency and vehicle payload while minimizing the turbine weight. The preliminary design used a si mplified model to provide the system response. A FCCD was applied for each turbine stage. A composite response surface was constructed using desirability functions, and an optimum design was located. The error between the actual response and the predicted response at the lo cated optimum design wa s high, indicating poor RSA accuracies. As a result, Papila et al. re duced the size of the design space by about 80%. The smaller design space resulted in substantial improvement in the RSA prediction accuracies. Reducing the size of the design space in this manner is depicted in Figure 24, and is an example of design space exploitation, and can lead to a good accuracy surrogate in a small region of interest. Shaping the design space based on feedback fr om the objective function values allows the designer to concentrate solely on the region of interest. Balabanov et al.70 pointed out that effort could be saved if the design space was shaped su ch that it only considered reasonable designs rather than using the standard boxshaped desi gn space. The study by Balabanov et al. used this
PAGE 57
57 method in the RSA of a highspeed civil transport model. The a pproach removes large areas of the standard boxshaped design leav ing more of a simplex or ellips oid design. Inexpensive, lowfidelity models are used to predict the system response, and the result is used to shape the design space into a reasonable design space. Points that would lie on the edge of the standard box in a traditional DOE procedure are moved inward to the boundary of the identified reasonable design space. By concentrating the data points on the region of interest, the accuracy of the response surface within that region is improved. The t echnique was used to separately estimate the structural bending material weight and the lift and drag coeffici ents. In these cases, the design space was reduced at least by half, resulting in improvements in RSA accuracy. Roux et al.73 employed techniques such as intermediate response surfaces and identification of the region of interest to improve surrogate accuracy in structural optimization problems. Linear and quadratic response surfaces were used to fit the data. The design space was iteratively reduced around the suspected optimum by fitting respons e surfaces on smaller and smaller subregions. It was found that surroga te accuracy was improv ed more for carefully selected points within a small s ubregion than for a large number of design points within a large design space. Both Balabanov et al.70 and Papila et al.82 combine multiple objectives into a composite objective function before optimizing. However, this procedure can limit design selection to one or a few designs, when mu ltiple other desirable designs might exist. Papila et al.77 pointed out that the inaccuracy in th e initial response surf ace may result in windowing to a region that does no t actually contain the optimum design. Papila et al. argued that particular attention must be paid to the regions of higher e rror in an RSA. Papila et al. explains that because a quadratic RSA is a loworder polynomial, the model can often be inadequate. This is termed as Â“bias error,Â” a nd can obviously be reduced by using higherorder
PAGE 58
58 approximating polynomials. The study by Papila et al. seeks to improve the accuracy of RSAs by using a design space windowing technique to focus only on the region of interest. At the same time, it seeks to exclude regions of the de sign space with high bias errors. Papila studied the effects of fitting three different RSAs: a standard quadratic RSA, a RSA within the region of interest using the design space windowing t echnique, and an RSA using the design space windowing technique while excluding regions with high error. Papila found that the design space windowing technique improved the accuracy of the RSA over the global RSA. However, using the design space windowing technique excl uding regions of high error did not improve accuracy over a standard design space windowing technique. Sequential response surfaces115,116 involve several design space windowing steps. This method is used in the search of a single optimum by coupling surroga te modeling with the method of steepest descent. After a surrogate is fitted, a new design space box is constructed at a new location by stepping in the direction of the steepest descent. A new DOE is applied, and a new surrogate is constructed for the new design space region. The obvious disadvantage to this method is that it could potentiall y require a large number of data points for several different local optimizations. RaisRohani and Singh85 used sequential response surfaces to increase the structural efficiency in various problems. In the study, RaisRohani and Singh employ global and local response surface models. They noted that the global response surface model required a large number of data points for sufficient accu racy. For the local model, a linear response surface model was fit to the local regions. A tota l of ten design subspaces were needed before the optimum was reached. The two methods obtained similar optimum points, indicating comparable accuracy. However, the global response surface required 1000 to 3000 design points, while the local response surface method re quired from 69 to 406 data points. The study
PAGE 59
59 shows that using a large overall design space region over a small, local design space is not always the best choice. Goel et al.117 reached a similar conclusion in the optimization of a diffuser vane shape. Goel et al. found that su rrogates constructed with in a smaller, refined design space were significantly more accurate th an surrogates constructed over the unrefined design space. It has been shown that design space refine ment techniques can increase the accuracy of RSAs. One obvious omission in current design space refinement techniques is a method of refining the design space in the presence of mu ltiple objectives when a satisfactory weighting criterion is not known. In the use of RSAs fo r multiobjective optimizat ion, it is difficult to locate a design space that can simultaneously represent the multiple objectives while retaining accuracy for each objective. One solution is to ensure that the design space is large enough to encompass all of the effects the design. Th is course of action, as Papila et al.78 noted, can introduce substantial error into RSA. Using an al ternative surrogate such as Kriging or neural networks can alleviate this problem. However, doing so may require begi nning the analysis from scratch to obtain a data set using a DOE that is amenable to the desired surrogate model. 2.4.2 Smart Point Selection for Second Phase in Design Space Refinement A popular technique in optimization is the us e of the Efficient Global Optimization (EGO) algorithm proposed by Jones et al.118 This technique uses the variance prediction provided by Kriging to predict the location of the optimum poin t. The procedure used to predict the optimum point is shown briefly in Figure 25 and is described in detail in Section 2.4.3, and involves maximizes the expected improvement of a sample d function. The optimization process usually proceeds by sampling the function at a limited number of points. Kriging used as the surrogate model fit to the data points, and the variance and expected improvement across the design space is calculated. The optimization th en proceeds in a linear fashion, using the value of the expected
PAGE 60
60 improvement to select the optimum. The predicted optimum is sampled, and then added to the data set. The iteration continue s until convergence is reached. Various attempts have been made to impr ove upon the process to make it suitable for multiobjective optimization. Jeong and Obayashi119 used EGO in the multiobjective optimization the pressure distribution at two operating conditions of a twodimensional airfoil shape. Genetic algorithms were used to develop the predicted Pareto front using the value of the expected improvement as the f itness criteria. The Pareto fr ont identifies points that are nondominated in the expected improvement of two objectives. Jeong and Obayashi choose three new data points based on the original Kriging mode l: 1) the predicted point with the maximum expected improvement in the first objective whic h lay at one end of the Pareto Front, 2) the predicted point with the maximu m expected improvement in the second objective at the other end of the Pareto front, and 3) a point with an expected improvement that lies in the middle of the Pareto front. The process was iterated multiple times until a converged optimum was achieved in both objectives. Th is straightforward optimizati on was concerned with and was successful in finding a single optimum, but did not address competing objectives. Knowles120 sought modify EGO for the use of co mputationally expensive multiobjective problems. The new algorithm called ParEGO by simultaneously using Pareto frontbased optimization with the EGO algorithm, similar to the procedure used by Jeong and Obayashi.119 In this case, the goal was to develop an accurate Pareto front using a smaller number of function evaluations than methods typically used to generate Pareto fronts such as the NSGAII113 algorithm. Again, EGO was used to select the points. However, this procedure relies on a few hundred function evaluations. This type of evaluation is only possible with relatively inexpensive function evaluations or a surrogate model that is already sufficiently accurate.
PAGE 61
61 FarhangMehr and Azarm121 used a different selection tech nique for constructing surrogate models for crash analyses. The technique also required a standard DOE and Kriging fit to initialize the response surface. The method then added points based on the predicted irregularities in the response surface. This method allowed for the response surface to be improved while requiring potentially fewer points overall. Thus, the design space was adaptively refined using points th at were selected based on th e predicted objective function, itself. FarhangMehr and Azarm cited the benefits of their approach in reducing the total number of points needed for optimizations based on computationally expensive simulations. The procedure is good for improving surr ogate accuracy, but does not ta ke into account the value of the function, meaning that points may be added in regions of noninterest. 2.4.3 Merit Functions for Data Selection and Reduction Using the concepts such as design space reduction (Section 2.4.1) along with the benefits of innovative point selection as in the previous section, merit f unctions, facilitated by Kriging, can be smoothly integrated into the multiobjective DSR process. Merit functions are statistical measures of merit that use information of th e function values and model uncertainty in the surrogate model to indicate the locations wh ere the function values can be improved and uncertainty reduced. In this research effort, mer it functions are used to select data points such that the accuracy of the surrogate model can be improved using a minimal number of data points. The merit functions rely on pred ictions of the function values and function uncertainty and attempt to balance the effects of the two. Krig ing is selected as the surrogate model for the optimization due to its inherent ability to provi de estimates of both the function value and the prediction variance. Although several merit functi ons are available, the strengths of a given merit function over another for the purpose of multiobjective optimization with an extremely limited number of function evaluatio ns has not been determined.
PAGE 62
62 Four different merit functions are considered. Because ordinary Kriging assumes that the response data is normally distribut ed, a statistical lowe r bound can be calculated. The first merit functions seeks to minimizi ng the statistical lower bound (MF1) given by 1ÂˆMFykVxxx (233) where k is an adjustable scaling factor that is set to 1 in this study.. The variance V is predicted as given in Equation 212. The statistical lower bou nd is simply the predicted function value minus the estimated standard error in the surrogate pr ediction at a given location, where the estimated error is a function of the distan ce between data points. Locations far away from any data points will have higher estimated error, or variance. The second merit function is maximizing the probability of improvement122 (MF2). The function is used to calculate the probability of improving the function value beyond a value T at any location. A point is selected where the probability of improvement is the highest. 2Âˆ Ty MF V x x x (234) where T is a target value given by minminTfPf (235) where fmin is the current predicted optimum, is the normal cumulative distribution function, and P is an adjustable scaling factor set to 0.1 in this study, which corresponds to an improvement of 10% over the current best f unction value. Depending on the value of P using the criterion of maximizing the probability of improvement ( MF2) can results in a highly localized search. Jones123 suggests that selecting several target s can help the search proceed in a more global manner. The third merit func tion is maximizing the expected improvement118 ( MF3). Maximizing the e xpected improvement ( MF3) is the criterion used in the Efficient Global
PAGE 63
63 Optimization (EGO) algorithm developed by Jones et al.118 and has been shown to be effective in searching globally for the optimum It calculates the amount of improvement that can be expected at a given location. minmin 3minÂˆÂˆ Âˆ fyfy MFfyV VV xx xx x xx (236) where is the normal density function. The fourth merit function to be considered is minimizing the expected value of th e minimum function value122 ( MF4). The expected value is simply the predicted value at a given lo cation minus the expected im provement. The criterion MF4 is similar to MF3, but the search tends to be more local. Sasena et al.122 suggests that MF4 should be used only when there is confidence that the optimum region has been found. minmin min 4ÂˆÂˆ ÂˆÂˆ ,0 00 fyfy yfyVV MF VV V xx xxx x xx (237) The procedure is developed such that it acc ounts for the presence of multiple objectives. In this way, a single point can be selected to improve the response surface across multiple objectives. The procedure involves applying a DOE with a large number of points across the design space. The points are then divided into cl usters such that the number of clusters equals the number of desired new points. Clustering the data prior to selection ensures that the data are spread sufficiently across the design space, reduci ng the possibility of choosing two points that are very near each other. For the purpose of mu ltiobjective optimization, the characteristics for each objective need to be combined before the best points can be selected. The points are selected based on a simple weighting function. In this study, the data is clustered based on proximity using the Matlab function kmeans For a given cluster m a rank R is given to each point n
PAGE 64
64 ,,min 1 ,max,minmm k knk m n i kkFF R FF (238) where F is the value of the selection criterion (merit function, function value, etc.), ,minm kF is the point with the smallest value of F within the cluster m n is a point from the cluster m and ,max kF and ,min kF are the maximum and minimum values F respectively, within the entire data set for objective k For MF1 and MF4, the point in each cluste r with the minimum value R is selected. For MF2 and MF3, the point in each cluster with the maximum value R is selected. The full procedure for using merit functions to select points is as follows: 1. Construct a response surface using Kriging fo r each objective using an initial sampling. 2. Determine if the design space size should be reduced. Construct a DOE using a large number of points M within the new design space. 3. Choose a merit function and calculate the merit function value for each objective k at each point M in the new DOE using the predicte d function value and variance. 4. Calculate R at each point M as determined by Equation 238. 5. Choose the desired number of points m to be sampled from the new DOE. The number of points m is selected by the user according to th e number of sample points that can be afforded by the simulation or experiment. The quantity M should be much larger than m 6. Group the points M into m number of clusters. 7. Select the point that has the best value of R from each cluster. The final set of m design points will satisfy three basic criteria: 1. The new points will be evenly spread across the new design space 2. The selected points will be selected such that they are predicted to improve the function values in each objective 3. The selected points are those in each cluster predicted to best reduce the errors in the surrogate models for each objective.
PAGE 65
65 2.4.4 Method of Alternative Loss Functions By modifying the least squares loss function, it is possible to fit several similar, but slightly different polynomial response surfac es to the same set of data. For adequately refined design spaces, it can be shown that the difference among these polynomials is negligible. For design spaces that require some type of refinement, larg e differences can exist. This property of using the differences in different surrogates highlight potential problems in the surrogate model or in the data is not new.75 However, the proposed method has an advantage in that surrogates of the same type can be compared requiring no change in the data set. This may enable a quantitative measurement in the degree of refi nement needed in a problem. For a polynomial response surface the prediction at any point i is given as 1Âˆk ijij jybx (239) where k is the number of terms in the polynomial response surface and xij are values from the matrix X from Equation 23. For example, at any point i a quadratic response surface with two variables can be given as 22 011213241252Âˆybbxbxbxbxxbx (240) or simplified to 01122334455Âˆybbxbxbxbxbx (241) The notation used for these and the followi ng equations are adapted from Myers and Montgomery.46 To determine the coefficients bj, the loss function to be minimized can be defined as 0 1111Âˆp nnnk pp iiiijij iiijLyyybbx (242)
PAGE 66
66 where is the error given by the differe nce between the actual response y and the predicted response and k is the number of terms in the response surface equation. Fi gure 27 illustrates the effect of changing the value of p in the loss function formulation. For a traditional least squares loss function, p = 2. The coefficient vector bj must be determined such that the loss function is minimized. To do so, the partial derivatives must be set equal to zero 1 1 0 1 1sgn0 sgn0,1,2,,n p ii i n p iiij i jL p b L p xjk b (243) resulting in a set of simultaneous equa tions that must be solved where sgn(i) = i / i  The coefficient vector b can be determined using any met hod for solving simultaneous nonlinear equations. In this study, the Matlab function fsolve can be used to obtain the coefficients. Preliminary tests were conducted on data from the preliminary optimization of a radial turbine. Details of this optimi zation are given in Chapter 3. This optimization problem involved the refinement of the design space for a problem with two objectives and involved screening data (Data Set 1) to determine the feasible design space (Data Set 2). Figure 28 shows the results of response surfaces fit to Data Set 2 and illustrate s how the sums of the squares of the errors (2 1sN i iSSE ) of a quadratic res ponse surface vary with p for two objectives. In general, SSE increases as p increases. The amount of change in SSE depends on the size of the design space. This can be better illustrated through use of Pareto fronts. Figure 29 shows the Pareto fronts construc ted using response surfaces from the two objectives for coordinating values of p. In this case, data is used from the preliminary optimization of a radial turbine, and the objectives are the turbine weight, Wrotor, and the totaltostatic efficiency, ts. Figure 29A shows a large difference in the Pareto fronts for an identical
PAGE 67
67 data set and identical forms of the RSAs The only difference was the value of p in the loss function. After the design space was refined (Data Set 3) to improve the accuracies of the RSAs, the Pareto fronts based on varying values of p are quite similar. To study the quantitative differences, the curves were integrated to obtai n the area under the Pareto front curves, and the results were compared to the Pareto front usi ng RSAs obtained from a standard least squares regression. Figure 210 shows that the differences in the Pareto fronts from the unrefined design space were very large. In contrast, the Pare to fronts from the refined design space differed by less that five percent. This may be a useful tool in determining wh ether a design space is adequately refined. Table 21. Summary of DOE and surrogate modeling references. Author Application DOE Surrogate(s) Key results Han et al.64 Multiblade fan/scroll CCD Quadratic RSA RSA using CCD required no special treatment, in this case Sevant et al.86 Subsonic wing CCD RSA CCD used to ensure equal variance from center of design sp ace. RSA used to smooth numerical noise RaisRohani and Singh85 Composite structures LHS, OA, random RSA Accuracy and efficiency of RSA varies depending on choice of DOE Kurtaran et al.88 Crashworthiness Factorial design with Doptimal Linear, elliptic, quadratic RSA Difference between RSAs reduces as design space size reduces Hosder et al.89 Aircraft configuration Doptimal RSA A 30 variable optimization required design space refinement and multifidelity analysis Mahadevan et al.92 Engineering reliability analysis IMSEa 1storder RSA, Kriging For very small number of points, Kriging performed better than 1storder RSA Forsberg and Nilsson93 Crashworthiness Doptimal RSA, Kriging Kr iging performs better than RSA in this case Jin et al.94 Analytic test functions LHS RSA, MARSb, RBF*, Kriging Kriging more susceptible to noise than other surrogates Rijpkema et al.95 Analytic test functions Full factorial RSA, Kriging Kriging better captures local perturbations, but causes problems when strong noise exists
PAGE 68
68 Table 21. Continued. Author Application DOE Surrogate(s) Key results Chung and Alonso96 Supersonic business jet CCD, MPDc RSA, Kriging Kriging better at fitting when there are several local optima Simpson et al.97 Aerospike nozzle OA RSA, Kriging Kriging and RSA produce comparable results Mahajan et al.102 Radiation shield Full factorial NN Gain information on system trends and shield performance Papila et al.78 Supersonic turbine FCCD, OA, Doptimal NN, RSA NN data used to supplement CFD data for RSA Shi and Hagiwara105 Crashworthiness CCD with Doptimal RSA, NN Maximized vehicle energy dissipation aIMSE Â– integrated mean square error92 bMARS Â– multivariate adaptive regression splines124 cMPD Â– minimum point design96 Table 22. Design space refineme nt (DSR) techniques w ith their applications and key results. Author Application DSR Technique Key Results Balabanov et al.70 Transport wing Reasonable design space Unreasonable designs eliminated and surrogate accuracy improved Roux et al.73 Structural optimization Design space windowing Polynomial RSA accuracy sensitive to design space size Papila et al.82 Supersonic turbine Design space reduction Design space reduction substantially improved RSA accuracy RaisRohani and Singh85 Structural optimization Sequential response surface Successive small design spaces more efficient than large design space BosqueSendra et al.125 Chemical applications Sequential response surface BoxBehnken design used to move and resize design space while including previously used points Papila et al.77 Preliminary turbine design Design space windowing Windowing improved RSA accuracy over global RSA Jeong and Obayashi119 2D airfoil shape Smart point selection using EGO Used expected improvement uncertainty parameter for Pareto front fitness to select new design points Knowles120 Arbitrary functions Pareto front refinement using EGO (ParEGO) ParEGO more efficient than random search for Pareto front construction FarhangMehr and Azarm121 Crash analysis Smart point selection based on function irregularities Increasing point density in irregular function regions improves surrogate accuracy using minimal points
PAGE 69
69 Figure 21. Optimization framework flowchart. Problem definition and optimization setup 2. Design space refinement Refinement necessary? 3. Dimensionality reduction check 4. Multiobjective optimization using Pareto fron t Two or more confliction objectives? Find/choose optimum Design of Experiments Numerical simulations at selected locations Construction of surrogate models (model Selection and identification ) Model validation 1. Surrogate modeling yes no no yes Iterate, if necessary
PAGE 70
70 A x 1 x2 B x 1 x 2 Figure 22. DOEs for noisereduc ing surrogate models. A) Ce ntral composite design and B) face centered cubic design for two design vari ables. The extreme points are selected around a circle for CCD and a square for FCCD. A x1 x2 B x1 x2 Figure 23. Latin Hypercube Sa mpling. A) LHS with holes in the design space. B) Orthogonal array LHS can help address this issue by filling the design space more evenly.
PAGE 71
71 Original design space x 1 x 2 Refined design space Figure 24. Design space window ing showing optimum based on original design space and the final optimum based on the refined design space. actual function kriging model merit function original data points new data points Figure 25. Smart point selecti on. The final Kriging approxim ation of the actual function is shown. The new point added in each cycle is shown along with the three points at the function ends and center that comprised the original Kriging appr oximation. In this case, points are selected based on the locat ion of the minimum merit function value.
PAGE 72
72 Figure 26. Depiction of the merit functi on rank assignment for a given cluster given by Equation 238 for two objectives where min mmm nnaFF Figure 27. The effect of varying values of p on the loss function shape. A traditional least squares loss function is shown by p = 2.
PAGE 73
73 1 2 3 4 5 1.7 1.8 1.9 2 2.1 2.2 SSE for objective 1 py1 1 2 3 4 5 0.06 0.08 0.1 0.12 0.14 0.16 SSE for objective 2 py2 Figure 28. Variat ion in SSE with p for two different responses in the preliminary optimization of a radial turbine for the unref ined design space (Data Set 2). A B Figure 29. Pareto fronts for RSAs constructed with varying values of p for A) the original design space and B) the refined design space. The blue dots represent actual validation data along the Pareto front cons tructed on the refined design space using standard least squares regression.
PAGE 74
74 Figure 210. Absolute percent difference in the area under the Pareto front curves for the original feasible design space (RS 2) and the refined design space (RS 3) for various values of p as compared to a traditional least squares loss function (p = 2).
PAGE 75
75 CHAPTER 3 RADIAL TURBINE OPTIMIZATION Surrogates can be used to obtain necessary information about the design space. This information can be used to help refine the de sign space to improve the acc uracy of the surrogate model. The following case study demonstrates how response surfaces can be used to provide constraint information necessary in refining the design space to pr event infeasible CFD runs. It also shows ways in which the design space can be refined to a small region of interest by using Pareto fronts to simultaneously satisfy two objectives. A response surfacebased dualobjective de sign optimization was conducted in the preliminary design of a compact radial turbin e for an expander cycle rocket engine. The optimization objective was to increase the effi ciency of the turbine while maintaining low turbine weight. Polynomial response surface ap proximations were used as surrogates and the accuracy of such approximations improves by limiting the size of the domain and the number of variables of each response of interest. This wa s done in three stages using an approximate, onedimensional model. In the first stage, a relativel y small number of points were used to identify approximate constraint boundaries of the feasib le domain and reduce th e number of variables used to approximate each one of the constraints. In the second stage, a small number of points in this approximate feasible domain were used to identify the domain wh ere both objectives had reasonable values. The last stage focused on obtaining high accuracy approximation in the region of interest with a large number of points. The approximations were used to identify the Pareto front and perform a global sensitivity an alysis. Substantial improvement was achieved compared to a baseline design. A second study was conducted on the radial turb ine data after the first optimization study was completed. The research effort was also used to test strategies of improving the efficiency
PAGE 76
76 of the DSR process by reducing the necessary nu mber of points for use in the multiobjective optimization of computationally expensive pr oblems. These improved DSR strategies were applied to the optimization of a compact radial turbine, which was originally optimized using traditional response surface methodology and Pareto fr onts. Merit functions were used to reduce the number of points in the analysis, while ensu ring that the accuracy of the surrogate was maintained. In particular, the use of merit f unctions in a multiobjective optimization process was explored. This research seeks to explore th e benefits of using merit functions to select points for computationally expens ive problems. The purpose of th e proposed analysis procedure is twofold. First, it must be en sured that using a given selection cr iterion, such as merit functions to select the data points, result s in better surrogate accuracy th an using a seemingly more obvious method, such as simply choosing points with the best function values. Second, it must be determined whether the use of one merit function for the purpose of computationally expensive multiobjective optimization is superi or over the other merit functions. 3.1 Introduction In rocket engines, a tu rbine is used to drive the pumps that deliver fuel and oxygen to the combustion chamber. Cryogenic fuel is heated, resu lting in a phase change to a gaseous state. The increased pressure drives the turbine. Howeve r, the survivability of turbine blades limits the degree to which the fuel can be heated. There is a limited amount of h eat available from the combustion process with which to preheat the fuel, resulting in low chamber pressure and temperature. This can be an advantage because lower fuel temperatures can improve turbine reliability. However, the low chamber pressure means that the turbine work output is also limited. Turbine work can be increased in two ways: increasing the available energy in the drive gas or improving the efficiency by which the turbin e can extract the available energy. Increasing
PAGE 77
77 the available energy for a given fluid is accomplishe d by increasing the turbine inlet temperature. To increase the turbine inlet temp erature in an expander cycle, a higher heat flux from the thrust chamber to the cooling fuel is needed. Obtaining this higher heat flux is problematic in several ways. First, materials and manuf acturing development is necessary to produce a thrust chamber with high heat flux capability. This work is an ongoing area of tec hnology development. Second, to enable a higher heat exchange, in creased surface area and co ntact time between the thrust chamber and cooling fuel is needed. Th ese requirements lead to a larger, and heavier, thrust chamber. In addition, significantly raising the turbin e inlet temperature defeats the expander cycleÂ’s advantage of maintaining a be nign turbine environment. Immich et al.126 reviewed methods of enhancing heat transfer to the combustion chamber wall in an expander cycle. The authors tried three methods of enhanc ing heat flux to the unburned fuel: 1) Increase the length of the combustion chamber, 2) in crease the combustion chamber wall surface area by adding ribs, and 3) increase the combusti on chamber wall roughness. The authors also mentioned that in the future they would expe rimentally investigate the influence of the distribution of the injector elem ent distribution and the effect of the distance of the injector element to the wall on the heat transfer to the wall. However, results of the future analysis are not available. The second approach to increasing turbine work is to improve turbine efficiency. If the turbine inlet temperature is held constant, an increase in turbine work is directly proportional to efficiency increase. If the required work can be achieved with moderate efficiency, an improvement in that efficiency can be traded for reduced inlet temper atures, providing better design environment margins. One way to improve tu rbine efficiency is to use a radial turbine. Radial inflow turbines perform better than axial turbines at high velocity ratios, exhibit better
PAGE 78
78 tolerance to blade incide nce changes, and have lo wer stresses than axial de signs. Radial turbines have been used successfully in automotive applications, but are not often used in rocket engines due to their relatively large size and weight. The size of the comp act radial turbines makes them applicable to rocket engine cycles wh en a high velocity ratio is involved. This research will focus on improving turbine efficiency. The radial turbine design must provide maximum efficiency while keeping the overall weight of the turbine low. This necessitates a multiobjective optimization. A response surface analysis46 provides an efficient means of tackling the optimization problem. The research presented here represents the preliminary optimization of a radial turbine using a simplified 1D radial turbine model adapted from the 1D Meanline127 code utilized by Papila et al.82 The Meanline code provides performance and geometry predictions based on selected input conditions. It is an approximate and inexpensive model of the actual processes. Because it is an approximate model, there is some degree of uncertainty involved, but it can provide a good starting poin t in the design process. Using response surface analysis, an accurate surrogate model was constructed to predict the radial turbine weight and the efficiency ac ross the selected design space. The surrogate model was combined with a genetic algorithmbas ed Pareto front construction and facilitates global sensitivity evaluations. Because the radial turbine represented a new design, the feasible design space was initially unknown. Techni ques including design constraint boundary identification and design space reduction were nece ssary to obtain an acc urate response surface approximation (RSA). The analysis used the optimization framewor k outlined in Chapter 2. The framework steps included in the analysis are 1) modeling of the objec tives using surrogate models, 2) refining the design space, 3) reduc ing the problem dimensionality, and 4) handling
PAGE 79
79 multiple objectives with the aids of Pareto fr ont and a global sensitivity evaluation method. Finally, the ability of Merit Func tions to improve the efficiency of the optimization process was demonstrated. 3.2 Problem Description The radial turbine performance was simulate d using the 1D Meanli ne code. Using a 1D code allowed for the availability of relatively inexpensive computations To determine whether using the 1D code was feasible for optimi zation purposes, a 3D ve rification study was conducted. Once the Meanline code was verified, th e radial turbine optimi zation could proceed. 3.2.1 Verification Study Threedimensional unsteady Navi erStokes simulations were performed for the baseline radial turbine design, and the pr edicted performance parameters were compared with the results of the Meanline analysis. Simulations were perf ormed at three rotational speeds: the baseline rotational speed of 122,000 RPM, a low speed of 103,700 RPM and a highspeed of 140,300 RPM. A 2vane/1rotor model was used. The simulations were run with and without tip clearance, and the computational grids contained approximately 1.1 million grid points. The PHANTOM code was used to pe rform the numerical simulations.128 The governing equations in the PHANTOM code are the threedi mensional, unsteady, NavierStokes equations. The equations have been written in th e Generalized Equatio n Set (GES) format,129 enabling it to be used for both liquids and gases at operating conditions ra nging from incompressible to supersonic flow. A modified BaldwinLomax tur bulence model is used for turbulence closure.130 In addition to the perfect gas approximation, the code contains two options for the fluid properties. The first option is based on the equations of state, thermodynamic departure functions, and corresponding state pr inciples constructed by J. C. Oefelein at Sandia Corporation in Livermore, California. The second option is based on splines generated from the NIST
PAGE 80
80 Tables.131 A detailed description of th e code/algorithm development, as well as its application to several turbine and pump te st cases, is presented in Venkateswaran and Merkle129 and Dorney et al.132 Figure 31 shows static pressu re contours (psi) at the midhe ight of the turbine for the baseline rotational speed of 122,000 RPM. This figur e illustrates the geometry of the turbine, and the contours indicate that the pressure decrease is nearly evenly divided between the vane and the rotor. In fact, the reaction was approximately 0.60 for each of the th ree 3D simulations as compared to 0.55 for the 1D simulation. Figur e 32 contains the predicted totaltostatic efficiencies from the Meanline and CFD analys es. The CFD results include values with and without tip clearance. In genera l, fair agreement is observe d between the Meanline and CFD results. The trends are qualitatively similar, but the Meanline analysis predicts higher efficiencies. There is approxima tely a fourpoint difference in the quantitative values. The quantitative differences in the results are not surp rising considering the lack of experimental data available to anchor the Meanline code. The differences in the pr edictions with and without tip clearance decrease with increasing rotational spee d. Figure 33 shows the predicted work from the Meanline and CFD analyses. The trends are again similar between the Meanline and CFD analyses, but the Meanline values are consiste ntly 5 Â– 6% higher than the CFD values. The similar trends between the 1D Meanline code and 3D CFD analys es indicate that the optimization can be confidently performed on the 1D Meanline code. It can be expected that for a given turbine speed the Meanline code will overpredict the totaltostatic efficiency by an expected degree. The predicted optimum point ba sed on the 1D Meanline code will likely yield overly optimistic results, but the predicted degr ee of improvement should translate to the 3D CFD analysis.
PAGE 81
81 3.2.2 Optimization Procedure The purpose of the design optimization is to maximize the turbine efficiency while minimizing the turbine weight. A total of six desi gn variables were identifie d. The ranges of the design variables were set based on current design practices. Additionally, five constraints were identified. Two of the five cons traints are structural constraints, two are geometric constraints, and one is an aerodynamic constraint. The aerodynamic constraints are based on general guidelines. The descriptions of a ll objectives, variables, and constr aints are given in Table 31. It is unknown in what way the constraints depend on the design variable s. It is possible that certain combinations of desi gn variables will cause a constraint violation. It is also unknown whether the selected ranges resu lt in feasible designs. The res ponse surface analysis can help clarify these unknown factors. The design of experiments (DOE) procedure was used to select the location of the data points that minimize the effect of noise on the fitted polynomial in a response surface analysis. A facecentered cubic design was used to generate a total of 77 data points within the selected ranges. 3.3 Results and Discussion 3.3.1 Phase 1: Initial Design of Experiments a nd Construction of Constraint Surrogates The values of the objective functions were obt ained using the Meanline code. Of the 77 solutions from the initial DOE, seven cases failed and 60 cases violated one or more of the five constraints, resulting in only 10 successful cases. Before the op timization could be conducted, a feasible design space needed to be identified. Because there was limited information on the dependencies of the output constraints on the de sign variables, response surface analyses were used to determine these dependencies. Response su rfaces were used to properly scale the design variable ranges and identify ir regular constraint boundaries.
PAGE 82
82 Quadratic response surfaces were fit to the output constraint s. The change in each response surface with respect to changes in some of design variables was small, indicating that the response surfaces were largely insensitive to ce rtain design variables. The accuracy of the response surfaces was not compromised when the identified design variables were neglected. Therefore, for each response surf ace, the effects of va riables that contribu ted little to the response surface were removed. In this wa y, each response surface was simplified. The simplified dependencies are shown in Equation 31. 22 11,, ,, ,, ANANAnsqrFrac TipSpdTipSpdU/Cisen Cx2/UtipCx2/UtipRPMU/CisenAnsqrFrac ReactU/CisenTipFlw Rsex/RsinRsex/RsinAnsqrFracU/CisenDhex% (31) The information obtained from the response surfaces about the output constraint dependences were further used to develop c onstraints on the design va riables. A quadratic response surface was constructed for each design va riable as a function of the output constraint and the remaining design variables, for example, 1,, R eactReactU/CisenTipFlw (32) The most accurate response surfaces (R2 adj 0.99) were used to dete rmine the design variable constraints. The output c onstraints were in turn set to th e constraint limits. For example, a constraint on React was applied to coincide with the constraint 1 0: 11 1,,0 0,, ReactU/CisenTipFlw R eactReactU/CisenTipFlw (33) Constraint boundary approximati ons were developed in this manner for each constraint. As can be seen from Equation 31, two of the five constraints (AN2 and Tip Spd) were simple limits on a single variable. For the single variable constraints, the variable ranges were simply
PAGE 83
83 reduced to match the constraint boundaries. Th e remaining constraints were more complex. However, one of the complex constraints (Rsex/Rsin) was automatically satisfied by the reduction of the variable ranges for the single vari able constraints. The f easible region is shown in Figure 34. The remaining constraints involved three vari ables each. It was discovered that many low values of RPM violated the Cx2/Utip constraint. The region of violation was a function of RPM, U/C isen, and AnsqrFrac as shown in Figure 35. The 1 constraint was found to be the most demanding and resulted in a feasib le design space as shown in Figur e 36. Much of the original design space violated this cons traint. It was also discovere d that the constraint surface representing the bounds for 1 40 lay outside of the origin al design variable range for React. In this case, the lower bound for React was sufficient to satisfy this constraint. The predictive capability of the constraints wa s tested using the available data set. The design variable values were input into the RSAs for the constraints. Using the RSAs, all points that violated the output constraints were correc tly identified. Now that the feasible design space was accurately identified, data points could be placed in the feasible data region. The results and summary of the prediction of constr aint violations are as follows: 1. Quadratic response surfaces we re constructed to determine the relationship between the output constraints and the design variables. 2. The variable ranges were adjusted based on information from the constraint surfaces. 3. A 3level factorial design (729 points) wa s applied within new variable ranges. 4. Points that violated any cons traints (498 / 729 points) were eliminated based on RSAs of constraints. Using the response surface constraint appr oximations, 97% of the 231 new data points predicted to be feasible lay w ithin the actual feasib le design space region after simulation using the Meanline code. The point s that were predicted to be feas ible, but were actually found to be
PAGE 84
84 infeasible, often violated the 1 0 requirement by a slight amount. The resulting data set contained 224 data points. 3.3.2 Phase 2: Design Space Refinement Plotting the data points in function space re vealed additional information about the location of the design space as shown in Figure 37. A large area of function space contained data points with a lower efficiency than was de sired. There also existed areas of high weight without improvement in efficiency. Additionally the fidelity of the response surface for ts was apparently compromised by the existen ce of a design space that was too large. These undesirable areas could be eliminated, and the response surf ace fidelity could be improved by refining the design space. The density of points could then be in creased within the region of interest, eliminating the possibility of unnecessary invest igation of undesirable points. The region of interest is shown in Figure 37. To further test the necessity of a reduced desi gn space, five RSAs each were used to fit the data for Wrotor and ts in the original feasible design space. These response surfaces were constructed using the general loss function given in Equation 242 for p = 1Â…5. The RSA constructed using the least square loss function (p = 2) was used as a reference point. As seen in Figure 38, regardless of the RSA used, the error in the RSA at the data points is high at high Wrotor and low ts. To improve RSA performance, additiona l data points could be added in these regions, or these data regi ons could be eliminated. Pareto fronts were constructed for each RSA se t. The results are show n in Figure 39. In this case, the Pareto fronts differ by as much as 20%. Because the results differ significantly depending on which RSA is used, this also indi cates that further design space refinement is necessary.
PAGE 85
85 The design variable bounds were further redu ced to match the new design space. The new design variable ranges are given in Table 33. Quadratic response surfaces were constructed for the turbine weight, Wrotor, and the turbine totaltostatic efficiency, ts, using the original feasible design space to screen data points. Points pred icted to lie outside of the refined design space would be omitted. For the refined design space, a third set of data was required. As seen in Figure 37, using a factorial desi gn tended to leave holes in functi on space. It was possible that this could hamper construction of an accurate Pareto front. To prevent this, Latin Hypercube Sampling was used to help close possible holes for the third data set. The points were efficiently distributed by maximizing the mini mum distance between any points. Of the best points from the second data set, it only RPM, Tip Flw, and U/C isen varied, while React and Dhex% remained constant at their lowest values, and AnsqrFrac remained constant at its maximum value among these points. To ensure that this effect was captu red in the third data set, additional points were added using a 5level factorial de sign over these three variables. The remaining variables were held constant according the values observed in the best tradeoff poi nts. Quadratic response surfaces previously constructed for the turbine weight, Wrotor, and the turbine total to static efficiency, ts, were used to screen the potential data poi nts. Points predicted to lie outside of the newly refined design space would be omitted from the analysis. In summary, 1. Only the portion of design space with best performance was reserved to allow for a concentrated effort on the region of interest and to increase response surface fidelity. 2. Latin Hypercube Sampling (181 / 300 feasible points) was used over all six variables and was supplemented by a 5level factorial design used over RPM, Tip Flw, and U/C isen (119 / 125 feasible points) to improve re solution among the best tradeoff designs. 3. Points that were predicted to violate constrai nts or lie outside of re gion of interest were omitted. The combination of the DOEs resulted in a total of 323 feasible design points.
PAGE 86
86 3.3.3 Phase 3: Construction of the Pareto Fr ont and Validation of Response Surfaces As in the second data set, five RSAs each with p = 1Â…5 were used to fit the data for Wrotor and ts for the third data set. Pareto fronts were constructed for each RSA set and are shown in Figure 310. In this case, the Pareto fronts differed by a maximum of only 5%. Because the difference in the response surfaces for varying values of p is small for the third data set, the design space was determined to be adequately refined. Function evaluations from the quadratic response surfaces (p = 2) were used to construct the Pareto Front shown in Figure 311. Within the Pareto front, a region was identified that would provide the best value in terms of maxi mizing efficiency and minimizing weight. This tradeoff region was selected for the validation of the Pareto front. The results of the subsequent validation simulations indicated that the respon se surfaces and corresponding Pareto front were very accurate. A notable improvement was attain ed compared to the baseline radial turbine design. The design selected optimum design had the same weight (Wrotor) as the baseline case with approximately 5% improvement in effici ency. The specifications for the optimum design are given in Table 34. Within the best tradeoff region, only RPM and Tip Flw vary along the Pareto front as seen in Figure 312. The other variables are constant within the tradeoff region and are set to their maximum or minimum value. This indicates that increasing the range of one of these variables might result in an increase in performan ce. The minimum value of the variable React was chosen as the only variable range that could reas onably be adjusted. The validation points were simulated again using a reduced React value. Reducing the minimum value of React from 0.45 to 0.40 increased the maximum efficiency only for Wrotor > 1. The maximum increase in efficiency improved from 4.7% to 6.5%, but this increase occu rred outside of the preferred tradeoff region.
PAGE 87
87 It was further determined that using a React value of 0.40 resulted in a turbine design that was too extreme. Based on the results of the current study, seve ral observations can be made concerning the design of highperformance radial turbines. 1. The more efficient designs have a higher velo city ratio based on the rotational speed as shown by U/C isen in Figure 312. This is a re sult of the inlet blad e angle being fixed at 0.0 degrees due to structural considerations. Classic radial turbine designs have velocity ratios of approximately 0.70. 2. The more efficient designs have a smaller tip radius and larger bl ades (AnsqrFrac) as compared to designs with a larger radius and smaller blades. Incr easing the annulus area leads to higher efficiencies. In this case, th e annulus area (AnsqrFrac) is set to the highest value allowable while respecting stress limitations. 3. The higher performing designs have higher rota tional speeds (as compared to the original minimum RPM value of 80,000 given in Table 31) as a result of th e smaller radius and large blades. The higher rotational speeds also lead to more efficient pump operation. 3.3.4 Phase 4: Global Sensitivity Analysis and Dimensionality Reduction Check A global sensitivity analysis was conducted using the response surface approximations for the final design variable ranges gi ven in Table 34. The results ar e shown in Figure 313. It was discovered that the tu rbine rotational speed RPM had the largest impact on the variability of the resulting turbine weight Wrotor. The effects of th e rotational speed RPM along with the isentropic velocity ratio U/C isen make up 97% of the variability in Wrotor. All other variables and variable interactions have minimal effect on Wrotor. For the total to static efficiency, ts, the effects of the design variables are more evenly distributed. The reaction variable, React, has the highest overall impact on ts at 28%. This information is useful for fu ture designers. For future designs, it may be possible to eliminat e all variables except RPM and U/C isen when evaluating Wrotor, whereas for ts, it may be necessary to keep all variables.
PAGE 88
88 It must be noted that the local variability at any given point can vary significantly from the global sensitivity values due to the nonlinear natu re of the RSAs. However, the global sensitivity analysis can be scaled to the region of in terest to explore more localized effects. 3.4 Merit Function Analysis Merit functions were used in an attempt to reduce the number of points in the DSR analysis as compared to the orig inal radial turbine optimization st udy. They were used to select data points such that the accura cy of the surrogate model can be improved using a minimal number of data points. The merit functions rely on predictio ns of the function values and variance and attempt to balance the effects of th e two. Kriging was selected as the surrogate model for the optimization. The merit function analysis demonstrates how the merit functions, facilitated by Kriging, can be integrated into the DSR process. 3.4.1 Data Point Selection and Analysis First, a Kriging model is fit to the 231 points that comprised the feasible design space constructed as the result of the first cycle in th e optimization process. This feasible design space is designated DS1. Next, the 323 points from the refined design space constructed during the second cycle of the optimization process are used as a databank. This reasonable design space is designated DS2, but no function valu es are initially made at these points. Out of the databank of data point locations from DS2, points are clus tered based on proximity, and the point from each cluster that is predicted to have the best charac teristics based on merit func tion values is selected for function evaluation. It is a ssumed that the points in DS2 lay very near the suspected optima. The procedure used in the anal ysis is given as follows: 1. Fit Kriging model to points in the feasib le design space (DS1) for each objective. 2. Using Kriging model from DS1, predict function valueÂˆ yand prediction variance V(x) at each point in reasonable design space (DS2) for each objective.
PAGE 89
89 3. Calculate merit function values at each point in DS2 based on Eqs. (233) through (237). 4. Normalize variables by the design variable range and cl uster the data points from DS2 into m groups, based on proximity. 5. Select 6 sets of m points each: a. Choose the data points with the best rank R from each cluster and predict objective value at those points for each of the four merit functions. b. Choose the points with the best rank R based on the minimum function value fmin from each cluster using Equation 238. c. Choose the data points that lie cl osest to the cluster centers. 6. Evaluate objectives at each point in each of the six data sets. 7. Fit Kriging model to the points in each of the six data sets. 8. Use Kriging model to predict objective values along Pareto front for each set. 9. Calculate the errors in the objec tive values for each scenario. 10. Repeat steps 1 through 9 for 100 times such th at the random variati on that occurs when clusters are selected sets is minimal. 11. Compare accuracy of different selection criteria among the 6 selection criteria. 12. Repeat steps 1 through 11 using m = 20, 30, 40, 50 clusters (with one data point selected per cluster) to check sensitivity to number of data points. 13. Compare accuracy among the 6 selection criteria for different numbers of data points. 3.4.2 Merit Function Comparison Results The points along the validated Pareto front gi ven by the blue curve in Figure 311 were predicted by each of the six new Kriging approximations based on MF1, MF2, MF3, MF4, fmin, or the cluster centers. The differences were small for data sets containing 50 data points, so the results shown in Figure 314 and Fi gure 315 are illustrated using the results of the smallest data set. The results for a selected cluster set a nd 20 data points are shown in Figure 314. Although only 20 points were selected by maximizing the probability of improvement (MF2) and used to generate the Kriging model, the predicted points only vary sligh tly from the actual values along
PAGE 90
90 most of the Pareto front. This indicates that the 20 data points selected by MF2 actually lie on or very near the Pareto front. Because Kriging inter polates the data points, the prediction of points near the 20 data points is very good which leads to the construction of an accurate Pareto front. It is likely that the pr ediction of points away from the Pare to front would be very poor. Other selection criteria were not as successful in predicting points along the Pareto front. The absolute scaled error at each point is calculated as maxminÂˆscaledyy e yy (34) and the absolute combined error is given as ,1 2rotortscombinedscaled Wee (35) The error distribution for data points along the Pare to front for each selection criterion is given in Figure 315. The use of MF2 results in the lowest combined e rror. Thus, it is demonstrated qualitatively that se lecting points using MF2 results in the highest accuracy for data sets with a low number of points. Due to the random nature of the Matlab function kmeans that is used to generate the clusters, the cluster sets can be slightly different depending on the starting point of the search, especially if the data set being clustered is well distributed. For this r eason, 100 different cluster sets were used to reduce effects due to cl uster selection. The mean and maximum error distributions were compared for various numbers of data points and th e different selection criteria and are shown in Figure 316 and Figure 318, respectivel y. Figure 317 and Figure 319 provide direct comparisons of a ll selection criteria us ing the median values from Figure 316 and Figure 318, respectively. By usi ng the probability of improvement MF2 as a selection criterion, the overall accuracy was maintained to a surprising degree. Except for MF1, using merit
PAGE 91
91 functions as selection criteria a ppeared to result in better accuracy than simply selecting the point with the minimum function value or centers from each cluster. Using MF1 gave results that were very similar to using fmin. When the maximum error was considered, however, all merit functions except MF2 resulted in higher maximum errors as the number of points decreased. Selecting points based on maxi mizing the probability of impr ovement resulted in the best performance given a small number of data points as compared to simply choosing the minimum function value or cluster centers. The results in the use of the remaining merit functions were mixed. Maximizing the probability of improveme nt, maximizing the expected improvement, or minimizing the expected value of the minimum f unction value resulted in lower average errors than selection based on minimi zing the statistical lower bound, minimum function value, or cluster centers. However, simply selecting th e cluster centers seemed to limit the maximum error better than any merit function except maxi mizing the probability of improvement. In this analysis, it was possible to reduce the number of points in the final op timization cycle by 94% while keeping the accuracy of points along the Pare to optimal front within 10% with an average error of only 3% or less. 3.5 Conclusion An optimization framework can be used to f acilitate the optimization of a wide variety design problems. The liquidrocket compact radial turbine analysis demonstrated the applicability of the framework: Surrogate Modeling. The radial turbine optimization process began without a clear idea of the location of the feasible design region. RSAs of output constraints were successfully used to identify the feasible design space. Design Space Refinement. The feasible design space wa s still too large to accommodate the construction of an accurate RSA for the prediction of turbine efficiency. A reasonable design space was defined by eliminating poor ly performing areas thus improving RSA fidelity.
PAGE 92
92 Dimensionality Reduction. A global sensitivity analysis provided a summary of the effects of design variables on objective variab les, and it was determined that no variable could be eliminated from the analysis. Multiobjective Optimiza tion using Pareto Front. Using the Pareto front information constructed using genetic algor ithms, a best tradeoff region was identified within which the Pareto front and the response surfaces used to create the Pareto front were validated. At the same weight, the RS optimization result ed in a 5% improvement in efficiency over the baseline case. Merit Functions. A methodology was presented to use merit functions as a selection criteria to reduce the number of points requ ired for multiobjective optimization. Using merit functions gives the ability of dramatical ly reducing the number of points required in the final cycle of multiobjective optimization. Through this study, a number of aspects from the framework were demonstrated, and the benefits of the various steps were made a pparent. The framework provides an organized methodology for attacking several issues that arise in design optimization. Table 31. Variable names and descriptions. Objective variables Description Baseline design Wrotor Relative measure of Â“goodnessÂ” for overall weight 1.147 ts Totaltostatic efficiency 85% Design variables MIN Baseline MAX RPM Rotational Speed 80,000 122,000 150,000 React Percentage of stage pressure drop across rotor 0.45 0.55 0.70 U/C isen Isentropic velocity ratio 0.50 0.61 0.65 Tip Flw Ratio of flow parameter to a choked flow parameter 0.30 0.25 0.48 Dhex % Exit hub diameter as a percent of inlet diameter 0.10 0.58 0.40 AnsqrFrac Used to calculate annulus area (stress indicator) 0.50 0.83 1.0 Constraints Desired range Tip Spd Tip speed (ft/sec) (stress indicator) 2500 AN2 Annulus area speed2 (stress indicator) 850 1 Blade inlet flow angle 0 1 40 2C/Utip Recirculation flow coefficient (indication of pumping upstream) 0.20 Rsex/Rsin Ratio of the shroud radius at the exit to the shroud radius at the inlet 0.85
PAGE 93
93 Table 32. Response surface fit statistics before (feasible DS) and afte r (reasonable DS) design space reduction. Feasible DS Reasonable DS Wrotor ts Wrotor ts R2 0.987 0.9170.9960.995 R2 adj 0.985 0.9050.9960.994 Root Mean Square Error 0.094 0.0200.02350.00170 Mean of Response 1.04 0.7711.040.844 Observations 224 224310310 Table 33. Original and final de sign variable ranges after constr aint application and design space reduction. Original ranges Final ranges Design variable Description MIN MAX MIN MAX RPM Rotational Speed 80,000150,000 100000 150,000 React Percentage of stage pressure drop across rotor 0.450.68 0.45 0.57U/C isen Isentropic velocity ratio 0.50.63 0.56 0.63Tip Flw Ratio of flow parameter to a choked flow parameter 0.30.65 0.3 0.53Dhex% Exit hub diameter as a % of inlet diameter 0.10.4 0.1 0.4AnsqrFrac Used to calculate annulus area (stress indicator) 0.50.85 0.68 0.85 Table 34. Baseline and optimum design comparison. Objectives Description Baseline Optimum Wrotor Relative measure of Â“goodnessÂ” for overall weight 1.147 1.147 ts Totaltostatic efficiency 85.0% 89.7% Design Variables Baseline Optimum RPM Rotational Speed 122,000 124,500 React Percentage of stage pressu re drop across rotor 0.55 0.45U/C isen Isentropic velocity ratio 0.61 0.63Tip Flw Ratio of flow parameter to a choked flow parameter 0.25 0.30Dhex % Exit hub diameter as a % of inlet diameter 0.58 0.10AnsqrFrac Used to calculate annulus area (stress indicator) 0.83 0.85
PAGE 94
94 Figure 31. Midheight static pressure (psi ) contours at 122,000 rpm. 0.74 0.76 0.78 0.8 0.82 0.84 0.86 0.88 0.9 0.92 100000105000110000115000120000125000130000135000140000145000150000RPMTotaltoStatic Efficiency Meanline CFD No tip clearance CFD W/tip clearance Figure 32. Predicted Meanline and CFD totaltostatic efficiencies. 1300.0 3200.0 100 110 120 130 140 150 0.92 0.9 0.88 0.86 0.84 0.82 0.8 0.78 0.76 0.74 TotaltoStatic Efficiency Meanline CFD Â– No tip clearance CFD Â– W/ti p clearance 103 RPM
PAGE 95
95 220 230 240 250 260 270 280 290 100000105000110000115000120000125000130000135000140000145000150000RPMWork (BTU/lbm) Meanline CFD No tip clearance CFD W/tip clearance Figure 33. Predicted Mean line and CFD turbine work. Feasible AN2 > 850 Tip Spd < 2500 Rsex/Rsin < 0.85 Infeasible AN 2 < 850 R se x /Rsin > 0.85 Tip Spd > 2500 U/C isen A ns q rFrac Figure 34. Feasible region and location of three constraints. Three of five constraints are automatically satisfied by the rang e reduction of two design variables. 290 280 270 260 250 240 230 220 100 110 120 130 140 150 Work ( BTU/lbm ) Meanline CFD Â– No tip clearance CFD Â– W/tip clearance 103 RPM
PAGE 96
96 Infeasible Cx2/Utip < 0.2 Feasible Cx2/Utip > 0.2 U/C isen AnsqrFrac RPM 105 Figure 35. Constraint surface for Cx2/Utip = 0.2. At higher values of AnsqrFrac and U/C isen, lower values of RPM are infeasible. Figure 36. Constraint surfaces for 1 = 0 and 1 = 40. Values of 1 > 40 lay outside of design variable ranges.
PAGE 97
97 Approximate region of interestNote: Maximum Etats 90% Approximate region of interestNote: Maximum Etats 90% Approximate region of interestNote: Maximum Etats 90% Figure 37. Region of interest in function space. (The quantity 1 Â– ts is used for improved plot readability.) A 0 0.2 0.4 0.6 0.8 1 1.2 012345 Wrotorerror p = 2 p = 5 p = 1 Figure 38. Error between RSA and actual data point. A) Wrotor and B) ts at p = 1, 2, and 5. Wroto r vs. ts Wroto r 1 Â– ts
PAGE 98
98 B 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.550.60.650.70.750.80.850.90.95ts error p = 2 p = 5 p = 1 Figure 38. Continued. Figure 39. Pareto fronts for p = 1 through 5 for second data set. (The quantity 1 Â– ts is used for improved plot readability.) Pareto fronts differ by as much as 20%.
PAGE 99
99 Figure 310. Pareto fronts for p = 1 through 5 for third data set. (The quantity 1 Â– ts is used for improved plot readability.) Pareto fronts differ by no more than 5%. Figure 311. Pareto Front with validation data. Deviations fr om the predictions are due to rounded values of the design variables (predi ction uses more signi ficant digits). The quantity 1 Â– ts is used for improved plot readability.
PAGE 100
100 Figure 312. Variation in desi gn variables along Pareto Front. A Other < 1%Dhex % < 1% Tip Flw 1% U/C isen*RPM 1%U/C isen 7% RPM 90%B Dhex % 3% AnsqrFrac 5% RPM 15% React 28% U/C isen 21% Tip Flw 23% React*RPM 2% Other 3% Figure 313. Global sensitivity analysis. Effect of design variables on A) Wrotor and B) ts.
PAGE 101
101 Figure 314. Data points predicte d by validated Pareto front compared with the predicted values using six Kriging models based on 20 selected data points. The validated Pareto front is labeled Â“Actual.Â” Figure 315. Absolute error distribution for poi nts along Pareto front using 20 selected data points each where the points were selected using (1) MF1, (2) MF2, (3) MF3, (4) MF4, (5) fmin, and (6) cluster centers.
PAGE 102
102 Figure 316. Average mean error distribution over 100 cl usters for (1) 50 point s, (2) 40 points, (3) 30 points, and (4) 20 points. As number of points decreases, merit functions 2, 3, and 4 perform better than MF1, fmin, or cluster centers with MF2 showing the best performance.
PAGE 103
103 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 10 20 30 40 50 Number of PointsMedian Average Mean MF 1 MF 2 MF 3 MF 4 fmin centers Figure 317. Median mean error over 100 clusters Merit function 2 shows the best performance as the number of points decreases. Figure 318. Average maximum error distribu tion over 100 clusters for (1) 50 points, (2) 40 points, (3) 30 points, and (4) 20 points.
PAGE 104
104 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 10 20 30 40 50 Number of PointsMedian Average Max MF 1 MF 2 MF 3 MF 4 fmin centers Figure 319. Median maximum error over 100 cl usters. Merit function 2 shows the best performance as the number of points decreases.
PAGE 105
105 CHAPTER 4 MODELING OF INJECTOR FLOWS The combustion chamber of a liqui d rocket engine is cooled by transferring heat to the unburned fuel that circulates ar ound the outside of the combustion chamber via a series of tubes or coolant channels. Rocket injectors deliver th e fuel and oxidizer to the combustion chamber. Commonly, the injector face is made up of a series of injector elements arranged in concentric rows. The row of injectors near the chamber wall can cause considerable local heating that can reduce the life of the combustion chamber. The local heating near each inj ector element takes the form of a sinusoidal wall heat flux profile cause d by the interactions of the outer row elements. Often, this effect is not consid ered during the design process. Reduc ing the intensity of this local heating is of prime importance. For example, in the design of the Space Shuttle Main Engine (SSME), the effect of the local heating in the combustion chamber was not considered during the design process. After construc tion, it was discovered that hotspots along the chamber wall severely reduced the expected chamber life. This resulted in an unforeseen increase in the reusability operating costs of the engine. Local hotspots can even cause actual burnout of the chamber wall. By accurately predicting potentially detrimental phenomena in advance, it may be possible that issues such as wa ll burnout can be avoided. CFD m odeling and validation efforts, in conjunction with the experimental data, can a ssist in the understanding of combustor flow dynamics, eventually leading the way to efficient CFDbased design. This chapter outlines current and proposed methodologies in the CFDbased optimization of liquid rocket engine components. First, a summary is presente d of past and present injector analysis techniques. Then, the basic governing equa tions are presented that are applicable in the simulation of turbulent reacting flow. Fina lly, a simplified onedimensional analysis is
PAGE 106
106 demonstrated to help define the relationship betw een flow parameters and heat transfer due to reacting flow. 4.1 Literature Review Injector research has been ongoing for over 40 year s. Injector inlet fl ow has a significant influence on combustor performance. The understandi ng of the injector charac teristics is critical in determining the nature of the flow within the combustion chamber. In particular, the inlet flow geometry and injector outlet diameter ha ve large influences on flow in the combustion chamber.26 However, at times, the research has raised more questions than answers. Work must still be done to explain certain elements of the flow within the combustion chamber. A general diagram of injector flow is given in Figure 41. Historically, injectors have been de signed using experi mental techniques133 and empirical calculations. A design was built and tested, and then improvements were made based on the results. For example, Calhoon et al.134 extensively reviewed standa rd techniques for injector design, including the analysis of coldfire and hotfire test ing to study general injector characteristics. The results of these tests we re used along with a number of additional multielement coldfire tests to design th e full injector. The full injector was then fabricated and tested for performance and heat flux characteristic s and combustion stability. However, the experimental design techniques were insufficient to predict for many conditions that could reduce injector or combustion chamber life. 4.1.1 SingleElement Injectors A significant portion of injector research has been conducted using experiments consisting of a single injector element. The single elemen t analysis is often used as a starting point in modeling full combustor flow. A test firing of a single element injector is shown in Figure 42. Hutt and Cramer135 found that if it is assumed that all of the injectors in the core are identical,
PAGE 107
107 then the measure of energy release from a single elementÂ’s flow field is a good approximation of the efficiency of the entire core. In particular the mixing characteristics of a single injector can be used to determine the element pattern and spacing needed fo r good mixing efficiency. Singleelement injectors are often used in in jector experimentation to approximate the mixing effects of the full injector. Calhoon et al.134 tested a multiple element injector chamber such that the radial spacing of the elements coul d be varied. In cold flow testing, it was found that mixing increased dramatically when multiple elements were used. However, it was determined that the element spacing only had a sli ght influence on the mixing efficiency. Yet, as the distance between the elements increased, the mixing efficiency increased. This was due to large recirculation regions that br ought flow from the wellmixed fa r region to the near field. When recirculation effects are compensated for, Calhoon et al. found that there is no significant effect of multiple element interactions compared to single element mixing. This, in theory, indicates that a satisfactory combustion analysis can be done using only a single injector. 4.1.2 MultiElement Injectors A multielement injector face is made up of an array of injector elements, and usually contains from seven to hundreds of injector elements as shown in Figure 43. Small changes in the design of the injector and the pattern of el ements on the injector f ace can significantly alter the performance of the combustor. Elements mu st be arranged to maximize mixing and ensure even fuel and oxidize r distribution. Gill26 found that the element diam eters and diameter ratios largely influence mixing in the combustion chambe r, and that small diameters lead to overall better performance. The type of elements need not be consistent across the entire injector face. The outer elements must be chosen to help provide some wall cooling in the combustion chamber. Gill suggests that using a coaxial type injector for the outer row of injectors provides an ideal nearwa ll environment.
PAGE 108
108 Heating effects on the combustion chamber wa ll due to the arrangement of injector elements is of prime importance. Injector placement can result in high local heating on the combustion chamber wall. Figure 44 shows the e ffect of local heating that resulted in the burnout of an uncooled combusti on chamber. Rupe and Jaivin136 found a positive correlation between the temperature profile al ong the wall and the placement of injector elements. Farhangi et al.137 investigated a gasgas injector and measur ed heat flux to the combustion chamber wall and injector face. It was found that the mixing of the propellants controlled the rate of reaction and heat release. Farhangi et al. suggested that the injector element patte rn could be arranged in a way that moved heating away from the injector face by delaying the mixing of the propellants. 4.1.3 Combustion Chamber Effects and Considerations Several studies have suggested that pressu re has little influence on the combustion dynamics. Branam and Mayer138 studied turbulent length scales by injecting cr yogenic nitrogen through a single injector to help in understanding cryoge nic rocket propellant mixing efficiency. It was found that changes in pres sure and injection velocity ha d very little effect on flow dynamics. Similarly, Calhoon et al. 134 found that the pressure effects on chamber heat flux were very small. Quentmeyer and Roncace139 found no change in the heat flux data with a change in pressure when calculating the heat flux to the wall of a plug flow calor imeter chamber. In addition, Wanhainen et al.140 found that changing the chamber pressure had no effect on combustion stability. While, Mayer et al.141 found that pressure is re latively constant throughout the combustion chamber, while Moon142 asserts that even the very small pressure gradients that exist near the injector exist can significantly al ter velocity profiles and mixing in the combustion chamber. The relatively constant pressure indi cates that density gradients in the combustion chamber must be due to temperature, rather than pressure. Thus, the flow can be considered incompressible. In addition to pressure independence, Preclik et al.143 experimentally measured
PAGE 109
109 heat fluxes alon the combustion chamber wall. It was found that the wa ll heat fluxes were largely independent of the mixtur e ratio in the analysis of inj ection patterns and velocities on wall heat flux when the oxygen mass flow ra te was held constant. Conley et al.144 also discovered a negligible influence on wall heat flux due to the mixture ratio. Several analyses consistently found that there is a sharp jump in the heat flux a certain ways down the combustion chamber wall. The expl anations for the jump have been varied. Quentmeyer and Roncace139 found a sharp jump in heat flux 3 cm from the injector face. They took this to mean that at this point, the reactants had mixed su fficiently for rapid combustion to begin. Calhoon et al.134 found a sharp increase in heat flux at 4.44 cm from the injector face. It was believed that this was due to the deterioration of the cooling effect of the fuel near the wall. Reed145 sought to understand mixing and heat transfer characteristics in small film cooled rockets using a gasgas hydrogen/oxygen combusto r. Reed also found a local maximum heat flux in the flow. Similar to Calhoon et al., Reed proposed that the loca l maximum heat flux was due to a breakdown of cooling due to increased mixi ng. Reed also suggeste d that the peak heat flux has to do with the mixture ratio and the amou nt of available oxygen pres ent to react with the cooling flow. One characteristic of reacting flow in a ro cket engine combustion chamber is that the reaction happens very quickly, and is limite d by the amount of tur bulent mixing in the combustion chamber. The combusting shear area is highly turbulent, so mixing, and thus reaction, can occur very quickly. Foust et al.146 investigated the use of gaseous propellants for the main combustion chamber by exploring the mixing characteris tics, combustion length, flame holding, injector face heating, and the potential for combustion chamber wall cooling. Foust found that reaction occurs almost immediately af ter injection, and the high velocity gaseous
PAGE 110
110 hydrogen jet decelerates rapidly after leaving the injector, then reaccel erates due to heat addition. Morren et al.147 found that the temperature profiles along the combustion chamber wall were caused exclusively by mixing and shear layer effect s, and not by pressure gradients. de Groot148 found that for gaseous injection, a high degree of mixing is cr itical. Additi onally, combustion must occur away from injector face to reduce injector face wall heating. Several efforts have been made to study inject or dynamics in detail, such that the results could be used to improve combustion chamber perf ormance. Experiments have only been able to provide a limited picture due to measurement difficulties posed by the high temperature, high pressure, and multiple species flow. Rupe and Jaivin136 studied the effect of injector proximity to the wall on combustion chamber heat flux. However, the available experimental techniques were not enough to fully understanding the bounda ry layer phenomena and were insufficient to allow accurate future predictions on heat flux based on the results. Strakey et al.149 experimentally examined the eff ects of moving the inner oxygen post offaxis and away from the combustion chamber wall to study the effects on wall heat transfer. While an attempt was made to balance between wall heating and combustion performance, the experimental results were not enough to accurately predict whether the wall heat transfer would be sufficiently low for the optimized designs. Smith et al.150 conducted hot fire tests of a 7element gasgas coaxial injector for different pressures to assess the effect on inj ector face heating, but we re unable to accurately evaluate chamber performance due to the pres ence on interior boundary layer cooling. CFD promises to fill in the gaps of knowledge left by experimental evaluations. The experimental procedures are currently di rected at providing va lidation data for CFD simulations. Experimental measurements are inva luable in the ongoing task of CFD validation. A benchmark injector experiment was run by Marshall et al.17 for the purpose of CFD validation.
PAGE 111
111 In this study, wall heat flux measurements were taken for a gasgas single element shear coaxial injector element. A singleelement injector was used rather than a multielement injector because the experimental set up is simpler and cheaper. The study was conducted using hot and ambient temperature gaseous oxygen and gaseous hydrogen at four chamber pressures. Preburners were used to supply th e hot gases. The preburners were used to simulate flow in a stagedcombustion cycle, such as the SSME. Th ey found that for a given geometry and injection velocity, the heat flux along the wall is pressure independent. Conley et al.22 investigated a GO2/GH2 coaxial shear injector to examine the heat flux to the combustion chamber wall for the use of CFD validation. The experi mental setup included optical access and the ability to measure heat fl ux, temperature, and pressure in the combustion chamber. In the experiment, Conl ey et al. investigated the eff ects of combustion chamber length, pressure, and mass flow rate on th e heat flux to the wall. A square chamber was used to allow for a window to be placed on th e combustion chamber. The optical access was in place for the purpose of providing information about combusti on dynamics by viewing the flame at various pressures. This was accomplished using OHPl anar Laser Induced Fluorescence (OHPLIF) to show the primary reaction region, or flame. The edges of the square combustion chamber were rounded to reduce stress on the chamber due to pr essure. Conley et al. found no dependence of the heat flux on pressure when the heat flux was scaled by th e mass flow rate. Specifically, Conley found that when the heat flux could be scaled by the inlet hydrogen mass flow as 2, s caledHinqqm the heat flux profile along the wall co llapsed to a single curve for a given combustion geometry. Conley concluded from th e experiments that th e amount of heat flux along the combustion chamber wa ll was primarily a function of the combustion chamber geometry.
PAGE 112
112 4.1.4 Review of Select CFD Modeling and Validation Studies The majority of CFD validation efforts to date have concentrated on single element coaxial injectors. Work has focused on LOX/GH2 or GO2/GH2 combustion. Most efforts focused on matching species or velocity profiles. Only recen tly have efforts been in place to work towards using CFD as a combustor design tool. These recent efforts focus on accurately modeling the heat characteristics of combustors. Liang et al.27 sought to improve the e ffects of multiphase mode ling that included liquid, gas, and liquid droplets. A 2D axisymmetric model was used along with a chemical model consisting of a 9equation kinetic model along wi th 4 equilibrium equations. The equilibrium equations were included to help anchor the flame during the comput ation by allowing for instantaneous reactions. In this case, equilibrium reactions were necessary due to the coarseness of the grid used. The turbulence model used is th e eddy viscosity model. In this case, no effort is made to couple combustion with the turbulen ce equations. For the simulation of a gasgas injector, a mixture fraction of unity was used for the propellants, and the combustion chamber was originally filled with oxygen. An artificia l ignition region was plac ed near the oxygen post tip. The computation was run for a physical equiva lent of 10ms. An attempt was made to use the 2D model to simulate LOX/GH2 multielemen t injector flow by changing the walls from a noslip to a slip boundary conditi on. Local peak temperature is around 2000K, and the average temperature is around 1500K. Liang et al. mentio ned that, based on the CFD results, atomization might be a ratecontrolling factor, but there was insufficient experimental data to confirm this fact. Liang et al. notes that grid resolution has a significant e ffect on flame ignition and flame steadiness, but no grid sensitivity study is conducted to explore the effects. Temperature, velocity, and mass fraction contours were obtained, but no comp arison was made to experimental results, so the accur acy of the simulation is unknown.
PAGE 113
113 Foust et al.146 modeled GO2/GH2 combustion using an 18 reac tion finiterate chemical model. The chemistry parameters were determ ined based on the temperature field. The kturbulence model was used. Preliminary co mputations were done to obtain the boundary conditions upstream of the inject or. The final computa tional model extends s lightly upstream of the injector exit. The model is 2D Cartesia n, and the grid was coarse at 10151. The computation found good agreement in the specie s concentration profiles and the velocity profiles. Using CFD, Foust confirmed the fl ame holding ability of th e oxidizer post tip over wide ranges of equivalence ratios for a given injector. Cheng et al.151 looked to develop a CFD spray comb ustion model to help understand the effects on wall erosion. The Finite Differe nce NavierStokes (FDNS) solver used a kmodel with wall functions, and the real fluids models were used for the multiphase flow. Heat and mass transfer between phases was neglected. The velocity and species concentrations were solved based on a constant pressure assumpti on, and then the density and temperature ere determined based on the real fluids model. The pressure was then corrected based on the newly determined density. Finiterate and equilibrium chemistry models were used. The GO2/GH2 computation used the 4equation equilibrium ch emistry model for hydrogenoxygen combustion. The shear layer growth was well predicted by the model. Good co mparisons were made between the CFD and experiment for velocity and species profiles. Disagreement in the H2O species profile was attributed to experime ntal measurement error. It is assumed that the grid was twodimensional, however, no grid information is provided. Schley et al.152 compared the Penn State University (PSU) code used by Foust et al.146 and the NASA FDNS code used by Cheng et al.151 with a third computat ional code by called Aeroshape 3D153 (AS3D). The purpose in the CFD development was to reduce combustor
PAGE 114
114 development time and costs. The CFD comput ations were all based on the experiment performed by Foust et al.146 Schley noted that the experimental combustion chamber had a square crosssection with rounded corners, while all three resear ch groups chose to treat it as an axisymmetric computational domain. None expected the nonasymmetry to impact the comparison, as other errors were thought to be more significant. The FDNS and AS3D codes included the nozzle in their computations, while the PSU code did not. The FDNS code did not include any upstream injector analysis, while the ot hers did at least some simulation of the flow within the injector. All three codes used the kturbulence model. Coarse grids were used, as the researchers found that the solution increased in unsteadin ess with increasing grid resolution. Despite the different codes used and the di fferent treatments in boundary conditions, all computations agreed reasonably wi th each other and the experiment. Ivancic et al.154 used the AS3D code to do simulate LOX/GH2 combustion using a coaxial shear injector. Ivancic et al. assumed chemical equilibrium for the combustion modeling with no coupling between turbulence and chemistry. A real gas model was used for the oxygen. An adaptive grid was used with 150,000 grid elem ents and was grid independent based on finer meshes. The simulated shear layer thickness and th e flame location were not consistent with the experimental results. It is s uggested that the discrepancy is due to the 2D simulation of 3D effects. Another reason for the discrepanc y was the overly simplifie d combustion model. Lin et al.155 specifically investigates the ability of CFD to predict wall heat flux in a combustion chamber for a shear coaxial single elem ent injector based on th e experimental results from the test case RCM1 of the 3rd Rocket Combustion Modeling (RCM) Workshop held in Paris, France in March, 2006. The experimental setup of the RCM 1 test case is shown in Figure 45. Lin et al. uses the FDNS code introduced by Cheng et al.151 The model was simulated
PAGE 115
115 using two codes: FDNS, a finitevolume pressureb ased solver for structured grids, and lociCHEM, a densitybased solver for generalized gr ids. The FDNS code has long been used by NASA for reacting flow simulation, but is cumbersome to use for complex geometries. LociCHEM, on the other hand, is a new code that is still in development. Because it is applicable to structured or unstructured grid s, lociCHEM is convenient for complex geometries. The code used a 7species, 9reaction finite rate hydroge noxygen combustion model. The computational grid encompassed the preinjector flow, combusti on chamber, and nozzle. Computations are 2D axisymmetric, and consisted of coarse grids with 61,243 points using wall f unctions, or fine grids with 117,648 points where the turbul ence equations were integrated to the wall using MenterÂ’s baseline model. The mass flow rates are fixed at the inlet to the computational zone with a uniform velocity profile. The experimental temp erature profile is appl ied along the wall of the computational domain. The computational domains were initially filled w ith steam, and for the LociCHEM code, an ignition point was specified The temperature contours for each code by integrating the turbulence equations to the wall ar e given in Figure 46. The best matching of the heat flux values came by using the lociCHEM code integrated to the wall using MenterÂ’s SST model. Prediction of the h eat flux was good in the recircul ation zone and poor beyond the reattachment point, as shown in Figure 47A. Lin et al. cited th e need for additional validation and improved turbulence m odels to improve results. West et al.156 expanded on the work by Lin et al.155 using an updated version of lociCHEM. West et al. performed a detailed grid refinement study on the CFD model of the RCM1 test case. In particular, West et al. found that a structured grid was prohibitive, in that a large number of grid points were requi red for adequate reso lution near the combus tion chamber wall. For this reason, West et al. us ed hybrid grids consisting of a st ructured grid near the wall and
PAGE 116
116 triangle elements away from the wall. The st ructured grid had 51,000 points, while the hybrid grids had between 500,000 and 1,000,000 points. West et al. separately expl ored the effects of grid resolution in the boundary layer region, the nonboundary layer region, and the region in the vicinity of the flame. The effect s of grid resolution were small, and the coarsest hybrid grid was found to be sufficient to accurate ly predict the peak heat flux. The study also compared three turbulence models. The use of MenterÂ’s SST model provided slightly better results than MenterÂ’s BSL model and Wilcox kmodel. The basic kwas found to provide the worst approximation, as the prediction of the peak heat flux was in the wrong lo cation, indicating that the model would not be appropria te for a design problem. All tu rbulence models were found to overpredict the heat flux downstr eam of the peak heat flux. Thakur and Wright157 tested the code lociSTREAM against the same RCM1 test case used by Lin et al.155 A 2D axisymmetric model was used for the computational domain. A 7species, 9reaction finite rate hydrogenoxygen combus tion model was used. The CFD model was initiated with steam filling the combustion cham ber. Similarly to Lin et al., different grids were investigated: a fine grid with 104,000 grid points and a co arse grid with 26,000 points for integrating the turbulence equations to the wa ll, and a grid with 104, 000 points that remains coarse near the combustion chamber wall to use with wall function formul ations. The maximum chamber temperature was 3565 K, and the temperat ure contours are shown in Figure 46. The pressure was found to be approximately constant from the injector inlet to the nozzle throat. Thakur and Wright found that th e use of wall functions lowered th e predicted heat flux at the wall, regardless of whether the grid was fine or co arse. It is suspected that this may be due to errors in the computational code, as this same phenomenon was corrected in Lin et al. 155 The heat flux profiles obtained when in tegrating to the wall are similar to the results by Lin et al., as
PAGE 117
117 shown in Figure 47. It was also determined that the wall heat flux was not sensitive to the grid distribution far from the combustion chamber wall, as long as the flame re gion is well resolved. Improving the fidelity of CFD computations is ongoing. Cu rrently, work is advancing with the use of threedimensional CFD models and simulation of multielement injector flow simulations. Tucker et al.158 reports on several worksinpr ogress at Marshall Space Flight center. This includes a simplifie d CFD model that uses a 7elem ent injector element flow along with bulk equilibrium assumption in the remainder of the 17 degree subsection to help approximate full multielement combustor dynamics of the Integrated Powerhead Demonstrator159 IPD injector. Sample results are gi ven in Figure 48A and Figure 48B. Additional simulations are being conducted on th e Modular Combustor Test Article (MCTA) to supplement the IPD injector studies as shown in Figure 48C. 4.2 Turbulent Combustion Model Combusting flows are highly complex, simultane ously involving chemical kinetics, fluid dynamics, and thermodynamics.160 For example, in a rocket combustion chamber, heat transfer to the walls depends on the boundary layer characteristics.161 Turbulence enhanced mixing along with the chemical kinetics can affect the rate of combustion.162 This section re views the basic governing equations, chemical ki netics, and equilibrium equations needed for analyzing combusting flow. 4.2.1 Reacting Flow Equations Because several species exist in combusting flow, the properties of each fluid must be accounted for. Mass fr actions are defined by k km Y m (41) where mk is the mass of species k in a given volume, and m is the total mass of gas.
PAGE 118
118 The total pressure and density are 11,whereandNN kkkk kk k R R p ppTpT WW (42) where Wk is the molecular weight of species k, R is the ideal gas constant, T is the temperature, and W is the mean molecular weight given by 1 1 N k k kY W W (43) The mass enthalpy of formation of species k at temperature T0 (the enthalpies needed to form 1 kg of species k at the reference temperature) is a property of the substance and is designated by 0 f kh (usually T0 = 298.15K). The enthalpy of species k is then given by 00 chemical sensible T kpkfk ThCdTh (44) where Cpk is the specific heat of species k. The heat capacities at constant pressure of species k are usually not constant with temperature in combusting flows. In fact, large changes are possible. They are usually tabulated as polynomial functions of temperature. The diffusion coefficient of species k in the rest of the mixture is Dk. Different gases diffuse at different rates. In CFD codes, this is usually resolved by using simplified diffusion laws such as FickÂ’s law: k kkkdY mYmD dx (45) In any chemical reaction, species must be c onserved. This conser vation is expressed in terms of the stoichiometric coefficients where kj and kj designate reactants and products, respectively:
PAGE 119
119 11 NN kjkkjk kkWW (46) The mass rate of reaction is given by: 1 M kkkjj jWQ (47) where N is the number of species, M is the number of reactions, and Qj is the rate of progress of reaction j which can be written as 11kjkjNN kk jfjrj kk kkYY QKK WW (48) where Kf and Kr are the experimentally dete rmined forward and reverse reaction rate coefficients. To solve for reacting flow, the conservation of mass, species, momentum, and energy is required. In this case, the governing equations in Cartesian coordinates are the conservation of mass, species, momentum, and energy, respectively. 0j ju tx (49) ,,1,k jkkjkk jjY uYVYkN txx (410) ji ij i jjiuu u p txxx (411) iij j j jjju q HpuH txxx (412) where is given by Equation 42, Vk is the diffusion velocity of species k, and 1 2ii H huu (413)
PAGE 120
120 0 1 0 111 N Tfkk k NNN Tkkskkfkk kkkh hhh (414) where hs is the sensible enthalpy. The heat flux qj can be written as 1 2PrPrPrii j LjLjLjuu hH q xxx (415) 4.2.2 Turbulent Flow Modeling The conservation equations given in equations (49) Â– (412) are applicable for laminar flow. For turbulent flow, the Reynolds Averaged Navier Stokes methods (RANS) use ensemble averaging to obtain the timeaveraged form of the conservation equations The flow properties are decomposed into their mean and fluctuati ng components where Reynolds averaging is used for the pressure and density components, while Fa vre averaging is used for the velocity, shear, temperature, enthalpy, and heat flux components. The mean fl ow governing equations are given as the conservation of mass, sp ecies, momentum, and energy: 0j ju tx (416) ,,1,k jkkjkjkk jjY uYVYuYkN txx (417) ji i jiij jijuu u p uu txxx (418) 11 22 1 2 iijjii j jjiijijiijjii jjjHuupuHuuu tx quhuuuuuuu xxx (419)
PAGE 121
121 where a tilde and double prime denote a Favreaverag ed mean and its fluctuation, respectively, and a bar and a prime denote a Reynoldsaverage d mean and its fluctu ation, respectively. The values of several terms in the mean averaged conservation equations are unknown, and must be modeled instead. Specifically, the terms are 1 2iikuu (420) where k is the kinetic energy per uni t volume of the turbulent velo city fluctuations, and the Reynoldsstress tensor ijuu that is denoted by ijijuu (421) These terms can be modeled using the Boussine sq eddy viscosity models (EVM). The EVM models the Reynoldsstresses by 22 33j il ijijttijij jilu uu uuk xxx (422) where t is the turbulent, or eddy, viscosity. The turbulent heat transport term juh is modeled as Prt j tjh uh x (423) where Prt is the turbulent Pra ndtl number, and the eddy conductivity is given by Prtp t tc k (424) In this study, the Prandtl numb er is kept constant at Prt = 0.9. The term 1 2 iijjiiuuuu from equation (419) is modeled as
PAGE 122
122 1 2 t iijjii kjk uuuu x (425) In the species equation, the term jkuY is modeled as tk jk ktjY uY Scx (426) The final forms of the conservation equations fo r mass, species, momentum and energy are given respectively as follows: 0j ju tx (427) ,1,k tk jkk jjktktjY Y uYkN txxScScx (428) 2 3ji j i il tij jijjiluu u u uu p txxxxxx (429) PrPr 2 3j jjLtj j ilt itij jjiljkjh HpuH txxx u uu k u x xxxxx (430) where 1 2 iiHhuuk (431) and the turbulent eddy viscosity t is modeled as a product of the velocity scale and a length scale multiplied by a proportionality constant C tC (432)
PAGE 123
123 The velocity and length scales can be determined using the kmodel163 for turbulence modeling. This model is based on idealized as sumption of homogenous is otropic turbulence and consists of a transport equation of kinetic energy of turbulence k and second for rate of dissipation of turbulent energy Note that in all subsequent e quations, bars and tildes have been dropped for convenience. t ik iikik kukP txxx (433) 2 12 t ik iikiuCPC txxxkk (434) where the production Pk of k from the mean flow shear stresses is defined as j ii kt jiju uu P x xx (435) and the velocity and length sc ales are given respectively by k (436) 33 42Ck (437) and the constants are given by 120.09;1.44;1.92;1.0;1.3kCCC (438) For turbulent flows, the sharp va riations in velocity and temperature profiles in the near wall region require special treatments for the wall boundary conditions. Two common methods are the use of a low Reynolds number k model and the use of wall functions. Examples of low Reynolds number k models can be found in Patel et al.,164 Avva et al.,165 and Chien.166 Wall functions assume that the turbulence calculations stop at some distance away from the wall. It
PAGE 124
124 applies loglaw relations near the wall.163 Further descripti on can be found in Pope167 and Tennekes and Lumley.168 4.2.3 Chemical Kinetics In rocket combustion, the fuel and oxid izer enter the combustion chamber in a nonpremixed fashion. The resulting diffusion flam es must form in the midst of turbulence fluctuations, leading to a highl y complex flow. One way of handling this is by making the assumption of Â“fast chemistry.Â” This assumption assumes that the reacti on rates of the species are very fast compared to the mixing time. This means that the instantaneous species concentrations and temperatures are functions of conserved scalar equations, such as the mixture fraction. The various average thermodynamic variables can be obt ained from scalar statistics using probabilitydensity functions Further details on using the conservedscalar approach to account for turbulencecombustion interactions can be found in Libby and Williams.169 The method used in this research is the finite rate chemistry assumpti on. In particular, a kineticcontrolled model is used that depends primarily on chemical kinetics as described previously. This method involves solving the co nservation equations desc ribing the convection, diffusion, and reaction sources. Exampl es can be found in Thakur et al.170 Turbulencecombustion models struggle in forms of accuracy and level of descriptio n, so new models are continually being developed. Chemical reaction rates control the rate of co mbustion and are related to flame ignition and extinction. Chemical kinetics ma p the pathways and rates of reaction from reactants to products. The global reaction of interest in this effort is the hydrogenoxyge n combustion reaction. Within the global reaction 2222HO2HO (439)
PAGE 125
125 there are several elementary reactions. A globa l reaction can possibly be made up of dozens of elementary reactions. However, the effects of some mechanisms are small at certain temperatures and pressures. In this case, the nu mber of reaction mechanisms used to represent a global reaction can be reduced. If necessary, the number of r eactions can be reduced as to contain only the key reactions. Often, the reaction mechanisms ar e Â“optimizedÂ” for this purpose. In this work, nine elementary reactions are considered for the reac tion of hydrogen and oxygen to water. The reaction mechanisms were s upplied by NASA Marshall. Each of the nine reactions has a rate coefficient K given by expb AuKTATERT (440) where Ru is the universal gas constant and A b and EA are experimentally determined empirical parameters given in Table 43. 4.2.4 Generation and Decay of Swirl Swirl number is the ratio of axial flux of angular momentum to the axial flux of axial momentum rd S R ud vA vA (441) where R is the hydraulic radius, and is the angular velocity. The swirl number can be approximated as W S U (442) where W is the mean angular flow velocity and U is the mean axial velocity. A swirl number of S < 0.5 indicates weak to moderate swirl.
PAGE 126
126 For the case of annular flow, it can be expected that the rate of sw irl decay will be much greater than that of pipe flow due to the increased fluid drag at the center core. If the difference between the diameters becomes suffic iently small, the flow is approximately equal duct flow. In the case of the hydrogen inlet flow, the outer to in let diameter ratio is 1.22. Any swirl should quickly be eliminated for di ameter ratios close to unity. For confined swirling flows, Weske and Sturove171 examined the decay of swirl and turbulence in a pipe flow. They found that the tu rbulence field decayed qui ckly, and that the rate of decay was a strong function of swirl number. Yajnik and Subbaiah172 found that swirl decay, in general, was not significant for pi pes less than 5 pipe diameters long. In general, it has been found that swirl deca ys exponentially. For laminar swirling pipe flow (Re < 4000), Talbot173 determined a theoretical formulatio n for decay of swirl in a pipe. The experimental values were found to fall within the theoretical values for Re from 100 to 4000. The solution to the swir l equation for pipe flow is ,,0zvrzvre (443) where z is the nondimensional length given by hzLD and is not a function of the initial flow distribution ,0 vr and is given by two theoretical limits: 1Re22.2 and 2Re74.3 The decay of swirl Ds can therefore be given by 11 ,0z svrz De vr (444) If it is simply assumed that the pipe flow example can be extended to annular flow, it can be estimated that the swirl decay in the hydroge n pipe will be between 47% and 88% without considering the effects of a fuel baffle. In actu ality, this amount woul d be greater for annular flow due to the growth of the boundary layer on the inner and outer surfaces of the annular pipe.
PAGE 127
127 If the amount of swirl generated in the pipe is small, it can be assumed that there will be negligible swirl at the exit of the hydrogen inlet. 4.3 Simplified Analysis of GO2/GH2 Combusting Flow As shown earlier, several factors are involved in injector and injecti on pattern design that are of interest in this research effort: A single injector can be analyzed as an approximation to full injector analysis The outer injector elements must be chose such that cooling is enhanced at the combustion chamber wall Injector placement can result in high local heating Pressure effects on combustion cham ber wall heat flux are very small Wall heat fluxes are largely inde pendent of the mixture ratio A sharp increase in heat flux occurs consiste ntly at a certain distance away from the injector face Analysis of the flow upstream of the injecti on may be necessary to determine the effect on combusting flow A preliminary combustion analysis can give some insight into the nature of H2/O2 reacting flow. In particular, the relationship between temperatur e, mass flow rate, pressure, heat flow rate, and equivalence ratio can be expl ored using a simplified versi on of the combustion chamber. A system can be considered wherein hydrogen and oxygen are injected such that the molar flow rate of hydrogen exceeds that of oxygen (Figure 49). The reaction is given as 22222HO2HO21H (445) and the following assumptions are made: Constant pressure Complete combustion with no dissociation Fuelrich combustion ( > 1) Steadystate All combustion products exit at the same temperature
PAGE 128
128 For this system the heat output in watts is given by 22222 ,, 22111 2 2HinOin outHHOHOH TT Tqmhhhh (446) where the species enthalpies h are equal to the molar enthalpies in J/mol divided by the molecular weight of each species. Using this equa tion, several observations can be made. First, the heat flow rate is dependent only on the specie s enthalpies at the inlet and outlet temperatures, the mass flow rates, and the equivalence ratio. Second, it must be noted that enthalpies are largely independent of pressure. This would indicat e that the heat flow ra te is also independent of pressure when the inlet temperatures, outle t temperatures, and mass flow rates are held constant. Table 41. Selected inj ector experimental studies. Author Injector type Propellants Key results Hutt and Cramer (1996)135 Single element oxidizerrich coaxial swirl injector LOX/GH2 Innovative measurements of LOX swirl spray Preclik (1998)143 Multipleelement coaxial injector LOX/LH2 Wall heat flux measurements Farhangi et al. (1999)137 GO2/GH2 Design, develop, and demonstrate a subscale hydrogen/oxygen gaseous injector Smith et al. (2002)150 Multipleelement coaxial for FFSC GO2/GH2 Evaluate performance of gasgas injector. Mayer (2002)141 Nonreacting jet LN2 Characterization of liquid jet flow Marshall et al. (2005)17 Singleelement shear coaxial GO2/GH2 Wall heat flux measurements Conley et al. (2007)22 Singleelement shear coaxial GO2/GH2 Wall heat flux independent of pressure, injection velocity
PAGE 129
129 Table 42. Selected CFD and numerical studies for shear coaxial injectors. Author Propellants P (MPa) T(K) Summary/ parameters compared Grid spec. Turbulence model Accuracy Liang et al. (1986)27 LOX/GH2 37.9 94.4/298, 118/ 154 Developed threephase combustion model N/R Eddyviscosity Unknown. Cited inadequate experimental knowledge. Foust et al. (1996)146 GO2/ GH2 1.3 297 Velocity and species comparison 2D Cartesian 10151 kGood species matching with inhouse experiment. Velocity matching degrades downstream. Cheng et al. (1997)151 LOX/GH2, GO2/GH2 3.1, 1.3 117/309, 290/298 Velocity and species comparison Axisym. 15181 kGood species matching to experiment by Foust et al. 146 Poor H2O species matching. Ivancic et al. (1999)154 LOX/GH2 6.0 127/125 Time/length scale investigation Axisym. 1.5105 elements kFair matching of radial OH distribution to inhouse experiment. Lin et al. (2005)155 GO2/GH2 5.2 767/798 Heat flux comparison (RCM1) Axisym. 61,243Â– 117,648 points Menter BSL Good peak heat flux prediction. Overpredicts downstream heat flux. Low grid dependence. West et al. (2006)156 GO2/GH2 5.2 767/798 Heat flux comparison (RCM1) Axisym structured/ hybrid 50,000 Â– 1,000,000 points Wilcox k, Menter BSL and SST Good peak heat flux prediction when adequate grid resolution in high gradient areas. MenterÂ’s SST model provides superior peak heat flux prediction. Thakur and Wright (2006)157 GO2/GH2 5.2 767/798 Heat flux comparison (RCM1) Axisym. 26,000Â– 104,000 points Menter SST Good peak heat flux prediction. Overpredicts downstream heat flux. Table 43. Reduced reaction mechan isms for hydrogenoxygen combustion. Reaction A ((m3/gmol)n 1/s) B EA (kJ/gmol) 22HOOHOH 1.7010100.00 2.41104 22HOHHOH 2.1910100.00 2.59103OHOHOHH 6.021090.00 5.501022HOOHH 1.80 1071.00 4.481032HOOHO 1.2210140.91 8.37103HOMOHM 1.001010 0.00 0.02OOMOM 2.5510121.00 5.941042HHMHM 5.00109 0.00 0.02OHHMHOM 8.401015 2.00 0.0
PAGE 130
130 H 2 O2 shear layer shear layer recirculation recirculation flame homogenous products oxygenrich core heat transfer reattachment p oint heat transfer Figure 41. Coaxial injector a nd combustion chambe r flow zones. Figure 42. Flame from ga seous hydrogen Â– gaseous oxygen single element shear coaxial injector.
PAGE 131
131 A B Figure 43. Multielement inject ors. A) Sevenelement shear coaxial injector and B) multielement IPD159 injector. Both injectors contain smaller holes at the outer edge that inject unburned fuel for wall cooling.
PAGE 132
132 Figure 44. Wall burnout in an uncooled combustion chamber. A Figure 45. Test case RCM1 injector. Test rig A) schematic and B) photo.
PAGE 133
133 B Figure 45. Continued. A B C Figure 46. Temperature cont ours for a single element injector. A) FDNS and lociCHEM,155 B) lociSTREAM,157 and C) lociCHEM156 codes. Plots are not drawn to same scale.
PAGE 134
134 A B C Figure 47. CFD heat flux results as compared to RCM1 experimental test case. A) FDNS and lociCHEM codes,155 B) lociSTREAM code,157 and C) lociCHEM code using hybrid grids and various turbulent models.156
PAGE 135
135 A B C Figure 48. Multielement injector simula tions. Reproduced from [158]. A) Integrated Powerhead Demonstrator (IPD) injector showing computational domain and B) a closeup of temperature contours. C) Norm alized temperature contours are shown for Modular Combustor Test Ar ticle (MCTA) injector.
PAGE 136
136 22,,,HinHinmT 22,,,OinOinmT q2,,HoutoutmT 2,,HOoutoutmT 22,,,HinHinmT 22,,,OinOinmT q2,,HoutoutmT 2,,HOoutoutmT Figure 49. Fuel rich hydrogen a nd oxygen reaction with heat release.
PAGE 137
137 CHAPTER 5 SURROGATE MODELING OF MIXING DYNAMICS A bluff bodyinduced flow is used as a m odel problem to help probe the physics and surrogate modeling issues rela ted to the mixing dynamics. Understanding of the mixing characteristics in the wake of the bluff body has im portant implications to reacting flow, such as in the shear reacting layer of in jector flows. The sensitivity of the mixing dynamics to the trailing edge geometry of the bluff body can be explored using CFDbased surrogate modeling. Plausible alternative surrogate models can le ad to different results in surrogatebased optimization. The current study demonstrates the ability of using multiple surrogate models to discover the inadequacy of the CF D model. Since the cost of constructing surrogates is small compared to the cost of the simulations, using multiple surrogates may offer advantages compared to the use of a single surrogate. Th is idea is explored for a complex design space encountered in a trapezoidal bl uff body. Via exploration of local regions within the design space, it is shown that the design space has small islands where mixing is very effective compared to the rest of the design space. Both polynomial response surfaces and radial basis neural networks are used as surrogates, as it is difficult to use a single surrogate model to capture such local but critical features The former are more accura te away from the highmixing regions, while the latter are mo re accurate near these regions. Thus, surrogate models can provide benefits in addition to simple model approximation. 5.1 Introduction The case study focuses on the mixing and tota l pressure loss characteristics of time dependent flows over a 2D blu ff body. Bluff body devices are often used as flameholding devices such as in afte rburner and ramjet systems. Blu ff body devices shoul d have good mixing capability and low pressure loss across them. Ch allenges exist in that the bluff body flow is
PAGE 138
138 unsteady and difficult to predict. There is recirculating flow in the nearwake region that decays to form a welldeveloped vortex street in the wa ke region. The instantaneous loss and degree of mixing changes over time and must be well resolved for accurate solutions. An earlier investiga tion of the optimization of a tr apezoidal bluff body was conducted by Burman et al.138 The study uses a relatively coarse me sh (9272 computational cells) that was selected by performing a grid sens itivity study on a single case ba sed on the drag coefficient. The effects of grid resolution on the measure of mixing were not inves tigated, so it is possible that not all of the flow effects were captured. Fo r higher Reynolds number flow, as observed by Morton et al.,174 grid resolution can have a marked eff ect on the prediction of unsteady flows and can substantially affect the fidelity of the surrogate model. Surrogate models are used to approximate the effect of bluff body geometry changes on total pressure loss and mixing effectiveness. The timeaveraged flow field solutions are compared by looking for common trends and co rrelations in the flow structures. 5.2 Bluff Body Flow Analysis In practical considerat ions, such as the aforementioned af terburner and ramjet combustor, the flow around many bluff body devi ces is turbulent. To simplif y the analysis, an effective viscosity is often estimated ba sed on engineering turbulence cl osures. The effective Reynolds number is therefore considerab ly lower than the nominal Reynolds number. For the purpose of this study, the flow is modeled using a Reynolds number in the range of the effective Reynolds number without resorting to turbulence models. Furthermore, without e ither enforcing the wall function or suppressing high gradie nt regions, both typically obse rved in engineering turbulent flow computations, the fluid fl ow tends to exhibit more unste ady behavior, rendering the flow computations more interesting and challenging. The bluff body geomet ry along with the flow
PAGE 139
139 area is described in detail. The computationa l domain and bluff body geometry specifications are based on the study performed by Burman et al.175 5.2.1 Geometric Description a nd Computational Domain The dimensions of the bluff body are given in Fi gure 52. The variables of interest are B b and h Altering the variables changes the slant a ngles of the upper and lower surfaces. The area, A of the bluff body is held cons tant and equal to unity. The fr ontal height is kept constant at D = 1. The value of H can be calculated from the afor ementioned variables and constants using the equation 2HAhbDDB (51) The computational domain consists of a trapezoid al bluff body within a rectangular channel. The flow area is illustrated in Figure 53. The fluid is incompressible, and the flow is laminar with a Reynolds number of 250 given by UD Re (52) where D is the frontal height of the bluff body, is the kinematic viscosity, and U is the freestream velocity. The upper and lower boundaries have Â“slipÂ”, i.e., zero gradient, boundary conditions. The left boundary is the inlet, and the right boundary is the outlet. The 2D, unsteady NavierStokes equations are solved using a CFD code called STREAM.176 The time dependent calculations were solved using the PISO (Pressure Implicit with Splitting of Operators) algorithm, and co nvective terms were calculated using the secondorder upwind scheme. Other spatial derivati ves are treated with th e secondorder central difference schemes. The grid was constructed using ICEMCFD177 software. The grid is nonuniform, with a higher density of grid points n ear the body and in the wake area as shown in Figure 54. Grids of varying resolution were used to determine the grid that provided the best
PAGE 140
140 balance in terms of solution run time and soluti on convergence in terms of grid resolution as detailed in Goel et al.178 Multiple grid refinements were investigated to identify the adequate resolution for this problem. The number of grid po ints in each grid is given in Table 51. While the difference in the total number of grid points between Grid 1 and Grid 2 is not large, the grid density in the near field and wake region for Grid 2 is higher than Grid 1. In Grid 3 and Grid 4, both the near field grid density and far field grid density are high, making them more refined. Overall, Grid 1 has the poorest resolution and Grid 4 has the best resolution. Grid 2 was selected for the analysis as it appeared to offer the best tradeoff in terms of grid resolution and computational run time. The run time for a si ngle case was approximately eight hours on a 16 CPU cluster with Intel Itanium proce ssors (1.3 GHz) and 16GB of RAM. 5.2.2 Objective Functions and Design of Experiments The two objectives are the tota l pressure loss coefficient, CD, and the mixing index, M.I.. The total pressure loss coefficient at any time in stant is equal to the su m of the pressure and shear forces on the body divide d by the drag force. The total pressure loss coefficient is averaged between time t*, and time t* + T. It is assumed that T is large compared to the oscillatory time period. The total pressure loss coefficient is given by 2 1 2 *11tT Dxixi tCpnndSdt TUD (53) where p is the pressure, ix is the viscous stress tensor, is the fluid density, and D is the frontal height of the trapezoidal bluff body. The measure of mixing efficacy is given by the laminar shear stress plus the unsteady stress. The mixing index, M.I., is chosen as the integral of the mixing efficacy over the entire computational domain averaged over a time range T. It is given by
PAGE 141
141 *11 ..tT V tU M IuvdVdt TVy (54) where is the fluid viscosity, U is the timeaveraged velocity, and u and v are the fluctuation velocities (, uuUvvV ). Equation 54 is related to the full Reynolds momentum conservation equation. The mixing is thus described as the combination of the momentum transfer due to viscosity and the momentum transfer by the fluctuation velocity field. The magnitude of the product of the fluctuation velocities is used to keep the sign of this term consistent throughout the data points. Only th e magnitude of the mixing index is important, so the absolute value of the timeaveraged quant ity represents the mixing index value. The constraints are given as 1 0.51.0 0.00.5 0.00.5 Bb B b h (55) The first constraint in Equation 55 is incorporated to keep the frontal he ight equal to unity so that the Reynolds number remains constant. The remaining constraints maintain convexity of the geometry. The constraints are identical to those used in Burman et al.175 A total of 52 data points were used in the analysis. The design of experiments (DOE) procedure was used to select the location of the data points that minimize the effect of noise on the fitted polynomial in a response surface anal ysis. A modified facecentered composite design (FCCD) is used to select 27 of the 52 data po ints. The facecentered composite design usually uses the corners and center faces of a cube for poin t selection, but in this case, half of the cube would violate the first constraint in Equation 55. This means that four points from the FCCD are infeasible and must be removed from the design. Additional points were added on the
PAGE 142
142 interior of remaining triangular section to compensate for the missing points. The remaining 25 data points were selected using Latinhyp ercube sampling to fill the design space. 5.3 Results and Discussion The CFD solutions were obtained for the 52 da ta points. These solutions were compared and categorized based on the mean flow field. Response surface and radial basis neural network prediction models were constructed from the data. Comparisons of fit were made between the surrogates. 5.3.1 CFD Solution Analysis Figure 55A shows typical instantaneous vort icity characteristics for a given trapezoidal bluff body geometry. The flow exhibits characte ristic vortex pair shedding. This oscillatory behavior is reflected in the instantaneous total pressure loss coefficient and mixing index values. The vortex pair shedding is asymmetric due to the asymmetry of the bluff body. However, in order to evaluate the overall device performan ce, the timeaveraged flow is a more concise quantity. Figure 55C shows the timeaveraged solution for the same geometry. In the time averaged solution it can be seen that the oscillatory behavior in the wake has been completely averaged out. Table 52 provides the range of the total pressure loss coefficient and mixing index for the 52 design points. From Table 52, it is ev ident the maximum mixing index value lies well outside of the range indicated by the standard deviation. This indicates evidence of possible outliers. Along closer inspection, it was discov ered that three of the 52 cases had mixing index values that lay far beyond the range bounded by the standard deviation. These were termed Â“extreme cases.Â” A complete analysis of th e extreme cases is given in Goel et al.178 The flow fields from the CFD runs for geom etries representing the full expanse of the design space can be divided into several major groups as illustrated in Figure 56. The flow
PAGE 143
143 fields are categorized based on the location and si ze of mean recirculation regions near the bluff body. These groups will provide a means of re ference for describing general trends in the sample data. Geometries in Group F had the lo west total pressure loss coefficient with an average value of 1.85. Group B had the highest mi xing index with an average value of 655. It can be predicted that the best designs in terms of high mixing and low pressure drop, and thus the focus of a possible future design refinement may lie in the region including and between Group B and Group F. 5.3.2 Surrogate Model Results When constructing the response surface models of the two objectives, a cubic response surface model was found to perform better than a simpler quadratic model. The full cubic model is given by ****2***2** 01234567 ***2*3*2***2*3*2* 891011121314 ****2***2**2**2*3 141516171819Âˆ yBbhBBbbBh hbhBBbBbbBh BbhbhBhbhbhh (56) where a starred value indicates the variable is no rmalized between 0 and 1. Insignificant terms were then removed from some cubic response surfaces, resulting in reduced cubic models. The models used in this analysis are given in Table 53. Problems were discovered in the mixing index response surface model fit: The response surface has a low R2 adj value and a high RMS error, so the fit is very poor. Roughly half of the coefficients in th e response surface are very large (> 1000). Three of the large coefficients are positive and four are negative. The last two symptoms indicate a polynomial with sharp gradients and oscillatory behavior. It was suspected that these symptoms were due to behavior exhibited by the extreme cases. The extreme cases were removed in an attempt to improve the response su rface fit for the mixing index. As seen from Table 53, after remova l of the extreme cases the response surface fit
PAGE 144
144 improved to 0.921, and the magnitude of the highest term dropped from 4780 to 1550. The number of terms needed for the response surface fit also reduced from 16 to 14 and the RMS error reduced from 39 to 12. Radial basis neural networks we re also constructed to model the data to determine whether it was a better surrogate model for the bluff bod y. For the analysis, 10 of the 52 points were reserved as test cases for crossvalidation. The values of SPREAD and GOAL that minimize the error in the test cases are given in Table 54. The neural network also had problems in fitting the data when all 52 data points were used. Indi cations of those problems are given as follows. The optimum value of SPREAD for the mixing index is quite low. A large number of neurons are needed to fit the data. This behavior is indicative of large local variations in the data. When the same three points that were removed in the response surface analysis we re removed in the neural network analysis, the number of neurons needed to fit the data re duced from 35 to 21. The RMS error was also reduced and the optimum spread consta nt increased from 0.14 to 0.35. The response surface and neural network predic tion surfaces for the total pressure loss coefficient are very similar. The most noticeable difference in the total pressure loss coefficient predictions occurs at high h values. This behavior can be observed in Figure 57. For example, at h* = 1, B* = 0, and b* = 0.5, the response surface predicts a relatively high value for the total pressure loss coefficient, while the neural network predicts a relatively low total pressure loss coefficient value. The differences between the prediction models fo r the mixing index are striking, as seen in Figure 58. The neural network clearly shows three points with mixing index values that are considerably higher than that of the surrounding area. The respon se surface attempts to fit the same three points, but the effects are wider spread. The result is that prediction is poor in areas
PAGE 145
145 near these points. The large differences be tween the response surface and neural network surrogates are useful in that they can alert us to the fact that both models may have problems in some parts of the design space, so that at a minimum they serve as a warning signal. The three points with disproportionately high values were the ones identified as extreme cases in the previous analysis. When these ex treme cases are removed, the response surface and neural network prediction surface are more consistent with each other as seen in Figure 59. Yet, large differences still exist in the over all shape of the prediction model surfaces. The RMS error values shown in Table 54 indicat e that, overall, the neural network is able to fit the data better than the response surface. The points reserved as test points in the construction of the neural network should have an RMS error that is near the RMS error for all 52 points. For the mixing index, the RMS errors for the test points in the neural network prediction surface are lower than the RMS errors of the response surfaces as seen in Table 55. The fit of the neural network to the data that includes extreme cases is only fair, but is considerably better than that of the response surface. Removal of the extreme cases further improves the fit. The error in the test data is co nsiderably reduced and is, in fact, lower than the total RMS error. The RMS error for the total pressure loss coefficient approximation in the test data is also less than the tota l RMS error and is considerably smaller than the response surface RMS error. This indicates that better overall pred iction ability is gained for using the radial basis neural network. 5.3.3 Analysis of Extreme Designs A grid resolution study was conducted by Goel et al.178 Because the quality of the response surface for the mixing index from Grid 2 was poor an outlier analysis was conducted to identify possible outliers. Three cases were detected as potential outliers. The suspect outlier designs were analyzed to identify possible issues. The results for the grids used had adequate time
PAGE 146
146 convergence, and these three potential outlier cases did not indicate any abnormal behavior. In addition to the three potential outliers, two other cases were also analyzed to observe the trends of grid refinement on other solutions. The total pressure loss coefficient and mixing index for these cases are given in Table 56. The to tal pressure loss coeffi cient did not change significantly but the mixing index further increased. Also the contrast of the outlier designs with the other designs became even sharper. This indicated that the cases detected as potential outliers might represent a true phenomenon. Thes e designs are therefore designated as extreme designs rather than outliers. Six additional designs between extreme Case 1 and Case 3 (see Table 56) were analyzed using Grid 2. Designs for Case 1 and Case 3 were mirror symmetric. Since B and h were at the lower extremes, b = 0.25 was the axis of mirror symmetry. This means that the designs with b = 0.05 and b = 0.45 were mirror images and so on. The to tal pressure loss coefficient and mixing index for mirror symmetric designs were expected to be the same and observed to be almost the same. The mixing indices for these intermediate designs did not exhibit the trend of high mixing indices. The timeaveraged flow fields also di d not reveal any abnormality. The sudden change in behavior near the extreme designs was not captured, and higher resolution was required. To investigate this sharp change in the mi xing index, additional simulations were conducted in the vicinity of the extreme design Case 1. Grid 3 was used to conduct the simulations to get a more accurate estimate of the total pressure loss coefficient and mixing index. In these designs, h was set at the lower limit and B wa s also fixed at the lower limit for most cases. Only b was varied to differentiate between desi gns. It was observed that the mixing index was very high for the designs near the extreme case up to b = 0.02 and there was a sudden drop in the mixing index for design with b = 0.03. The results are shown in Table 57.
PAGE 147
147 Extreme design cases 1, 3, 52 and regular ca ses 7 and 17 were analyzed using Grid 4 to assess the convergence trends with respect to grid size. The results for these representative cases for different grids are shown in Figure 510. The mixing indices for different cases showed trends of achieving asymptotic behavior with im provements in grid resolutions. However, it is apparent that the Grid 2 solution has not reache d grid convergence, and th is may be contributing to the problem with the response surface fit. 5.3.4 Design Space Exploration The extreme cases cannot be dismissed as outliers, as Goel et al.178 determined, as they were found to be properties of the simulated fl ow. Thus the behavior of the prediction surfaces constructed using the extreme cases needed to be further investigated. The RMS errors indicate that when the extreme cases are included in th e surrogate model construc tion, the radial basis neural network fits the data better than the respon se surface. In reality, by inspection of Figure 58, the neural network appears quite deficient in areas away from the outlier. The fits of the response surface and radial basis neural networ ks constructed without the extreme cases were determined to be quite good. Thus, these surfaces could be compared to the prediction surfaces constructed with the ex treme cases to determine areas where the prediction was poor by using the equation with extremeswithout extremes...... MIMIMI (57) which leads to the contour plots in Figure 511. In areas away from the extreme cases, the change in the mixing index M.I. for the response surface is small. In contrast, even away from the extreme cases, there are areas of relatively high M.I.. This suggests that the response surface is better than the neural network at predicting the response away from the extreme cases.
PAGE 148
148 To test if the response surf ace fit was better away from the extreme cases, an exploration of the design space was conducted. Figure 512 plots the response su rface and radial basis neural network predicted values as co mpared to the actual CFD data from Table 57 along the line M.I. = f(B* = 0, b*, h* = 0). Of the cases used to construct the surrogate models, only the extreme cases (B,b,h) = (0,0,0) and (B,b,h) = (0,0,1) fall on this line. The radial basis neural network predicts both of these points very well, while the response surface underpredicts the mixing index at b = 1. The variation in the data is very high, with a sharp drop in the mixing index value over a sh ort distance near the extreme cases and very smooth behavior away from the extreme cases. Neither surface captures the complete behavior. In this case, the response surf ace predicted the interior points very well, as expected, but performed poorly near the extreme cases. The neur al network, on the other hand, better predicts the trends in the data near the extreme cases. Areas of large differences within the desi gn space between the tw o surrogates correspond to regions where the accuracy in the CFD model is compromised. Particularly, these regions are concentrated around the extreme designs. These designs were found by Goel et al.178 to have not reached grid convergence. The results indicate that the extreme designs may require a greater degree of grid refinement than other design points. 5.4 Conclusions Polynomial response surface and radial basis ne ural networks were used as surrogates for flow over a trapezoidal bluff body. Both the resp onse surface and the radial basis neural network approximations adequately predicted the tota l pressure loss coefficient. However, both surrogates poorly fit the mixing index and were substantially different from each other. The large differences in the surrogate approximations served as a warning signal and prompted further investigation. Inspection of the surroga te modelsÂ’ statistics re vealed that the fitting
PAGE 149
149 problems were due to high variations in the da ta in localized regions. Removing three extreme cases greatly improved the fits. Further investig ation revealed that the CFD simulations were not converged, and may have been contributing to the inability of the surrogates to properly approximate the model. Simulations of complex flows can sometimes re sult give unexpected results. Here, three cases had considerably higher mixi ng indices than the other cases. This behavior could not be accurately captured by a single surrogate model al one. The radial basis neural network was found to better approximat e the response near the extreme cases due to its local behavior, while the response surface provided be tter prediction of the response away from the extreme cases. Refinement of the design space near the extreme designs revealed that the differences in the surrogate models corresponded to the locations of the low accuracy CFD designs. Thus, surrogate models and design space refinement tech niques can be used to identify the presence and location of inaccurate CFD models. Table 51. Number of grid points used in various grid resolutions. Grid # Number of grid points 1 37,320 2 44,193 3 74,808 4 147,528 Table 52. Data statistics in the grid comparison of the CFD data. CD M.I. Max 2.21857 Min 1.79515 Range (maxmin) 0.42342 Mean 2.03617 Standard deviation 0.1166
PAGE 150
150 Table 53. Comparison of cubic response surf ace coefficients and response surface statistics. Terms CD M.I. M.I.a Intercept 1.95 801576 B* 0.368 149060.6B* 0.672 562795H* 0.337 186830B*2 1.24 2900394B*b* 3.29 47800B*2 0 5151550B*h* 0.507 8420H*b* 0.633 494556H*2 0.700 039.9B*3 0.844 1660360B*2b* 4.01 4620902B*b*2 4.08 36600B*3 0.686 0765B*2h* 0.597 71580.7B*b*h* 0.740 1000319B*2h* 0 769484B*h*2 0.517 00B*h*2 0 2680H*3 0.340 00 Response Surface Statistics R2 adj 0.963 0.6540.921RMS error 0.021 3912 # of terms 17 1614 Mean response 2.03 617604a extreme cases omitted normalized variable Table 54. Comparison of radialbasis neural network parame ters and statistics. Parameters CD M.I. M.I.a SPREAD 0.65 0.140.35 GOAL 0.0055 30005500 # of neurons 29 3521rms error 0.0109 15.611.6a extreme cases omitted Table 55. RMS error comparison for respons e surface and radial basis neural network. Objective function Quadratic RS rms error Cubic RS rms error RBNN rms error at test points RBNN rms error at all points CD 0.035 0.0220.0096 0.014 M.I. 54 4129 16M.I.a 16 1310 12a extreme cases omitted
PAGE 151
151 Table 56. Total pressure loss coefficient and mi xing index for extreme and two regular designs for multiple grids. CD M.I. Case # B b H Grid 1 Grid 2 Grid 3 Grid 1 Grid 2 Grid 3 1a 0.500 0.000 0.000 2.001.951.95477801 850 3a 0.500 0.500 0.000 2.001.961.96477805 853 7 0.500 0.250 0.250 2.092.172.14538635 643 17 0.875 0.125 0.375 1.931.921.91434541 579 52a 0.690 0.200 0.460 1.982.152.14502857 977a suspected outliers Table 57. Total pressure loss coefficient and mixi ng index for designs in the immediate vicinity of Case 1 using Grid 3. Case B b h CD M.I. 1 0.50 0.00 0.00 1.95850 I 0.50 0.01 0.00 1.95858 II 0.50 0.02 0.00 1.94886 III 0.50 0.03 0.00 1.96639 IV 0.50 0.04 0.00 1.98650 Figure 51. Modified FCCD. y x B b D =1 H h Figure 52. Bluff body geometry. The three design variables are B, b, and h. The parameter D is the frontal area which is kept constant at 1.
PAGE 152
152 6.5 D=1 32 5 14 Figure 53. Computational doma in for trapezoidal bluff body. X Y0102030 5 0 5 Figure 54. Computational grid for trapezoidal bluff body. A Figure 55. Bluff body streamlines and vorticity contours. A) Typical instantaneous vorticity contours, B) typical instantaneous streamlines, and C) timeaveraged streamlines for flow past a trapezoidal bluff body (B = 0.5, b = 0.25, h = 0.25) at Re = 250.
PAGE 153
153 B C Figure 55. Continued. Figure 56. Flow characterizat ion in design space on the Bb axis. At b = 0, the angle of the lower bluffbody surface is zero. At B = 0, small angles exist on lower and upper bluff body surfaces and the sum of the magn itude of these angles is at a maximum. Along the line B + b = 1, the angle of the upper surface is equal to zero.
PAGE 154
154 Figure 57. Comparison of res ponse surface (top row) and radial basis neural network (bottom row) prediction contours for total pressure loss coefficient at h* = 0 (left), 0.5 (center), and 1 (right). Normalized variables are shown. Figure 58. Comparison of res ponse surface (top row) and radial basis neural network (bottom row) prediction contours for mixing index (including extreme cases) at h* = 0 (left), 0.5 (center), and 1 (right). Norm alized variables are shown.
PAGE 155
155 Figure 59. Comparison of res ponse surface (top row) and radial basis neural network (bottom row) prediction contours for mixing index (excluding extreme cases) at h* = 0 (left), 0.5 (center), and 1 (right). No rmalized variables are shown. A B Figure 510. Variation in objective variables with grid refinement. A) CD and B) M.I.
PAGE 156
156 Figure 511. Difference in predicted mixing i ndex values from response surface (top row) and radial basis neural network (bottom row) prediction cont ours constructed with and without extreme cases at h* = 0 (left), 0.5 (center), and 1 (right). Normalized variables are shown. Figure 512. Comparison of response surface and radial basis neural network prediction contours for mixing index at B* = 0 and h* = 0. The CFD data is additional data and is not used to construct the surrogate models.
PAGE 157
157 CHAPTER 6 INJECTOR FLOW MODELING 6.1 Introduction Heating effects due to the arrangement of in jector elements is of prime importance. Injector placement can result in high local heating on the combustion chamber wall. A multielement injector face is made up of an array of injector elements. The type of elements need not be consistent across the entire injector face. The outer elements must be chosen to help provide some wall cooling in the combustion chamber. Sm all changes in the design of the injector and the pattern of elements on the injector face can significantly alter th e performance of the combustor. The elements must be arranged to maximize mixing and ensure even fuel and oxidizer distribution. For example, Gill26 found that the element diameters and diameter ratios largely influence mixing in the combustion chambe r, and that small diameters lead to overall better performance. Gill suggests that using a coax ial type injector for the outer row of injectors provides an ideal environment due to the outer flow being fuel. Rupe and Jaivin136 found a positive correlation between the temperature profil e along the wall and the placement of injector elements. Farhangi et al.137 investigated a gasgas injector and measured heat flux to the combustion chamber wall and injector face. It was found that the mixing of the propellants controlled the rate of reaction and heat release. Fa rhangi et al. suggested th at the injector element pattern could be arranged in a way that moved heating away from the injector face by delaying the mixing of the propellants. A CFD model is constructed based on experi mental results provide d by Conley et al.22 The CFD model is used to help understand the flow dynamics within the GO2/GH2 liquid rocket combustion chamber. In particular, the CFD model attempts to match the experimental heat flux data at the wall. The CFD simulation is also used to help provide answers to questions that could
PAGE 158
158 not be easily answered using an experiment alone. These issues include (1) the ability to show flow streamlines, (2) twodimensional wall heat fl ux effects, and (3) the relationships between flow dynamics and heat flux to the combustion ch amber wall. Finally, a grid sensitivity analysis is conducted to determine the effects of grid resolution and density on wall heat flux and combustion length. 6.2 Experimental Setup The validation exercise is based on an experimental setup by Conley et al.22 The injector itself was constructed of an outer hydrogen tube w ith an oxygen tube within. The oxygen tube is concentric to the hydrogen tube and is stabilized by a baffle. This baffle keeps the oxygen tube centered within the hydrogen tube and helps to evenly distribute th e hydrogen flow. The temperature and pressure of the propellants when they are injected is not known from the experiments. In modeling flow in the combustion chamber, Schley et al.152 compared the results from three different codes with experimental results with good agreement for gasgas combustion. One primary concern for all simulati ons was the lack of inlet turbulent conditions for the combusting flow. All contributing parties simply made an approximation of suitable inlet conditions. By analyzing the pr einjector flow, some of the gu esswork can be removed. A preinjector analysis is used to determine the flow conditions within the injector itself, as well as to determine how the flow conditions affect the flow entering the combustion chamber. The combustion chamber is allowed to co ol between tests. Conley et al.22 determined in a preliminary analysis that the temperature along the combustion chamber wall would reach a maximum of 450 K after 10 seconds of combustio n. It was estimated that the actual wall temperature reached up to 600 K. For a simila r CFD simulation, this would mean that the combustion chamber wall would be considerably cooler than the bulk flow, where the flame
PAGE 159
159 temperature can reach up to 3000 K. Thus, the th ermal boundary layer plays a significant part in the estimation of the heat flux via CFD. 6.3 Upstream Injector Flow Analysis Because it was determined that the geometry of the injector could have a large effect on combustion dynamics, the flow upstream of the injector must be characterized for flow properties and effects. A CFD st udy was conducted based on expe rimental results of a coaxial injector built and test ed by Conley et al.22 In particular, the CFD si mulation will seek to glean information from the injector that is difficu lt or impossible to meas ure using conventional experimental methods. The simulation will in vestigate the possibility of swirl development upstream of the injector, as well as determine the pressure drop across the injector. Finally, the shape of the velocity profile at the injector exit is investigated. 6.3.1 Problem Description The geometry consists primarily of an annula r flow region with the hydrogen entering at a direction perpendicular to the primary flow. A flow baffle, that exists in the experimental setup and is meant to center the oxygen tube within the hydrogen tube, has been omitted in this analysis. The flow undergoes a sharp reduction in area near the exit of the geometry. The flow parameters of interest are those of the exiting flow. For the hydrogen inlet flow, two separate cas es were examined. The first case assumes that the inlet flow is parallel to the walls, or = 0. This assumes that the hydrogen is being drawn from quiescent flow such as a tank, for example. The second case assumes that the flow enters at a slight angle, or = 5.7. This case represents a pe rturbation that may have existed in the original flow. The two cases are compared and analyzed for similarities and differences as well as for the potential for swirl development and subsequent decay.
PAGE 160
160 The flow regime was assumed to be turb ulent, compressible, and unsteady. The computational domain and boundary conditions are given in Figure 62. The computational mesh consisted of 2.36105 nodes and the mesh density was increas ed near the no slip walls and areas of flow curvature as shown in Figure 63. The computation was run using the lociSTREAM157 code using six 1.3GHz processors. Each computation was run long enough to correspond to a physical time of 1.6 seconds. 6.3.2 Results and Discussion Figure 64 shows the vorticity contours for the hydrogen flow given both inlet conditions. At x = 0, the flow makes a 90 turn while simultan eously flowing around the inner pipe. This results in the development of two vorticies in the xplane on either side of the y axis. These vorticies are convected down stream while they gradually decay. Th is decay is due to a relatively low Reynolds number combined with a long nondimensional chamber length of 10.5. However, in the converging section of the annular pipe, the remaining weak vorticies strengthen due to a sharp increase in the local Reynolds nu mber. Due to the small tangential velocity and small hydraulic diameter, the viscous effects result in a dramatic decrease in xplane vorticity, thus reducing the overall swirl value to neglig ible levels. The nondi mensional length of the portion of the pipe with a small hydraulic diameter is comparable to that of the larger portion of the pipe at 10.7. This allows the flow to reach almost fullydeveloped flow conditions at the exit of the pipe. Figure 65 shows the overall swirl numbers that arise as a result of the individual vorticies. The overall swirl is very weak (< 0. 1) outside the entrance region at the x = 0 plane. The magnitude increases momentarily at the convergin g section of the pipe, then peaks as viscous effects overcome the accelerating effect cause by the convergence and retard the flow in the tangential direction. The swirl number then re duces to a very low, and negligible, amount.
PAGE 161
161 Figure 66 shows the average velocity for increasing x for each inlet condition. The average mean velocity in the x direction, u, is nearly identical for each of the two inlet conditions. However, the inlet conditions result in very different values for the velocity in the direction, v. While the = 0 inlet condition results in negligible flow rotation, the = 5.7 causes a noticeable, but still small maximum v velocity of 0.6m/s. The tangential velocity v continues to decrease until the converging porti on of the pipe where there is a sharp increase. As the pipe area again becomes constant, the tangential velocity decreases again to a value near 0.3. An examination of the Reynolds number reveals th at practically no difference occurs in the Reynolds numbers for the two inlet conditions, as shown in Figure 67A. The Reynolds number in the tangential direction for the = 0 inlet condition is nearly zero along the length of the pipe. For the the = 5.7 inlet condition, the Reynold s number in the tangential direction continually decreases until its value nearly reaches that of the the = 0 inlet condition. Based on these results, it can be concluded that perturbi ng the inlet flow ultimately has little effect on the resulting exiting flow. Because little difference exists in the flow ex iting the injector, the efforts can now focus on the single case based on uniform inlet fl ow. There is a pressure drop of 5.5103 MPa across the injector, nearly all of which occurs in the conve rging region and final pipe section, as shown in Figure 68. The hydrogen exit velocity profile, k, and are shown in Figure 69. The hydrogen exit velocity profile is nearly symmetric with a maximum value of 60.6 m/s and a Reynolds number of 2.8103. A more distinct asymmetry can be seen in the turbulent kinetic energy k profile. Peaks in k occur near the pipe walls. Near the walls, the velocity gradients are high, leading to a large production of k. At the walls, the velocity is zero, so the turbulent kinetic energy is also set to zero. There is a decay of k away from the walls as the turbulent production
PAGE 162
162 of k again becomes zero at the centerline where the u velocity gradient is equal to zero. The production of turbulent dissipation is also dependent on the velo city gradients. However, at the wall, the value of turbulent dissipation is un known, and must be guessed. The value of turbulent dissipation at the wall is therefore set based on the cell distance to the wall. For an infinitely small cell, the turbulen t dissipation would be set to an infinite value. The free stream turbulent dissipation boundary condition is based on the free stream velocity and characteristic length. Therefore, there is a gradual reduction of from the walls to th e free stream value. Similar simulations were run for the oxygen flow. Full details are not given, as the geometry was very simple. The results are thos e of standard turbulent pipe flow. The oxygen flow Reynolds number is an orde r of magnitude higher at 3.7104 than the hydrogen flow due largely to a larger hydraulic diameter. The ma ximum velocity for the oxygen flow was 25.4 m/s. The shapes of the k and profiles are similar to those of the hydrogen flow, but are of lower magnitude than the hydrogen flow. 6.3.3 Conclusion Regardless of the characteristics of the flow entering the hydrogen inlet, the swirl at the outlet was found to be negligible. It was found that by biasing the inlet flow, a small amount of swirl developed, but it was not enough to sustai n through the final flow section. This flow reduction had two effects: (1) the flow accelerated in the xdirection to such a degree such that the swirl number became very small, and (2) the flow profile at th e outlet was slightly asymmetric. This asymmetry may or may not have a noticeable effect on the combustion chamber dynamics.
PAGE 163
163 6.4 Experimental Results and Analysis No temperature measurements are taken direc tly at the combustion chamber wall. Instead, equations were used to extrapolate temperature and heat flux values to the combustion chamber wall. The heat flux is measured along one of the combustion chamber walls and is determined using sensors that measure the temperature at tw o nearby locations, as shown in Figure 610. The heat flux equation was calculated based on the timedependent experimental results to account for the unsteady effects of heat transfer. The heat flux measured along one of the combustion chamber walls and is determined using sensors that measure the temperature at tw o nearby locations. The steadystate (SS) heat flux was then calculated using ,linearSSk qT y (61) where conduction coefficient k of Copper 110, the material used for the combustion chamber wall, is k = 388 W/mK y is the distance between the ther mocouple holes as shown in Figure 610, and T is the difference in temperature between the two thermocouple holes. The heat flux profile using this definition was thought to be inaccurate. Thus, the equation was altered to include a heat absorption correc tion which accounts for unsteady (U S) effects of heat transfer. ,2,1 ,,2,22oo linearUSioTT kcy qTT yt (62) where the subscript i denotes the thermocouple location cl osest to the combustion chamber wall and o denotes the outer thermocouple location. The subscripts 1 and 2 represents an initial and final time, respectively. The the density and heat capacity c of Copper 110 are given by = 8700 kg/m3 and c = 385 J/kgK. The chamber wall temperat ures were determined using the heat flux data and the thermal conductivity of the wall.
PAGE 164
164 linearSS walllineariqy TT k (63) In the experimental documentation, the steady and unsteady heat fluxe s are calculated using Equations 61 and 62, respectively, and are shown in Figure 611. Equations 61 through 63 assume that the heat flow is onedimensional.179 For a square duct, the onedimensional linear assumption may not be correct. The actual equation would need to be determined through a numerical solution of the twodimensional heat conduction equation. The correct heat flux equation would change the experimentally documented values of heat flux at the wall. However, because at each cross section only two temperature measurements are made, there is not enough data to do a full twodi mensional analysis of heat transfer through the combustion chamber wall. Additional temperature m easurements taken within the wall would be required for an accurate estimation of wall heat flux. Therefore, a onedimensional axisymmetric approximation was included in addition to the onedimensional linear approximation. If it is assumed that the isotherms within the combustion chamber wall are axisymmetric, a second onedimensional assumption can be conducted. The geometry of the isotherms is given in Figure 612. This is reasonable due to th e approximate axisymmetr y of the temperature contours within the combustion chamber wall, a nd the high thermal conductivity of the copper. For 1D axisymmetric heat c onduction, the steady state heat flux approximation is given by ,lnaxisymmSS walloikT q rrr (64) where rwall is the distance from the center of the combustion chamber to the combustion chamber wall, and ri and ro are the distances from the center of the combustion chamber to the inner and outer thermocouples, respectively, as shown in Figur e 612. Using this equa tion, the heat flux at the wall centerline can be determined based on the 1D heat conduction equation, giving
PAGE 165
165 ,,lnwalloi wallaxisymmlinearSS iwallrrr qq rr (65) For the given combustion chamber diameter, this assumption would result in a wall heat flux value that is everywhere 28% be low the linear heat flux assumption. This is reasonable, as the assumption of axisymmetric isotherms would resu lt in the lowest possible heat flux given the temperatures at the thermocouple locations. The temperature at the wall would be given by ln lnaxisymmo iooTr TrT rrr (66) An unsteady approximation is determined by adding the second term from equation (62) to the steady state heat flux approxima tions in Equation 65. Figure 613 shows the estimated wall temperature profiles based on the experimental thermocouple readings. The estimated wall temperatures are very similar between the line ar and axisymmetric approximations, indicating that the linear assumption used in the experime ntal documentation was reasonable in the case of the wall temperature. Finally, an approximation of the wall heat flux was conducted using a numerical twodimensional unsteady heat conduction analysis. Vaidyanathan et al.180 completed a similar analysis on a different injector and provided the conduction analys is results for the injector by Conley et al.22 The crosssections at each thermo couple pair location were analyzed independently. Because the experiment lasted only seven to eight seconds it was estimated that only negligible heat escaped through the outer combustion chamber wall. Therefore, the outer wall was assumed to be adiabatic. The thermal propagation time scale for the combustion chamber wall is equal to tP = 2/ = 3.13 s where is the thickness of the combustion chamber wall equal to 0.019 m, and is the thermal diffusivity and is equal to 1.158 104 m2/s. The
PAGE 166
166 thermal propagation time scale is on the same orde r of magnitude as the e xperimental run time of 7.75 s, indicating that the adiaba tic assumption is reasonable. Th e inner wall heat flux, taken to be uniform along the inner combustion chamber wa ll, was adjusted until the computational and experimentally measured temperatures obtaine d at 3.15 and 9.54 mm fr om the inner chamber walls (see Figure 610), respectively, are matched. The resulting temperature contours are given in Figure 614 for the prescribed boundary cond itions for the combustion chamber crosssection at x = 84 mm. Significant differences are seen in the comparis ons of the estimated wall heat fluxes shown in Figure 615. The shapes of the profiles given by the linear and axisymmetric assumptions are the same, albeit the axisymmetric assumption gives a proportionally lower heat flux value. The peak heat flux value predicted using the axisymmetric assumption is similar to that predicted by the 2D unsteady conduction analysis. The locati on of the peak heat flux from the conduction analysis is shifted from that of the 1D approximations. Conley et al.22 found that the point of maximu m heat release using the linear approximation is at 58 mm from the injector fa ce, whereas the maximum temperature occurs further downstream at 84 mm from the injector face. On the other hand, the point of maximum heat release using the unsteady 2D analysis gives a maximum heat flux at the same location as that of the maximum temperature. 6.5 Injector Flow Modeling Investigation A CFD modeling investigation was conducted ba sed on the experimental results presented by Conley et al.22 of a singleelement shear coaxial GO2/GH2 injector. The CFD model is used to supplement the experimental results, and to ga in insight into releva nt flow dynamics. The modeling effort is particularly useful in capturing the variatio n of heat flux along the combustion chamber wall due to threedimensional flow effects.
PAGE 167
167 The CFD simulations were conducted using lociSTREAM157 which is a pressurebased, finiterate chemistry solver for combustion fl ows for arbitrary grids developed by Streamline Numerics, Inc. at the University of Florida. The code numerically solves the 3D unsteady, compressible, NavierStokes equations. MenterÂ’s SST model is used for turbulence closure. A sevenspecies, nineequation finite rate chemistry model181 was used for gaseous hydrogen and gaseous oxygen. 6.5.1 CFD Model Setup The computational domain, grid, and boundary conditions are shown in Figure 616. The values for specific flow conditi ons are given in Table 61. The unstructured grid contains 1.7106 elements and 3.7105 nodes. The computational domain represents a threedimensional, oneeighthsection of the full combustion chambe r. Noslip conditions are specified at the combustion chamber wall, the injector face, and the oxygen post tip. Symmetry conditions are specified along the planes of symmetry and along the centerline of the combustion chamber. The nozzle is not included in the analysis; instead, a constant pressure outlet condition is specified, as the pressure is nearly constant throughout the combus tion chamber. The initial temperature was set to 2000 K to place initial heat flux value within the range of the experimental measurements. A short ignition region with T = 3000 K was used in the recirculation region of the shear layer to start combustion. The ignition was turned off after the flame was selfsustaining, and the computation was run until a steadystate condition was reached based on the boundary conditions given. Applying an adiabatic c ondition to the combustion chamber wall was considered, but this condition would erase th e thermal boundary layer along the wall. The thermal boundary layer is essential to determining the temperature gradient at the wall, and thus, the heat flux. A constant temper ature condition was considered a su itable alternative, and would represent the case wherein the temperature along the wall was unknow n. A constant temperature
PAGE 168
168 of T = 500 K was specified along the combustion chamber wall. This temperature value was arbitrarily selected as it is w ithin the approximate range of the experimental value, but is not equal to the average experimental temperature at the wall. Results when the experimental temperature profile is used along the combustion chamber wall are not included, as there are no obvious differences in the flow contours whethe r the experimental temp erature profile or a constant temperature is applied at the wall. The injector face and the oxidizer post tip region between the hydrogen and oxygen inlets are set to adiabatic conditions. 6.5.2 CFD Results and Experimental Comparison of Heat Flux In the computation, the heat flux at a location j is calculate d using the following equation: jwallfluid jT qk y (67) The calculated heat flux based on Equation 67 wa s calculated in the CFD simulation and then compared to the experimental values. Figure 617 shows the presence of a strong recirculation region with relatively slow fluid flow. The hydrogen flow enters at a greater velo city than the oxygen flow, but quickly slows due to mixing. Due to the low Mach number of 0.04 and the relativ ely constant pressure throughout the combustion chamber, the density changes th at occur within the combustion chamber are essentially due to changes in temperature. Figu re 618 shows the temper ature contours along the z = 0 plane. The combustion length, or the dist ance from the injector face where 99% of the combustion is complete, is at 38 mm. The steady state, onedimensional heat transf er approximation results in a linear heat flux and temperature relationship. If one accounts for the area variations via a steady state, axisymmetric treatment, then there is a different heat flux and temperatur e profile relationship. Finally, if one chooses to treat the outer wall boundary as adiabatic, to reflect the short
PAGE 169
169 experimental run time, then one can develop an unsteady, twodimensional representation between the heat flux and the temperature distribu tion as described in Section 6.4. These three cases, labeled as (i) linear onedimensional, (ii) axisymmetric, and (iii) 2D unsteady adiabatic are shown in Figure 619 along with the co mputed result. Figure 620 shows the twodimensional nature of the heat flux along the wa ll. Overall, the agreement between the CFD results and the experiment fair The overall trends from the CFD data and the 2D conduction analysis are similar, but shifted. The shift in th e CFD data as compared to the results of the 2D conduction analysis indicates a fast er heat release in the CFD simulation than may actually occur in the experiments. In all cases, the CFD data underpredicts the experimental data. Clearly, there are substantial uncertainties in converting limite d temperature measurements within the wall to heat flux profiles on the wall. There is also uncertainty in the CFD results that may be due to the resolution of the computational domain. The eff ects of the grid resolu tion on the experimental injector element are expl ored in Section 6.6. 6.5.3 Heat Transfer Characterization The threedimensional simulation allows for a view of the heat flux characteristics across the flat wall. Figure 620 shows the twodimensi onal nature of the predicted heat flux along the wall. Also shown in Figure 620 is the eddy con ductivity. The location of peak heat flux can be clearly seen in Figure 620A, and corresponds to the maximum eddy conductivity value along the combustion chamber wall. The heat flux contour s show where the heated gases are circulated back towards the injector face located at the left of the plot. Additional measurements would be required for more accurate comparisons of the twodimensional heat flux and temperature. Across x = 24 mm mark, there is a sudden jump in heat flux magnitude that corresponds to the change in xvelocity from positive to negative as shown in Figure 621. The sudden jump in heat flux is shown in Figure 622A. The locat ions of stagnation region resulting from the x
PAGE 170
170 velocity change and the streamlin e reattachment points are clearly apparent in the dips in the y+ values shown in Figure 622B. The effect of the velocity change is echoed in the temperature values near the wall (y+ ~ 2.5) as seen in Figure 623A. The mixing action of the shear layer transports heat from the flame to the wall. The point of maximum heat flux occurs at 63 mm. This corresponds to the point of maximum eddy conductivity that also occurs at 63 mm from the injector face, shown in Figure 623B. Similar to the experimental results, the peak wall temperat ure does not occur in the same location as the maximum heat flux. On the other hand, the ma ximum temperature along the wall occurs at 72 mm, but is essentially constant between 64 mm a nd 79 mm. This is immediately upstream of the reattachment point which is 80 mm from the inje ctor face. The disconnection between the peak heat flux and the peak wall temperature can be re lated to the ratio of eddy viscosity to laminar viscosity. The wall heat flux prof ile corresponds to the eddy viscos ity, and similarly the thermal conductivity. The temperature profile on the other hand corresponds to the laminar, or molecular viscosity. 6.5.4 Species Concentrations From the temperature profile given in Figure 618 at y = 0, it can be seen that much of the reaction occurs before 72 mm. This is also evident in the ma ss fraction contours in Figure 624. The reaction proceeds beyond that at a much sl ower rate as the system moves toward equilibrium. The species concentration thr ough the flame center is shown in Figure 625. Figure 625 shows the diffusion of hydrogen towa rds the centerline between 0 and 22mm. The mole fraction of H2 increases at y = 0 due to the consumption of O2. The y = 2 profile is located just beyond the H2 inlet, so the increase in H2 reflects the spreading of the H2 stream. Because the reaction is fuel rich, the hydrogen is not comp letely consumed. There is almost the almost complete disappearance of gaseous oxygen by approxima tely 53mm. This is also consistent with
PAGE 171
171 the rise in temperature that o ccurs at this location. This consumption of oxygen and the high temperatures gives rise to intermediate species. Species concentrations and temperatures near the wall (y+ ~ 2.5) show a jump near 24mm where vx = 0 similar to what was seen in the temper ature and enthalpy profiles. This effect is still found somewhat away from the wall. Fi gure 626 shows the homogeneity of the combustion products beyond approximately 72 mm. By this point, the products are comprised almost completely of steam and unburned hydrogen. 6.6 Grid Sensitivity Study For the grid sensitivity study, the injector element used is the same as that from the modeling study along with boundary conditions that simulate a single injector element near a combustion chamber wall. In preparation for th e later parameter study in Chapter 8, the CFD model attempts to approximate a nearwall inj ector element. The subdomain is approximated using a rectangular shaped computational domai n with width and height of 4.03 mm, and a length of 100 mm. Only onehalf of the square cross section is si mulated, with a symmetry plane defined along the symmetric boundary. The boundary conditions for all cases are give n as shown in Figure 627. In the CFD model, the distance to the wall is held equal to the distance to the inner slip. The purpose of the inner slip boundary is to keep the computational doma in approximately symmetric. A constant temperature of T = 500 K was specified along the combustion chamber wall. The remaining boundary conditions are the same as those used in Section 6.5.1. The injector face and the oxidizer post tip region between the hydrogen and oxygen inlets were set to adiabatic conditions. The nozzle is not included in the CFD model, and a constant pressure condition is specified at the outlet. A constant mass flow rate for hydroge n and a constant mass flow rate for oxygen are
PAGE 172
172 specified at the inlets, with gaseous oxygen ente ring through the center core of the injector and hydrogen entering through th e outer annular region. The effects of modifying the grid resolution we re investigated via a grid sensitivity study. The computational grids investigated containe d from 23,907 points for the coarse grid 103,628 points for the finest grid. The number of points an d smallest grid density used for each grid is given in Table 62, and each grid had a grid distribution as shown in Figure 628. A layer of prism elements were used near the wall so that the turbulence equations could be integrated to the wall without the use of wall functions which might reduce accuracy. The prism layers were constructed with an exponential distribution with an increasing number of prism layers used as the initial wall layer height was decreased. The specifications for the initial height of the prism layers are given in Table 62. The coarsest grid solution (grid 6) predicts a mu ch sharper rise in the wall heat flux that of the fine grid solution (grid 1), a nd underpredicts the peak heat flux as seen in Figure 629. This indicates that performance could suffer when usi ng a very coarse grid. The largest difference in the results can be seen in the length of the co mbustion zone in Figure 630. The coarsest grid solution (grid 6) predicts a much shorter combustion zone than the finest grid solution. Using the coarse grid leads to much more smearing along the reaction zone than the two finest grid solutions, grid 1 and grid 2. The species diffuse fast er in grid 6 as compared to the grids 1 or 2. However, the velocity field is less affected by th e grid density than the species diffusion. Thus, the mixing that is responsible for convecting heat to the combustion chamber wall leads to heat flux profiles that are less sensitive to grid de nsity. Significant differences are seen in the CFD solutions as the grid resolution is increased to its finest level. Numerical unsteadiness was seen in
PAGE 173
173 the finest grid solution (grid 1) shown in Figure 630. For this reason, the peak heat flux recorded for grid 1 is not considered the Â“trueÂ” peak heat flux. In this case, true grid independency cannot be confirmed. Howeve r, there is relative consistency between the values of the peak heat flux, mean heat flux, and combustion length for the intermediate grids. Over all, the intermediate grid s with 31,184 nodes and 43,680 nodes provide arguably the best tradeoffs between r un time, numerical steadiness, and accuracy as compared with other grid resolutions. A grid with a resolution between that of grids 1 and 2 was used for the CFD modeling study in the Section 6. 5. This indicates that significant improvement in the heat flux profile would not necessarily be obtained by increasing the grid resolution. For the same injector element and mass flow rates used for the grid sensitivity study and for the modeling study, the reduction in the com bustion chamber area appears to result in a reduction of the combustion length. However, because of the numerical unsteadiness of the finest grid, it is difficult to determine whether this is due to the combustion chamber geometry, or due to inadequate grid resolution. It is possible th at the combustion length may be, for the most part, only a function of the injector geometry and mass flow rate, and not a function of the combustion chamber geometry. The parameter study in Chapter 8 tries to determine some of the geometric effects on injector flow. 6.7 Conclusion The CFD model was able to qualitatively ma tch the experimental results. Uncertainty exists in the prediction of the peak heat flux and heat rise. However, the predicted location of the peak heat flux was consistent with the experime nt. The heat flux downstream of the reattachment point was well predicted. In addition, the CFD resu lts were able to provide additional insight into the combusting flow. In particular, the relationshi p between the flow dynamics and heat transfer could be characterized. It was found that the e ffect of the 3D square geometry shaped the
PAGE 174
174 recirculation in a way that noticeably affected the heat transfer profile. The heat flux along the wall is further shaped due to the recirculati on of fluid that has had time to cool along the combustion chamber wall, resulting in a relatively cool region of fluid n ear the injector face. Beyond the reattachment point, the ho mogenous fluid cools at a relati vely constant rate. The heat transfer within the combustion chamber can be characterized via the eddy conductivity which further relates to the peak heat flux at the wall. Additional experimental temperature measurements could improve the CFD modeling process by providing some measure of the 2D wall heat transfer effects. A dditional measurements could also improve the accuracy of the estimated wall heat flux and temperatures. A grid sensitivity study on the injector el ement used for the CFD modeling remained inconclusive. A grid independent solution could not be obtained by refining the grid. Numerical unsteadiness resulted with the finest grid. The co arsest grid predicted a sharper heat flux rise, a lower peak heat flux, and a shorter combustion le ngth than the finer grids. With the exception of the coarsest grid and the finest grids, the predicted heat flux remained largely unaffected. This indicates that a finer or coarser grid would likely have little eff ect of the heat flux predicted in the modeling study. However, the predicted com bustion length varied si gnificantly based on the grid resolution. Additional studies would be requ ired to determine the reason for the numerical unsteadiness. Table 61. Flow regime description. Inlet temperature, T 300 K Exit pressure, pe 2.75 MPa Mass flow rate, m 0.187 g/s Max Reynolds number, Remax 3000 Min Reynolds number, Remin 960
PAGE 175
175 Table 62. Effect of grid resolution on wall heat flux and combustion length. Grid # Grid points Minimum element size in shear layer/near wall (mm) Run time (min) qmax (MW/m2) qmean (MW/m2) Combustion length (mm) 1 103628 0.15 / 0.03109205.412.11 31.1 2 72239 0.2 / 0.0331505.082.03 34.7 3 43680 0.5/ 0.017444.802.09 30.7 4 31184 0.5 / 0.035884.932.06 30.7 5 25325 0.5/ 0.053904.732.07 30.6 6 23907 1 / 0.033924.472.03 23.2 Table 63. Flow conditions. Inlet hydrogen mass flow rate, 2 H m 0.396 g/s Inlet oxygen mass flow rate, 2Om 1.565 g/s Equivalence ratio, 2.0 Chamber pressure, p 2.75 MPa Initial combustion chamber temperature, T 2000 K Average Mach number, M 0.04 Figure 61. Blanching and cr acking of combustion chamber wa ll due to local heating near injector elements.
PAGE 176
176 y INLET OUTLET x N OSLIP N OSLIP N OSLIP z Figure 62. Hydrogen flow geometry. Figure 63. Hydrogen inlet mesh.
PAGE 177
177 A B Figure 64. Zvorticity contours. A) = 0 and B) = 5.7. 0.10 0.08 0.06 0.04 0.02 0.00 0.02 0.04 0.06 0.08 0.10 61434547494 xSwirl Number, S0 50 100 150 200 250area (mm2) = 0 = 5.7 area Figure 65. Swirl number at each x location.
PAGE 178
178 0 10 20 30 40 50 60 20020406080100 xu0.10 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70v u, = 0 u, = 5.7 v = 0 v = 5.7 Figure 66. Average axial velocity u and average tangential velocity v with increasing x. The axial velocity is two orders of magnitude higher than the tangential velocity. A 0 500 1000 1500 2000 2500 3000 20020406080100120 xReu Reu, = 0 Reu, = 5.7 B 0 100 200 300 400 500 600 700 20020406080100120 xRev Rev, = 0 Rev, = 5.7 Figure 67. Reynolds number profiles. A) ax ial velocity and B) tangential velocity. 0.0E+00 5.0E04 1.0E03 1.5E03 2.0E03 2.5E03 20020406080100120 x(ppo)/po turbulent, = 0 turbulent, = 5.7 Figure 68. Nondimensional pressure as a function of x.
PAGE 179
179 A 0.00 0.20 0.40 0.60 0.80 1.00 1.20 1.40 1.60 0.0020.0040.0060.0080.0 0 vXr B 0.00 0.20 0.40 0.60 0.80 1.00 1.20 1.40 1.60 0510152025 kr C 0.00 0.20 0.40 0.60 0.80 1.00 1.20 1.40 1.60 1.E+041.E+051.E+061.E+071.E+081.E+09 omega r Figure 69. Hydrogen inlet flow profiles. A) Axial velocity, B) turbulent production k, and C) turbulent dissipation 25.4 63.5 + 3.2 19.1 19.1 9.5 15.9 1.6 1.6 Thermocou p les Figure 610. Combustion chamber crosssectiona l geometry and thermocouple locations. Units are given in millimeters.
PAGE 180
180 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 1 1.5 2 2.5 3 3.5 4 x(m)q (MW/m2) Steadystate approximation Unsteady approximation Figure 611. Estimated wall heat flux using li near steadystate and unsteady approximations. To Twall T i rwall r i ro Thermocou p les Figure 612. 1D axisymmetric assumption for heat conduction through combustion chamber wall.
PAGE 181
181 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 400 450 500 550 600 xTwall(K) linear axisymmetric Figure 613. Estimated wall temperatures using linear and axisymmetric approximations. Linear and axisymmetric assumptions give nearly equal wall temperature estimations. Temperature values at x = 0m and x = 0.16m are estimated based on nearest temperature values. A 0 0.005 0.01 0.015 0.02 0.025 0.03 0.01 0.005 0 0.005 0.01 0.015 0.02 B 0 0.005 0.01 0.015 0.02 0.025 0.03 0.01 0.005 0 0.005 0.01 0.015 0.02 C 0 0.005 0.01 0.015 0.02 0.025 0.03 0.01 0.005 0 0.005 0.01 0.015 0.02 D 0 0.005 0.01 0.015 0.02 0.025 0.03 0.01 0.005 0 0.005 0.01 0.015 0.02 550 300 Figure 614. Temperature (K) contours for 2D unsteady heat conduc tion calculations at x = 0.084 m at A) 1 s, B) 2 s, C) 6.5 s, and D) 7.75 s. Units are given in meters. Onequarter of the combustion chamber cross sec tion is shown. The temperature scale is consistent across plots.
PAGE 182
182 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0 0.5 1 1.5 2 2.5 3 3.5 4 x(m)q (MW/m2)Unsteady approximation unsteady linear unsteady axisymmetric 2D unsteady adiabatic Figure 615. Experimental heat flux va lues using unsteady assumptions. The x values are given in meters. A B C Figure 616. Computational model for singleelem ent injector flow simulation. A) Combustion chamber boundary conditions, B) grid closeup, and C) representative combustion chamber section.
PAGE 183
183 Figure 617. Velocity contours vx(m/s) and streamlines. The injector is located in the lower right corner of the combustion chambe r. The reattachment point is at x = 77 mm. Figure 618. Temperature (K) cont ours. Injector center is at x,y = (0,0). Distances are given in millimeters. 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0 0.5 1 1.5 2 2.5 3 3.5 4 x(m)q (MW/m2) 1D linear axisymmetric unsteady adiabatic CFD Figure 619. CFD heat flux values as compar ed to experimental heat flux approximations.
PAGE 184
184 A zx q (MW/m2) 0 20 40 60 80 100 120 140 160 10 0 10 0 0.5 1 1.5 2 2.5 B C Figure 620. Wall heat transf er and eddy conductivity contour plots. A) Heat flux along combustion chamber wall (xzplane at y = 12.7) B) eddy conductivity at y = 12.3, and C) eddy conductivity at z = 0 with wall at y = 12.7 mm. Eddy conductivity contours are oriented perpendicular to wall he at flux contours. Horizontal axis is xaxis. (Solutions mirrored across x = 0). Length units are in mm. The injector is centered at x,y,z = (0,0,0), and the flow proceeds from left to right. Figure 621. Streamlines and temperature contours at plane z = 0. All lengths are in millimeters.
PAGE 185
185 A 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0 0.5 1 1.5 2 2.5 3 x(m)q (MW/m2)Unsteady approximation vx = 0B 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0 0.5 1 1.5 2 2.5 3 3.5 4 x (m)y+ vx = 0 reattachment point Figure 622. Heat flux and y+ profiles along combustion chambe r wall. A) Heat flux profile along the wall at z = 0 showing sharp increase in heat flux corresponding to the location of velocity vx = 0. B) The y+ values along the near wall cell boundary dip at stagnation regions. A B Figure 623. Temperature and e ddy conductivity profiles at various y locations on plane z = 0. A) Temperature. B) Eddy conductivity.
PAGE 186
186 Figure 624. Mass fraction c ontours for select species. H2 O2 OH H2O
PAGE 187
187 A B Figure 625. Mole fractions for all speci es along combustion chamber centerline (y = 0, z = 0). A) uniform and B) log scaling.
PAGE 188
188 A B Figure 626. Select species mole fraction profiles. A) H2 and B) H2O at plane z = 0.
PAGE 189
189 Figure 627. Sample grid and boundary conditions. Figure 628. Computational gr id along symmetric boundary. 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0 1 2 3 4 5 6 7 xq (MW/m2) 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 xy+ 23907 31184 72239 103628 23907 31184 72239 103628 Figure 629. Wall heat flux and y+ values for select grids. The distance from the injector face x is given in meters.
PAGE 190
190 Figure 630. Comparison of temp erature (K) contours for grids with 23,907, 31,184, 72,239, and 103,628 points, top to bottom, respectively. The finest grid shows timedependent oscillations.
PAGE 191
191 CHAPTER 7 MULTIELEMENT INJECTOR FLOW MODE LING AND ELEMENT SPACING EFFECTS 7.1 Introduction Because of the arrangement of injector elemen ts near the wall of the combustion chamber, the heat flux peaks near each injector element. Chamber design and potential material failure is based on peak heat flux, so the peak heat flux mu st be minimized to increase component life and reliability. By controlling the injection pattern, a layer of cool gas near the wall can be created. One method of controlling the inj ection pattern is to offset the oxidizer post tip away from the wall. This results in a higher percentage of unburned fuel near the wall, and this technique was used in the SSME. The cooler layer near the wa ll can possibly further minimize the peak heat flux. Another method of controlling the heat flux is to change the spacing of injector elements near the combustion chamber wall. This research effort looks to quantify selected geometry effects by directly exploring the sensitivity of wall heating and injector performance on the injector spacing. Shyy et al.29 and Vaidyanathan et al.63 identified several issues inherent in CFD modeling of injectors, including the n eed for rigorous validation of CF D models and the difficulty of simulating multielement injector flows due to lengthy computational times. They chose to simulate single injector flow, as this allowed fo r the analysis of key co mbustion chamber life and performance indicators. The lif e indicators were the maximum temperatures on the oxidizer post tip, injector face, and combustion chamber wall, and the performance indicator was the length of the combustion zone. These conditions were expl ored by varying the impinging angle of the fuel into the oxidizer. In other words, the injector type was gradually varied from a shear coaxial injector to an impinging injector element. Re sponse surface approximations were used to model the response of the injector system, and a co mposite response surface was built to determine the
PAGE 192
192 injector geometries that provide d the best tradeoffs between performance and life. It was determined that a shear coaxial injector provi des the lowest maximum temperatures on the injector face and combustion chamber wall, but had the lowest performance. An impinging element provided the lowest ma ximum oxidizer post tip temperat ure and had high performance due to the short combustion length, but because the maximum injector face and combustion chamber wall temperatures were high, the life of the injector was compromised. The study confirmed that small changes in the injector geom etry can have a large impact on performance. The optimization procedure was successful in spec ifically accessing the effects of a small change in the injector geometry. This is important, as it allows for greater confidence in choosing the next design. This information can be coupled with the experi ence of injector designers to accelerate the design process. The purpose of the proposed research is to im prove the performance of upper stage rocket engines by raising the temperature of the fuel en tering the combustor. A multielement injector model is constructed. An effort is made to capture the 3D effects of a full multielement injector face in a simplified 3D model. The step s to improve the injector design are twofold: 1) the mean heat flux to the combustion chamber wall must be increased to increase fuel temperature and subsequently incr ease engine performance, and 2) the peak heat flux to the wall must be decreased to diminish the effects of lo cal heating on the chamber wall. The injector is optimized to best satisfy the objectives. The feasible design space is determined th rough exploration of the design space using available design space refinement and selection techniques. Afte r the feasible design space is determined, the optimization framework described in Chapter 2 is used to guide the injector
PAGE 193
193 element spacing study. The design space is refine d as needed to obtain an accurate surrogate model. 7.2 Problem SetUp This study explores the effects of geometry on the peak heat flux and combustor length. Approximately 90% of the heat flux to the combusto r wall is due to heat transfer from the outer row of injector elements. The second row of inj ector elements is respon sible for approximately 10% of heat flux to the wall. This information wa s used to simplify the computational model of a GO2/GH2 shear coaxial multielement injector to make it suitable for a parametric study. The injector face simplification of the full inj ector shown in Figure 71 is outlined below: 1. For the CFD model, only the outer row of elemen ts is considered. Currently, the effects and interactions of the outer el ement row with interior elements are simplified by imposing symmetric conditions. 2. The multielement injector effects are modeled by imposing symmetric conditions between injectors with the computational domain si zes corresponding to the actual radial and circumferential spacing of the full injector. Thus, the CFD model is a scaled down version of what would occur in an actual combustion chamber. The CFD model attempts to capture e ffects near the outer edge of the injector. The variables are the circumfere ntial spacing (number of injector elements in outer row), wherein the total mass flow rate through the outer row injector elem ents is held constant, and the radial spacing, such that the distance to the wall is held equal to the distance to the inner slip boundary. By this assumption, the mass flow rate through a single injector is equal to the total mass flow rate through the outer ro w of injectors divided by the num ber of injectors in the outer row. The purpose of the inner slip boundary is to keep the computationa l domain approximately symmetric, as the contributions and interactions from inner elements are neglected in the analysis. It is desired that the CFD model be ab le agree qualitatively with related experiments as well as establish the sensitivity of certain geometry changes on wall heating.
PAGE 194
194 The CFD simulations were conducted using lociSTREAM157 (Thakur and Wright 2006). The CFD model attempts to approximate the sing leelement section near the chamberÂ’s outer wall given in Figure 71. The singleelement section was approximated using a rectangular shaped computational domain. Only onehalf of the domain is simulated. The grid resolution used is equivalent to the intermediate grid 3 fr om the grid sensitivity study of Section 6.6. This grid was chosen for its balance of accuracy a nd relatively short runtime. The solutions are expected to give good heat flux accuracy, but the combustion lengths are expected to be somewhat underpredicted. In summary, the objectives are to minimize the peak heat flux qmax(N*,r*) and minimize the combustion length LC(N*,r*) where the independent variables are explored within the range max0.751.25 0.251.25baselineN r NNN (71) where **,baselinebaselineNNNrrr (72) where N is the number of injector elements, Nmax is the maximum possible number of injector elements in the outer row, and r is given as shown in Figure 71. The combustion length is the length at which combustion is 99% complete and a shorter combustion length indicates increased performance. A low peak heat flux would reduce the risk of wall burnout. Although the number of injector elements in the outer row changes, the total combined mass flow rate to the outer elements also remains constant, so the mass flow rates for the in dividual injectors are a function of the number of injector elements N. The distance w of the computational domain in Figure 71B is given by
PAGE 195
195 tan 2 D wNR N (73) where R is the distance from the center of the injector to the center of an outer element injector element and D is the injector element diameter. The baseline injector element geometry is ba sed on the experimental test case RCM1 from Section 4.1.4, while the overall injector a nd element spacing is based on the Integrated Powerhead Demonstrator159 (IPD) main injector. For this injector, the maximum number or injector elements Nmax allowable in the outer row is 101. The flow conditions and injector geometry for the baseline injector are given in Table 71. The inlet flow represents incompletely burned flow from upstream preburners, so both th e fuel and oxidizer contain some water as a constituent. 7.3 Feasible Design Space Study The design points based on r* and N* are shown in Figure 72. The first seven design points were selected based on a pr eliminary design sensitivity study. Point 16 lies at the center of the design space. Latin Hypercube Sampling was then used to select the seven remaining design points within the variable ranges. Figure 73 shows the values of the objectives at each design point. It can be seen that the mean heat flux across the combustion chamber wall stays relatively constant across all of the design points. The maximum heat flux, however varies depending on the combustor geometry. The design points have an averag e mean heat flux of 8.2 MW/m2 and an average maximum heat flux of 37.3 MW/m2. In general, as the maximum heat flux becomes lower, the combustion length is longer, indicating slower combustion. For design purposes, there would need to be some balance between the maximum heat flux an d combustion length. From Figure 73B, it can be seen that case 6 shows good balance between minimizing the peak heat flux and minimizing
PAGE 196
196 the combustion length, while case 12 gives argu ably the worst characteristics due to its high maximum heat flux and long comb ustion length. Figure 73 also shows the maximum heat flux and combustion length as a function of the hydroge n mass flow rate. It is apparent that the maximum heat flux is not a function of the mass fl ow rate alone, but that there must be other influencing factors. There does seem to be an ove rall decrease in the maximum heat flux with the mass flow rate, as well as an overall increase in combustion length with increasing mass flow rate, which is a function of N* No obvious relationship was seen between the maximum heat flux or combustion length and th e crosssectional area. This i ndicates that, a smaller crosssectional area does not automatically translate into a longer combustion length, as one might expect. This was also demonstrated by the grid sensitivity study. The CFD results for four cases, represen ting the baseline case along with cases demonstrating good or poor heat flux characteristics, are shown in Figure 74. It can be seen that the flame is flattened somewhat based on the sh ape of the computational domain. Another major observation is of the location of the recirculation region. Instead of hot gases being directed at the wall, as in the baseline case and case 12, th e recirculation region of case 2 and case 6 is located between injector elements, and the str eamlines are largely parallel to the combustion chamber wall. This led to an observation that a ll of the CFD solutions resulted in one of two scenarios. Eleven of the cases showed a sharp pe ak in the wall heat flux near the injector face with the downstream heat flux considerably lowe r than the peak heat flux. This is termed Â“Pattern 1,Â” and corresponds to cases that have ho t gases directed at the wall as in Figure 75A. The remaining five cases had a relatively low he at flux near the combustion chamber wall that grew gradually to a slight p eak further downstream. This gr oup is termed Â“Pattern 2,Â” and corresponds to the cases that have a recircula tion zone that would exist between injector
PAGE 197
197 elements. The two groups can be clearly seen in Figure 76 in the plot of maximum heat flux versus r* Figure 77 shows the heat flux profile and contours for case 12 of Pattern 1, and case 6 of Pattern 2. In Pattern 2, the heat flux is sp read across the combustion wall, rather than being localized as were the cases from Pattern 1. The best performing designs w ith regards to heat flux appear to be those with small distances between the injector and the wall. However, of the cases in Pattern 2, the worst cases are cases 7 and 9 th at have smaller domain widths as compared to the other cases. This would suggest that the peak heat flux might be a function of the domainÂ’s aspect ratio, rather than separate ly being a function of radial or circumferential spacing, alone. The domainÂ’s aspect ratio is given by AR = r* / w* where w* = w / wbaseline, and wbaseline is determined based on N = Nbaseline in Equation 73. It can be seen in Figure 78 that the value of the peak heat flux appears to be strongly depe ndent on the aspect ratio. The designs with the lowest peak heat flux also have th e lowest aspect ratio. This indi cates that it may be desirable to have a low ratio of wall spacing (radial spacing) to injectorinjector sp acing (circumferential spacing) to reduce the peak heat flux. Due to the presence of case 5, it is difficult to determine whether the cases approximately follow a linear tre nd and case 5 is an ou tlier, or if the actual trend is quadratic. It appears that the peak heat flux is highly sensitive to the combus tion chamber geometry as evidenced by Figure 78. The re lationship of the combustion length to geometry is less clear. It appears that there is some sensitivity to the ge ometry due to the fact that equivalent injector elements with equal mass flow rates can result in combustion lengths that can differ significantly (see points 2, 4, and 6 in Figure 73D). In gene ral, the cases in the region defined by Pattern 1 have shorter combustion lengths than those in Pattern 2. This indi cates that increasing the size of the recirculation zone, and hence the mixing, may increase the combustion efficiency. An
PAGE 198
198 individual analysis on a single injector element of the eff ect of the combustion chamber geometry on combustion length with constant mass flow rate may be necessary to determine the relationship. The best tradeoff points are located in one or the other region of the design space, as shown in Figure 79. Because the behavior of th e design points in the different regions are very different, this indicates the need for two sepa rate design spaces. The dividing line for the new design space is given approximately by the dash ed line in Figure 79B. The design space will also be expanded slightly to *0.601.25 0.201.25 N r (74) due to the best tradeoff poi nts of pattern 2 having low N* and r* values. While it does appear that the combustion length is definitely affected by the geometry, the results of the feasible design space study were not able to reveal the direct relationship between combustion length and geometry. Based on a comput ational model for an injector element near the combustion chamber wall of a multielement in jector, it was found that an injector element that is located far from the combustion chamber wall does not always result in the best heat transfer characteristics. The resu lts of the feasible design spa ce study suggest that by increasing the spacing between injector elem ents of the outer row, while re ducing the distance of the outer row to the wall, that the heat tr ansfer results could be better c ontrolled. This configuration helps to direct heat away from the wall, rather than towards it, and results in an even distribution of heat across the combustion chamber wall. To a lesser degree, increasing the distance to the combustion chamber wall can also result in reduced peak heat fl ux, but the required spacing may possibly be prohibitively large.
PAGE 199
199 7.4 Design Space Refinement Using the merit function smart point selec tion procedure introduced in Section 2.4.3, along with an assigned rank using the formulation for multiobjective problems introduced in Equation 238, additional points were added to the newly expanded design space. The merit function that maximizes the probability of improvements MF2 was chose for point selection, as it resulted in the least error for a small number of sampling point s as was shown the radial turbine analysis in Section 3.4. The design space refinement proceed ed over several steps. The steps are listed briefly, then descri bed in more detail: 1. Kriging RS was constructed for each objective over 16 initial design points. Based on the results initial Kriging RS, the design space vari able ranges were expanded due to favorable function values on the edge of design space. Four points were selected via the multiobjective merit functionbased point selection crit eria to add to the original 16 points for a total of 20 designs. 2. The design space was separated based on flow patterns into two regions called Pattern 1 and Pattern 2. A Kriging RS was construc ted for each objective for the 12 out of 20 designs in Pattern 1. Three new points were selected using MF2 for a total of 15 designs in Pattern 1. 3. A Kriging RS was constructed for each objective for the 8 out of 20 designs in Pattern 2. Four new points were selected using MF2 for a total of 12 designs in Pattern 2. 4. A Kriging RS was constructed for each obj ective for the 15 desi gns in Pattern 1. 5. A Kriging RS was constructed for each obj ective for the 12 desi gns in Pattern 2. 6. For visualization of the full design space, a Kriging response surfaces were fit for each objective to all 27 design points. The error statistics for all six steps ar e given for each objective in Table 72. A Kriging surrogate was fit to the entire de sign space before the design space separation. The merit function value was calculated for point s sampled across the design space as shown in Figure 710. The densely sampled poi nts were divided into eight clusters based on location. The point with the best rank in each cluster was select ed. Out of the eight resulting points, four were
PAGE 200
200 chosen based on their favorable objective function values predicted by th e surrogate model. The Kriging surrogate results along w ith the four selected points ar e shown in Figure 711. While Kriging can provide an adequate surrogate model for the peak heat flux, Figure 711B shows that the design space is too large for an appropriate su rrogate fit for the combustion length. When the PRESSrms is calculated as a percent of the objectives sampling range or the standard deviation of the objectives, it can be seen that the PRESSrms error for the LC surrogate fit is about twice that of the qmax surrogate. The design space was thus divited into two smaller regions to try to improve the surrogate fits. Fifteen of the 20 design points were located in the region for Pattern 1, while eight of the 20 design points were located in the region for Patte rn 2. Kriging fits to the individual regions increased the scaled prediction error in each region. In particular the Kriging scaled prediction error for Pattern 2 increased significantly due to the small number and poor distribution of the data points located in the region. To improve the surrogate fits, additional data points were added to each region. Three points were added to the Pattern 1 region, and four points were added to the Pattern 2 region based on merit function evaluations and favorab le objective function values. The Kriging surrogate fit and merit function contours for Patte rn 1 and Pattern 2 are given in Figure 712 and Figure 713, respectively. Figure 712 indicates the locations of the th ree points added to the design space in the region of Pattern 1. Figur e 713 indicates the locatio ns of the four points added to the design space in the region of Pattern 2. These point s were chosen out of the points selected by the merit function analysis based on high uncertainty and high promise of function improvement. Regions where the predicted function values were very poor were avoided. This would have the overall effect of refining the design space only in the region of the best tradeoff
PAGE 201
201 designs. The resulting Kriging f its including the new data points are shown for Pattern 1 and 2 in Figure 714 and Figure 715, respectively. Th e response surface for peak heat flux of Pattern 1 in Figure 714 shows smooth contours, indicating a good surrogate fit. The prediction errors of the Kriging fits echo what can be casually observe d from the contour plots. For Pattern 1, the addition of points slightly decreases the scaled error in qmax, and significantly decreases the scaled error in the prediction for LC. The prediction of the com bustion length of Pattern 2 was improved by separating the design space into two re gions as evidenced by the scaled prediction error. The fit of the peak heat flux became wo rse for Pattern 2 after splitting the design space into two regions, but this may be due to th e sparseness of data points in the region. When the data points were plotted as combus tion length versus peak heat flux in Figure 716, it was discovered that all of the newly a dded points dominated the points in the original Pareto front. Overall, the heat flux was re duced by up to 20% and the combustion length was reduced by up to 5% among the eleven new points as compared to the best tradeoff points from the original data set of 20 points. This indicates the success of using merit functions to select points for design space refinement. This is a cas e where a surrogate based on a simple DOE can be used to positively identify the Â“badÂ” regions so that they may be avoided. In this case, the original surrogates were not accurate in regi ons away from the original design points, particularly in the prediction of the combustion le ngth, but they are suffici ent to identify regions that are obviously bad. By ignoring obviously ba d regions of the design space, new data points could be concentrated in the best tradeoff re gion. The original surrogate model was also successful in identifying the regions of the de sign space that may result in function improvement, as six of the eleven new points lay in the expanded portion of the design space.
PAGE 202
202 A Kriging fit to all points of the design space shown in Figure 717 show a smooth response surface for the prediction of the peak h eat flux. The highest peak heat fluxes are predicted in Pattern 2. This is to be expected, as these cases have a recirculation zone that directs hot gases to the combustion chamber wall. This feature allows for a new approximation of the boundaries between the two regions of the design space. Cases that lie very near the division, such as case 20 at N* = 0.6 and r* = 0.9, show characteristics of both regions with a recirculation zone evenly situated about the injector, rather th an in the vertical or horizontal directions, as shown in Figure 718. The longest combustion le ngths, or worst combustion efficiencies, occur in the points in Pattern 2. This is likely due to the smaller crosssectional area combined with mass flow rates that are larger, in general, than the designs in Pattern 1. However, for very low r values, the combustion efficiency improves. Based on Figure 716, cases 18 and 19 were chos en as the best tradeoff designs. These designs exist on opposite sides of the design space and in different pattern regions, as shown in Figure 719, but have similar number of injectors in the outer row of the multielement injector, as shown in Figure 720. The cases also have si milar heat flux values and combustion lengths. Compared to the baseline case, case 19 predicts a 45% reduction in heat flux as compared to the baseline case while maintaining the combustion lengt h. Case 19 predicts a 38% reduction in the heat flux and a 5% reduction in the combustion length as compared to the baseline case. Figure 721 shows the predicted heat flux profile for the selected tradeoff cases as compared to the baseline cases. The grid reso lutions of the baseline case a nd one tradeoff solution were moderately increased and compared to the solutions used for the surrogate modeling. The coarser grid solution required less than one day to obtain a solution, wh ile increasing the number of grid points by a factor of tw o increased the run time to approxi mately five days. A moderate
PAGE 203
203 increase in the grid resolution has a minimal eff ect on the baseline peak heat flux, but increases the downstream heat flux, as shown in Figure 722A. However, the combustion length is 15% longer using a finer grid as compared to the grid used to construct the su rrogate model. For case 19, the tradeoff case from Pattern 1, and increase in the grid resolution results in a 45% increased prediction of peak heat flux over the coarser grid and a 10% increase in the combustion length. When comparing the fine grid solutions of the baseline case and case 19, the peak heat flux is only reduced by 16% and the combus tion length reduced by approximately 6%. Significant differences in the prediction can occur as a result of the grid resolution. However, the improvement in the finer grid solutions indi cates that the surrogate model can still provide significant information on the general trends in the data with which to improve the design. The peak heat flux has a strong dependence on the combustion geometry. The effects of geometry can be stronger than the mass flow rate dependence. It was observed by Conley et al.22 (2005) that the heat flux of an injector was strongly dependent on the injectorÂ’s mass flow rate, with the heat flux increasing proportionally with th e square of the mass flow rate. The injector spacing analysis, however found that the peak heat flux actually decreased with increasing mass flow rate, as shown in Figure 723. This shows th at there are definite effects of geometry on the peak heat flux. In contrast, the combustion leng th predictably increases with the mass flow rate, although the spread in the data indicates geometric sensitivity. 7.5 Conclusion The optimization framework was used to de termine the effects of combustion chamber geometry on the wall heat flux and combustion length. Although, the accuracy of the surrogate models may have been affected by the coarse gr id resolutions used in the study, the surrogate models provided an approximation of the de sign space that was sufficient enough to give predictions of variable combina tions that would result in improve ment over the baseline case.
PAGE 204
204 When additional points were selected based on larg e probabilities of improvement given as given by the merit function MF2, improvements in the function valu es were obtained for every newly selected variable combination as compared with th e original data set. This indicates that the search for best tradeoff designs can proceed efficiently with car eful selection of design points using merit functions. Based on a computational model for an inject or element near the combustion chamber wall of a multielement injector, it was found that an in jector element that is located far from the combustion chamber wall does not always result in the best heat transfer characteristics. The results of the multielement injector spacing study suggest that by increasing the spacing between injector elements of the outer row, wh ile reducing the distance of the outer row to the wall, that the heat transfer results could be bett er controlled. This config uration helps to direct heat away from the wall, rather than towards it, a nd results in an even distribution of heat across the combustion chamber wall. Sufficiently incr easing the distance to the combustion chamber wall can also result in reduced p eak heat. In the future, further exploration into each phenomenon should provide additional insight. Table 71. Flow conditions and baseline com bustor geometry for parametric evaluation. Fuel annulus outer diameter (mm) 7.49 Fuel annulus inner diameter (mm) 6.30 Oxidizer post inner diameter (mm) 5.26 Oxidizer post tip recess (mm) 0.43 Combustion chamber diameter (mm) 262 rbaseline (mm) 6.59 Nbaseline 60 Total fuel mass flow rate in outer row (kg/s) 1.986 Total oxidizer mass flow rate in outer row (kg/s) 5.424 Fuel temperature (K) 798.15 Oxidizer temperature (K) 767.59 Chamber pressure (MPa) 5.42 H2 mass percentage of fuel 41.3 H2O mass percentage of fuel 58.7 O2 mass percentage of oxidizer 94.62 H2O mass percentage of oxidizer 5.38
PAGE 205
205 Table 72. Kriging PRESSrms error statistics for each design space iteration. # Data set # of pts qmax a LC a PRESSrms, qmax PRESSrms, LC PRESSrms/ range( qmax) PRESSrms/ range( LC) PRESSrms/ std( qmax) PRESSrms/ std( LC) 1 Original 16 6.93 5.33 3.32 5.63 12.14 25.39 0.41 1.02 2 Pattern 1 12 7.34 4.67 5.18 5.08 17.77 34.67 0.67 1.04 3 Pattern 1 15 4.28 5.90 6.59 4.62 14.97 12.26 0.40 0.35 4 Pattern 2 8 9.89 4.29 1.93 2.24 21.56 31.54 0.65 1.05 5 Pattern 2 12 4.26 6.06 4.59 3.78 33.34 20.72 1.03 0.66 6 Final 27 10.13 5.15 4.96 5.01 15.87 22.59 0.51 0.97 range( qmax) = max( qmax) Â– min( qmax) range( LC) = max( LC) Â– min( LC) a = 2 = square root of the process variance given in Equation 211 A B C Figure 71. Injector element subsection. A) Oute r row representation of baseline injector with computational subsection shown at top and B) in closeup. C) Computational domain and boundary conditions on injector elemen t face for injector subsection. One element in the outer element row is simula ted. The circumferential spacing w is a function of the number of injector elements in the outer row.
PAGE 206
206 0.2 0.4 0.6 0.8 1 1.2 0.7 0.8 0.9 1 1.1 1.2 1.3 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 r*N* Figure 72. Design points selected for design sp ace sensitivity study. Point 1 indicates the baseline case. A 0 5 10 15 0 10 20 30 40 50 60 pointq qmax qmean B 20 30 40 50 6 0 75 80 85 90 95 100 105 qmax(MW/m2)combustion chamber length (mm)1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 C 0.025 0.03 0.035 0.04 0.045 20 25 30 35 40 45 50 55 mdot H2 (kg/s)qmax(MW/m2)1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 D 0.025 0.03 0.035 0.04 0.045 75 80 85 90 95 100 105 mdot H2 (kg/s)combustion chamber length (mm)1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Figure 73. Effect of hydrogen mass flow rate on objectives. A) Heat flux and combustion length data for original sixt een points, B) Combustion le ngth versus maximum heat flux, and C) maximum heat flux and D) combustion length versus the hydrogen flow rate.
PAGE 207
207 A B C D Figure 74. Oxygen isosurfaces and hydrogen cont ours. A) Baseline case (case 1). B) Worst overall case (case 12). C) Case with the lo west peak heat flux (case 2). D) Best overall case (case 6). The combustion cham ber wall is at the top of each figure. Approximately half of the designs showed solutions similar to A) and B), while the remainder was similar to C) and D). Solutions are mirrored across the z plane. A B Figure 75. Hydrogen contours and st reamlines. A) Case 12 and B) Case 6. The wall is located at the top of each plot. Case 6 shows the pe rsistence of a gaseous hydrogen layer near the combustion chamber wall.
PAGE 208
208 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 20 25 30 35 40 45 50 55 r*qmax(MW/m2)1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Pattern 1 Pattern 2 Figure 76. Maximum h eat flux for a changing radial distance r *. Point 1 is the baseline case. A 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0 10 20 30 40 50 60 x (m)q (MW/m2) case 6 case 12 B z(mm)q (MW/m2) 5 0 5 z(mm)x(mm) 0 5 10 15 20 25 30 35 40 5 0 5 0 5 10 15 20 25 30 35 40 45 50 Figure 77. Heat flux distributi on. A) Center of combustion chamber wall for case 6 and case 12. B) Heat flux distribution al ong first onethird length of entire wall for case 12 (top) and case 6 (bottom). Horizontal axis is x axis.
PAGE 209
209 0 0.5 1 1.5 2 2.5 20 25 30 35 40 45 50 55 AR = r*/w*qmax(MW/m2)1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Figure 78. Maximum heat flux as a function of aspect ratio. A 20 30 40 50 6 0 75 80 85 90 95 100 105 qmax(MW/m2)combustion chamber length (mm)1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 20 30 40 50 6 0 75 80 85 90 95 100 105 qmax(MW/m2)combustion chamber length (mm)1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 20 30 40 50 6 0 75 80 85 90 95 100 105 qmax(MW/m2)combustion chamber length (mm)1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 B Figure 79. Design points in function and desi gn space. A) Combustion length versus maximum heat flux, and B) the location of design point s in the design space. The best tradeoff points within each pattern are circled by a dashed line.
PAGE 210
210 A 0.5 0.49 0.48 0.47 0.46 0.45 r*N* 0.2 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 r*N* 0.2 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 0.5 0.49 0.48 0.47 0.46 0.45 r*N* 0.2 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 r*N* 0.2 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 B r*N* 0.2 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 r*N* 0.2 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 0.5 0.4995 0.499 0.4985 0.498 0.4975 r*N* 0.2 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 r*N* 0.2 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 0.5 0.4995 0.499 0.4985 0.498 0.4975 Figure 710. Merit function ( MF2) contours for A) qmax and B) LC. Dark blue indicates regions of lowest uncertainty. White points are data points. Black points are points chosen using merit function analysis. Points circled by dashed li nes indicate best tradeoff points. Circled black points are new points chosen for simulation. A r*N*qmax (MW/m2) 0.2 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 25 30 35 40 45 50 r*N*qmax (MW/m2) 0.2 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 25 30 35 40 45 50 B r*N*LC (mm) 0.2 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 80 82 84 86 88 90 92 94 96 98 100 r*N*LC (mm) 0.2 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 80 82 84 86 88 90 92 94 96 98 100 Figure 711. Kriging surrogates base d on initial 16 design points. A) qmax and B) LC. White points are data points. Black points are poi nts chosen using merit function analysis. Points circled by dashed lines indicate best tradeoff points. Ci rcled black points are new points chosen for simulation.
PAGE 211
211 A r*N*qmax (MW/m 2 ) 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 25 30 35 40 45 50 r*N*qmax (MW/m 2 ) 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 25 30 35 40 45 50 B r*N*LC (mm) 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 80 82 84 86 88 90 92 r*N*LC (mm) 0.4 0.6 0.8 1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 80 82 84 86 88 90 92 C D Figure 712. Kriging surrogates and merit function contours for 12 design points in Pattern 1. Kriging fit for A) qmax and B) LC and merit function values for C) qmax and D) LC. White points are data points. Black point s are points chosen using merit function analysis. Circled black points ar e points chosen for simulation.
PAGE 212
212 A r*qmax (MW/m2) 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.6 0.7 0.8 0.9 1 1.1 1.2 20 22 24 26 28 30 32 r*qmax (MW/m2) 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.6 0.7 0.8 0.9 1 1.1 1.2 20 22 24 26 28 30 32 B r*N*LC (mm) 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.6 0.7 0.8 0.9 1 1.1 1.2 84 86 88 90 92 94 96 98 100 r*N*LC (mm) 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.6 0.7 0.8 0.9 1 1.1 1.2 84 86 88 90 92 94 96 98 100 C 0.5 0.49 0.48 0.47 0.46 0.45 0.44 r*N* 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.6 0.7 0.8 0.9 1 1.1 1.2 r*N* 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.6 0.7 0.8 0.9 1 1.1 1.2 0.5 0.49 0.48 0.47 0.46 0.45 0.44 r*N* 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.6 0.7 0.8 0.9 1 1.1 1.2 r*N* 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.6 0.7 0.8 0.9 1 1.1 1.2 D 0.5 0.499 0.498 0.497 0.496 0.495 0.494 r*N* 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.6 0.7 0.8 0.9 1 1.1 1.2 r*N* 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.6 0.7 0.8 0.9 1 1.1 1.2 0.5 0.499 0.498 0.497 0.496 0.495 0.494 r*N* 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.6 0.7 0.8 0.9 1 1.1 1.2 r*N* 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.6 0.7 0.8 0.9 1 1.1 1.2 Figure 713. Kriging surrogates and merit functi on contours for 12 design points in Pattern 2. Kriging fit for A) qmax and B) LC. Merit function values for C) qmax and D) LC. White points are data points. Black points are point s chosen using merit function analysis. Circled black points are points chosen for simulation. Points in Pattern 1 region are ignored.
PAGE 213
213 A r*N*qmax (MW/m2) 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 20 25 30 35 40 45 50 B r*N*LC (mm) 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0.6 0.7 0.8 0.9 1 1.1 1.2 80 82 84 86 88 90 92 Figure 714. Kriging fits for all 15 design points from Pattern 1. A) qmax and B) LC. A r*N*qmax (MW/m2) 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.6 0.7 0.8 0.9 1 1.1 1.2 20 22 24 26 28 30 32 B r*N*LC (mm) 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.6 0.7 0.8 0.9 1 1.1 1.2 82 84 86 88 90 92 94 96 98 100 102 Figure 715. Kriging fits for all 12 design points from Pattern 2. A) qmax and B) LC.
PAGE 214
214 Figure 716. Pareto front based on original 16 data points (dotted line) and with newly added points (solid line). Use of surrogate model plus merit functions as selection criteria along with refinement of the design space resu lted in all eleven new points (circled) dominating all original sixteen points.
PAGE 215
215 A B Figure 717. Approximate division in the de sign space between the two patterns based on A) peak heat flux. Division is also shown for B) combustion length. Designs near the division have qualities of both patterns. Figure 718. Variation in flow streamlines and hydrogen contours in design space.
PAGE 216
216 Figure 719. Location of best tradeoff points in each pattern group in design space. A 150 100 50 0 50 100 250 200 150 100 50 0 zy Figure 720. Injector spacing for selected best tradeoff design point for A) Pattern 1 (case 19 shown with 50 injector elements) and B) Pattern 2 (case 18 shown with 54 injector elements).
PAGE 217
217 B 150 100 50 0 50 100 200 150 100 50 0 zy Figure 720. Continued. 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0 5 10 15 20 25 30 35 40 45 x (m)q (MW/m2) case 1 case 18 case 19 Figure 721. Predicted heat flux profiles for baseli ne case (case 1), best tradeoff from Pattern 1 (case 19), and best tradeoff from Pattern 2 (case 18).
PAGE 218
218 A 0 0.05 0.1 0.15 0 10 20 30 40 x (m)q (MW/m2) 2.75 105 6.03 105 B 0 0.05 0.1 0.15 0 10 20 30 40 x (m)q (MW/m2) 2.53 105 5.94 105 Figure 722. Heat flux profiles for different grid resolutions. The coarser grid was used to construct the surrogate mode l. A) Baseline (case 1). B) Case 19 of Pattern 1. 0.025 0.03 0.035 0.04 0.045 0.05 0.055 0.06 10 20 30 40 50 60 mdot H2 (kg/s)qmax(MW/m2)1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 0.025 0.03 0.035 0.04 0.045 0.05 0.055 0.06 75 80 85 90 95 100 105 mdot H2 (kg/s)combustion chamber length (mm)1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 2627 Figure 723. Peak heat flux and combustion leng th as a function of hydrogen mass flow rate. Peak heat flux decreases and combustion length increases with an increase in mass flow rate.
PAGE 219
219 CHAPTER 8 CONCLUSIONS CFDbased optimization can improve rocket engine component design by predicting the effect of design changes on performance and life indicators. The development of a comprehensive and efficient multiobjective optimization framework for CFD analysis is necessary in this effort. Additionally, design sp ace refinement techniques can help direct the design optimization by guiding the process in mean ingful directions. Design space refinement techniques were used to reduce the cost of CFDbased optimization th rough the practice of removing infeasible regions from the design spac e using lowfidelity analyses. Coupled with tools including merit func tions and Pareto fronts, design space refinement is a useful tool in improving the accuracy and applicab ility of a surrogate model. Design space refinement techniques were de veloped to account for optimization problems with multiple objectives. This includes the use of Pareto fronts to identify favorable design space regions in both objectives to allow for the removal of unnecessary regions. Additionally, a merit function was selected based on its ability to provide a good accuracy surrogate with a minimum number of points. A rank formulation was applied with data clustering to choose data points that were simultaneously (a ) in regions of high surrogate uncertainty, (b) have good values in each function, and (c) are spread evenly across the design space. The design space refinement techniques were used in conjunction with problems of in terest to liquid rocket engine design, looking at techniques to improve the mixing in flameholding devices, improve turbine efficiency, and improving performance and life ch aracteristics of a rocket engine combustion chamber. The results show that even simple anal yses can provide significant insight into rocket engine component flow dynamics that can s upplement the existing experimental studies.
PAGE 220
220 8.1 Radial Turbine Efficiency and Weight Optimization The compact radial turbine preliminary optim ization identified several factors that can improve its efficiency. The lowcost surrogate m odel analysis revealed that reducing the radius of the turbine while increasing blade size has a positive impact on efficiency without impacting the turbineÂ’s weight. The most efficient desi gns have higher rotationa l speeds and isentropic velocity ratios. A grid sensitivity analysis re vealed that the rotational speed and the velocity ratio accounted for 97% of the vari ability in the rotor weight, wh ile the turbine efficiency was sensitive to all variables included in the analysis. The applicability of DSR techniques was succes sfully demonstrated in the preliminary optimization of a radial turbine optimization. 1. In the case where the feasible design space was unknown, response surfaces proved invaluable in determining design constraints. 2. After the feasible design space was determined, information from a Pareto front was used to refine the design space to improve su rrogate accuracy without compromising the performance of either objective. For the radial turbine optimization, the infeas ible points were not thrown away. Instead, they served a useful purpose by pr oviding the location of the bounds of the feasible design space. For a sixvariable design space, surrogate models were necessary in extracting these boundsÂ—a task which would have been very difficult, othe rwise. In addition, the study addressed the problem of refining the design space in the presence of multiple objectives. To prevent missing the region of interest in one objective or the other when the design space was refined, Pareto fronts were used to simultaneously locate the regions of interest in both objectives. Finally, the radial turbine case was analyzed again using merit functions to aid the analysis. It was discovered that the use of merit func tions could improve the efficiency of the optimization process by improving point selectio n during design space refinement. Using merit
PAGE 221
221 functions also means that fewer points are required during each st age in the analysis. Because the radial turbine optimization was low cost, comparisons could be made of the accuracy between the selection of a la rge number of points using standard DOE techniques of Latin Hypercube Sampling and factorial designs and the selection of a small number of points using merit functions. By using the merit function that maximized the probability of improvement of the function values, the number of points in the final optimization cycle could be reduced by 94% while maintaining a Pareto front accuracy w ithin 10%. This substantial reduction in the number of points required meant that the use of merit functions would be a suitable method of point selection in the optimiza tion cycles of a computationall y expensive problem, such as injector flow. 8.2 Bluff Body Mixing Dynamics A study of the sensitivity of flameholder mi xing to the bluff body geometry illustrated the multiple roles of surrogatebased optimization. 1. In addition to facilitating the optimization, surrogates can be used to warn against an inadequate CFD model. In this case, th e large difference between an RSA and RBNN approximation showed that a problem existed. 2. Complete improvement cannot always be gained using an alternative surrogate. A neural network provided a better approximation in so me regions, while an RSA performed better in other regions. 3. It may be necessary to improve the accuracy of the CFD model or refine the design space to obtain a better result. In this case, problem s in the CFD model must be resolved before the analysis can continue. 4. Exploration of the design space may be necessa ry to determine the deficiency in the CFD model. Additional data points were added in su spect regions to reveal the true nature of the CFD model response. These conclusions are particular ly important, as similar issues can arise in other CFDbased design optimizations with complex flow situatio ns. Without the use of surrogate models and
PAGE 222
222 DSR, the problem of insufficient grid refine ment in the CFD model might not have been discovered. Understanding the mixing dynamics and the effects of the geometry on mixing is important in the design of combustion devices such as injector flow. In this case, the CFDbased surrogate model was able to reveal important aspects on the sensitiv ity of the mixing to the bluff body geometry. It was found that the mixing coul d be increased by changing the shape of the trailing edge of the bluff body. This increased mixing often came at the expense of increasing drag on the body, however, locations of extreme cases revealed designs that gave favorable mixing characteristics and low drag. 8.3 SingleElement Injector Flow Modeling The injector flow modeling was conducted as a preliminary exercise to the combustion chamber flow optimization. Flow was analyzed within the injector itself to determine combustion chamber inlet conditions. Model verification was provided by developing a computational model of an experimental combus tion chamber. Finally, a grid sensitivity analysis was conducted. It was discovered that the combustion le ngth was more sensitive to the grid resolution than the wall heat flux. Grid i ndependency is not reached in the present study. The wall treatment of the turbulence model reduces the sensitivity of wall heat transfer computations. A grid resolution was selected fo r the injector flow op timization that provided good tradeoff between heat flux profile accuracy and computational run time. 8.4 MultiElement Injector Flow Modeling The CFD model sought to approx imate the effects near a sing le injector element in the outer injector row of multielement combustor flow. Geometric design variables were selected to improve the performance and life of a ro cket engine combustion chamber. The spacing
PAGE 223
223 between injector elements and the distance to th e wall were adjusted to determine the effect on the peak heat flux profile and combustion length. Determination of the feasible design space reveal ed interesting flow characteristics of the initial data set. It was found that each point in the design space resulted in injector flow that exhibited one of two characteristics: 1. The outer shear layer is directed towards the combustion chamber wall, resulting in a sharp peak in heat flux near the injector face. This occurs due to a large distance to the wall as compared to the interelement distance. A lthough these designs, in general, had lower mass flow rates per injector, overa ll, this scenario resulted in very high peak heat fluxes. 2. The outer shear layer is directed towards th e neighboring element within the same row, resulting in a low heat flux near the inject or face that gradually increases to a peak downstream. This occurs due to an injectort owall distance that is small compared to the interelement distance. These designs had highe r mass flow rates per injector, but the heat transfer was spread more evenly across the combustion chamber wall. The behavior of the different flow patterns a nd trends dictated that separate analyses be conducted based on the characteristic flow pattern. The accuracy of the initial surrogate was low, indicating the need for a design space refinement. In the radial turbine analysis, mer it functions were shown to have the ability to minimize the number of points needed to improve the surrogate model without significantly compromising the future accuracy of the surroga te. For this reason the merit function that maximizes the probability of improvement of the surrogate model was used to select additional points for each region of the design space. The re finement process consisted of a change in the design variable bounds as well as an additi on of points based on the multiobjective merit functionbased point selection procedure. The DSR was applied to each of the two regions of the design space based on the predicted flow characterist ics. All of the new points selected using the merit functionbased DSR dominated all of the original design points. The multiobjective
PAGE 224
224 surrogatebased DSR using merit functions for po int selection was thus successful in providing optimal design using a small number of design points. 8.5 Future Work This study provides a starting point for CFDba sed design of rocket engine components. In particular, this study provide s a new method of attacking comp lex rocket engine component design. Future studies may include 1. Full 3D simulations of the selected optimum radial turbine design and a comparison with the 1D Meanline code results to determine if the efficiency gains are comparable over the baseline case. 2. A threedimensional blade shape optimization of the selected optim um radial turbine design. 3. Adding additional design points near the selected designs of the multielement injector analysis. Doing so would provide a more accu rate representation of the design space near the selected designs. 4. Improving the accuracy of the multielement in jector model by including the second row of injector elements in the CFD model. In th is case, the best designs as selected in the current study can provide a star ting point for the additional an alysis. The effect of the injector interactions can be quantified, and the applicabil ity of the slip boundary and symmetric computational domain can be determined. 5. Systematic experimental invest igation of the effects of the wall distance and/or the interelement injector spacing on peak heat flux and combustion efficiency for a multielement injector. Experimental studies on the selected best, or similar, multielement injector designs and a comparison with the baseline case would provide expe rimental confirmation of the results of the simple study.
PAGE 225
225 REFERENCE LIST 1. Jue, F. and Kuck, F., Â“Space Shuttle Main Engine (SSME) Options for the Future Shuttle,Â” NASA CP3758, 2002. 2. Sarker, B., Â“Operations Analysis of Sp ace Shuttle System,Â” Research Reports: 2001 NASA/ASEE Summer Faculty Fellowshi p Program, NASA/CR2002211840, XLVI1 XLVI, 20020701; July 2002. 3. Enyinda, C., Â“Second Generation Reusable Launch Vehicle Deve lopment and Global Competitiveness of US Space Transportati on Industry: Critical Success Factors Assessment,Â” Research Reports: 2001 NASA/AS EE Summer Faculty Fellowship Program, NASA/CR2002211840, July 2002. 4. Jones, D. R., Schonlau, M. and Welch, W. J ., Â“Efficient Global Optimization of Expensive BlackBox Functions,Â” Journal of Global Optimization Vol. 13, No. 4, 1998, pp. 455492. 5. Tucker, P., Klem, M., Smith, T., Farhangi, S ., Fisher, S., and Santoro, R., Â“Design of Efficient GO2/GH2 injectors Â– a NASA, Indus try and University Cooperative Effort,Â” AIAA Paper 19973350, 1997. 6. Griffin, M. D. and French, J. R., Space Vehicle Design American Institut e of Aeronautics and Astronautics, Inc., Washington, DC, 1991. 7. Vigier G., Dufour A. and Peypoudat V., Â“Pre ssureFed Stages for Low Cost Expandable Launchers,Â” AIAA Paper 20034816, 2003. 8. Quinn, J. E. Â“Cost TradeOff for a Pressure Fed Liquid Rocket to LEO,Â” AIAA Paper 19992622, 1999. 9. Cotta, R. B., Eckmann, J. B., and Matuszak, L. W., Â“Upper Stage Options for Reusable Launch Vehicles,Â” AIAA Paper 19963015, 1996. 10. Wagner, W. R., Â“Future Higher Perfor mance O2/H2 Engine Combustion Cycle Alternatives,Â” AIAA Paper 19760709, 1976. 11. Erickson, C., Pinkowski, S., Anderson, M., Li nner, H., Obermaier, G., and Taubenberger, G, Â“High Performance Gas Generator Cycle Concept for Storable UpperStage Applications,Â” AIAA Paper 19992605, 1999. 12. Crocker, A. M. and Peery, S. D., Â“System Sensitivity Studies of a LOX/Methane Expander Cycle Rocket Engine,Â” AIAA Paper 19983674, 1998. 13. Krach, A. E. and Sutton, A. M, Â“Another Look at the Practical and Theoretical Limits of an Expander Cycle, LOX/H2 EngineÂ” AIAA Paper 19992473, 1999. 14. Parsley, R. C. and Crocker, A. M., Â“The RL200 Incorporating a Balanced Approach for Reusable Propulsion Safety,Â” AIAA Paper 20003287, 2000.
PAGE 226
226 15. Immich H. and Caporicci, M., Â“FESTIP Technology Developments in Liquid Rocket Propulsion for Rreusable Launch Vehicles ,Â” AIAA Paper 19963113, 1996. 16. Paster, R. D. and Stohler, S. L. Â“ SSME evolution,Â” AIAA Paper 19892618, 1989. 17. Marshall, W. M., Pal, S., Woodward, R. D., and Santoro, R. J., Â“Benchmark Wall Heat Flux Data for a GO2/GH2 Single Elemen t Combustor,Â” AIAA Paper 20053572, 2005. 18. Pavli, A. L., Â“Design and Evaluation of High Performance Rocket Engine Injectors for use with Hydrocarbon Fuels,Â” NASA TM69319, 1979. 19. Knuth, W. H., Gramer, D. J., and St. Clair, C. P., Â“Premixing Rocket Engine Injectors for High Pressure Gaseous Oxygen and Hydrogen,Â” AIAA Paper 19992491, 1999. 20. Rupe, J., Â“An Experimental Correlation of th e Nonreactive Properties of Injection Schemes and Combustion Effects in a LiquidPropell ant Rocket Engine,Â” NASA TR 32255, 1965. 21. Zinner, W., Haeseller, D., Mding, C., Rubins kji, V., Gorochov, V., Hrisanfov, S., and Nikulin, G., Â“Development of Advanced Technologies for Future Cryogenic Thrust Chamber Applications,Â” AIAA Paper 19973312, 1997. 22. Conley, A., Vaidyanathan, A., and Segal, C ., Â“Heat Fluxes Measurements in a GH2/GO2 SingleElement, Shear Injector,Â” Journal of Spacecraft and Rockets, Vol. 44, No. 3, 2007, pp. 633Â–639. 23. Osborn, G. H., Gordon, R., Coplen, H. L., and James, G., Â“LiquidHydrogen Rocket Engine Development at Aerojet, 1944 Â– 1950,Â” NASA N7733064, 1977. 24. Peterson, D. A., Winter, J. M., Pavli, A. J ., Â“Design, Development, and Testing of a 1000 Pound (4450 N) Thrust FloxPropane Ablativ e Rocket Engine, NASA TMX1873, 1969. 25. Injector orifice study, Apollo service propul sion system Summary report, 7 Apr. 1967 Â– 31 Jul. 1968 NASACR99684, 1968. 26. Gill, G. S., Â“Liquid Rocket Engine Injectors,Â” NASA/SP8089, 1976. 27. Liang, P., Fisher, S., Chang, Y., Â“Com prehensive Modeling of a Liquid Rocket Combustion Chamber,Â” Journal of Propulsion Vol. 2, No. 2, 1986, pp. 97Â–104. 28. Sindir, M. and Lynch, D., Â“Overview of the StateofthePractice of Computational Fluid Dynamics in Advanced Propulsion Syst em Design,Â” AIAA Paper 19972124, 1997. 29. Shyy, W., Tucker P. K., and Vaidyanathan R., Â“Response Surface and Neural Network Techniques for Rocket Engine Injector Optimization,Â” Journal of Propulsion and Power Vol. 18, 2001, pp. 391401. 30. Vaidyanathan, R., Tucker, P. K., Papila N., and Shyy, W., Â“CFDBased Design Optimization for a Single Element Rocket Injector,Â” AIAA Paper 20030296, 2003.
PAGE 227
227 31. Brunner, J. J., Â“Fluid Dynamics Comput er Programs for NERVA Turbopump,Â” NASA CR132222, 1972. 32. Chyu, M. K., Â“A Performance Loss Evaluation of SSME HPFTP Turbine,Â” NASA 8921735, 1988. 33. Dorney, D. J., Sondak, D. L. and Marcu, B ., Â“Application of a Real Fluid Turbomachinery Analysis to Rocket Turbopump Ge ometries,Â” AIAA Paper 20051007, 2005. 34. Snodgrass, R., Dorney, D. J., Sandgren, E., Me rz, L. F., Â“MultiObjective Shape Optimal Design of a Supersonic Turbine,Â” AIAA 20004860, 2000. 35. Hedges, D. and Welsh, B., Â“The Boei ng 777300/PW4098 Flying TestBed Program,Â” AIAA and SAE 1998 World Aviation Conferen ce, Anaheim, CA, Sept. 2830, AIAA19985550, 1998. 36. Greer, W., Hein, J., Hinson, M., Stour, W ., Â“Designing the Learjet Model 45 Greer,Â” AIAA Paper 19953951, 1995. 37. Stojanowski, M., Â“Business Jet Type Conf iguration Testing at ETW,Â” AIAA Paper 2005460, 2005. 38. He B., Ghattas O., and Antaki J. F., Â“Computational Strategies for Shape Optimization of TimeDependent NavierStokes Flow s,Â” Technical Report CMUCML97102, Computational Mechanics Lab, Department of Civil and Environmental Engineering, Carnegie Mellon University, 1997. 39. Rodriguez D. L. Â“A Multidisciplinary Optimization Method for Designing Boundary Layer Ingesting Inlets,Â” AIAA Paper 20025665, 2002. 40. Jones, W. and SamarehAbolhassani, J., Â“A Gr id Generation System for MultiDisciplinary Design Optimization,Â” NASA CP3291, 1995. 41. Jameson, A., Sriram, Martinelli, L., and Ha imes, B., Â“Aerodynamic Shape Optimization of Complete Aircraft Configurations usi ng Unstructured Grids, AIAA Paper 20040533, 2004. 42. Hussaini M. M. and Korte, J. J., Â“Investig ation of LowReynoldsNumber Rocket Nozzle Design Using PNSBased Optimizati on Procedure,Â” NASA TM110295, 1996. 43. Lesieutre, D., Dillenius M., and Lesieutre, T., Â“Multidisciplinary Design Optimization of Missile Configurations and Fin Planform s for Improved Performance,Â” AIAA Paper 19984890, 1998. 44. Nadarajah S., Jameson A., and Alonso J. J. Â“Sonic Boom Reduction Using an Adjoint Method for WingBody Configurations in Supersonic Flow,Â” AIAA Paper 20025547, 2002.
PAGE 228
228 45. Kim, S., Leoviriyakit, K., and Jameson, A., Â“Aerodynamic Shape and Planform Optimization of Wings Using a Viscous Reduc ed Adjoint Gradient Formula,Â” Second M.I.T. Conference on Computational Fluid a nd Solid Mechanics at M.I.T, Cambridge, MA, 2003. 46. Myers, R. H., and Montgomery, D. C., Response Surface Methodology New York: John Wiley & Sons, Inc., 2002, pp. 308, 393Â–399, 17Â–20. 47. Krige, D., Â“A Statistical Approach to So me Basic Mine Evaluation Problems on the Witwatersrand,Â” Journal of the Chemical, Metallurg ical and Mining Society of South Africa Vol. 52, 1951, pp. 119Â–139. 48. Oliver M. and Webster, R ., Â“Kriging: A Method of Inte rpolation for Geographical Information System,Â” International Journal of Ge ographical Information Systems Vol. 4, No. 3, 1990, pp. 313Â–332. 49. Journel, A. and Huijbregts, Ch., Mining Geostatistics New York, Academic Press, 1978. 50. Mitchell, T., Machine Learning McGrawHill, 1997, pp. 81Â–127. 51. Burgee, S., Giunta, A. A., Narducci, R., Watson, L. T., Grossman, B. and Haftka, R. T., Â“A Coarse Grained Parallel VariableComplexity Multidisciplinary Optimization Paradigm,Â” The International Journal of Supercomput er Applications and High Performance Computing Vol. 10, No. 4, 1996, pp. 269Â–299. 52. Queipo N., Haftka R. T., Shyy W., Goel T ., and Vaidyanathan R. Â“SurrogateBased Analysis and Optimization,Â” Progress in Aerospace Sciences Vol. 41, 2005, pp. 1Â–28. 53. Kontoravdi, C., Asprey, S. P., Pistikopoulos, E. N., and Mantalaris, A. Â“Application of Global Sensitivity Analysis to Determine Goals for Design of Experiments Â– An Example Study on AntibodyProducing Cell Cultures,Â” Biotechnology Progress Vol. 21, No. 4, 2005, pp. 1128Â–1135. 54. Marsden, A., Wang, M., Dennis, J., Moin, P ., Â“Optimal Aeroacoustic Shape Design Using the Surrogate Management Framework,Â” Optimization and Engineering Vol. 5, 2004, pp. 235Â–262. 55. Jansson, T., Nilsson, L., Redhe, M., Â“Using Surrogate Models and Response Surfaces in Structural Optimization Â– With Application to Crashworthiness Design and Sheet Metal Forming,Â” Structural Multidisciplinary Optimization Vol. 25, 2003, pp. 129Â–140. 56. Joyner, C. and Sabatella, J., Â“Launch Ve hicle Propulsion Optimization Using Response Surface Methodology,Â” AIAA Paper 19902433, 1990. 57. Shyy, W., Tucker P. K., and Vaidyanathan R., Â“Response Surface and Neural Network Techniques for Rocket Engine Injector Optimization,Â” Journal of Propulsion and Power Vol. 18, 2001, pp. 391Â–401.
PAGE 229
229 58. Simpson T. W. and Mistree, F., Â“Kriging M odels for Global Approximation in SimulationBased Multidisciplinary Design Optimization,Â” AIAA Journal Vol. 39, No. 12, 2001. 59. Dornberger, R., Bche, D., and Stoll, P., Â“Multidisciplinary Optimization in Turbomachinery Design,Â” European Congress on Computational Me thods in Applied Sciences and Engineering, Barcelona, September, 2000. 60. Iman, R.L., Davenport, J. M., and Zeigler, D. K., Latin Hypercube Sampling (A Program UserÂ’s Guide) Technical Report SAND 791473. Albuque rque, NM: Sandia Laboratories, 1980. 61. Simpson T. W. and Mistree, F., Â“Kriging M odels for Global Approximation in SimulationBased Multidisciplinary Design Optimization,Â” AIAA Journal Vol. 39, No. 12, 2001. 62. Rodriguez D. L. Â“A Multidisciplinary Optimization Method for Designing Boundary Layer Ingesting Inlets,Â” AIAA Paper 20025665, 2002. 63. Vaidyanathan, R., Tucker, P. K., Papila N., and Shyy, W., Â“CFDBased Design Optimization for a Single Element Rocket Injector,Â” Journal of Propulsion and Power Vol. 20, No. 4, 2004, pp. 705Â–717. 64. Han S. Y., Maeng, J. S., and Yoo D. H., Â“Shape Optimization of Cutoff in a Multiblade Fan/Scroll System Using Response Surface Methodology,Â” Numerical Heat Transfer Part B, Vol. 43, 2003, pp. 87Â–98. 65. Keane, A. J., Â“Wing Optimization Using De sign of Experiment, Response Surface, and Data Fusion Methods,Â” Journal of Aircraft Vol. 40, No. 4, 2003. 66. Alexandrov, N., Lewis, R., Gumbert, C., Gr een, L., and Newman, P., Â“Optimization with variable fidelity models applied to wing design,Â” AIAA Paper 20000841, 2000. 67. Knill, D. L., Giunta, A. A., Baker, C. A., Gr ossman, B., Mason, W. H., Haftka, R. T., and Watson, L. T., Â“Response Surface Methods Comb ining Linear and Euler Aerodynamics for Supersonic Transport Design,Â” Journal of Aircraft Vol. 36, No. 1, 1999, pp. 75Â–86. 68. Vitali, R., Haftka, R., Sankar, B., Â“Multifidel ity Design of Stiffened Composite Panel with a Crack,Â” Journal of Structural and Mu ltidisciplinary Optimizaiton Vol. 23, 2002, pp. 347Â–356. 69. Venkataraman, S. and Haftka, R. T., Â“Desi gn of Shell Structures for Buckling using Correction Response Surface Approxima tions,Â” AIAA Paper 984855, 1998. 70. Balabanov V. O., Giunta A. A., Golovidov O., Grossman B., Mason W., Watson L. T., and Haftka R. T., Â“Reasonable Design Space Approach to Response Surface Approximation,Â” Journal of Aircraft Vol. 36, No. 1, 1999, pp. 308Â–315. 71. Haftka, R.T., Â“Combining Global and Local Approximations,Â” (Technical Note), AIAA Journal Vol. 29, No. 9, 1991, pp. 1523Â–1525.
PAGE 230
230 72. Chang, K. J., Haftka, R. T., Giles, G. L., and Kao, P. J., Â“SensitivityBased Scaling for Approximating Structural Response,Â” Journal of Aircraft Vol. 30, No. 2, 1993, pp. 283Â– 287. 73. Roux W. J., Stander N., and Haftka R. T. Â“Response Surface Approximations for Structural Optimization,Â” International Journal for Nume rical Methods in Engineering Vol. 42, 1998, pp. 517Â–534. 74. Madsen, J. I., Shyy, W., and Haftka, R. T., Â“Response Surface Techniques for Diffuser Shape Optimization,Â” AIAA Journal Vol. 38, 2000, pp. 15121518. 75. Mack, Y., Goel, Y., Shyy, W., Haftka, R., Qu eipo, N., Â“Multiple Surrogates for the Shape Optimization of Bluff BodyFacilitat ed Mixing,Â” AIAA Paper 20050333, 2005. 76. Jeon, Y., Jeon, K., Lee, J., Byun, Y., Â“Pre mixture Optimization of the SuperDetonative Ram Accelerator,Â” AIAA Paper 20004874, 2000. 77. Papila, M., Papila, N., Shyy, W., Haftka, R ., and FitzCoy, N., Â“ErrorBased Design Space Windowing,Â” AIAA Paper 20020539, 2002. 78. Papila, N., Shyy, W., Griffin, L., and Dorney D. J., Â“Shape Optimization of Supersonic Turbines Using Global Approximation Methods,Â” Journal of Propulsion and Power Vol. 18, 2002, pp. 509Â–518. 79. Shyy, W., Papila, N., Vaidyana than, R., and Tucker, K., Â“G lobal Design Optimization for Aerodynamics and Rocket Propulsion Components,Â” Progress in Aerospace Sciences Vol. 37, 2001, pp. 59Â–118. 80. Simpson T. W. and Mistree, F., Â“Kriging M odels for Global Approximation in SimulationBased Multidisciplinary Design Optimization,Â” AIAA Journal Vol. 39, No. 12, 2001, pp. 2233Â–2241. 81. Schreur, B., Â“Presenting the Shuttle Main Engine Software,Â” NASA Center for AeroSpace Information (CASI), 19980101; 1998. 82. Papila, N., Shyy, W., Griffin, L., Huber, F. and Tran, K., Â“Preliminary Design Optimization for a Supersonic Turbine for Rocket Propulsion,Â” AIAA Paper 20003242, 2000 83. Myers, R. H., and Montgomery, D. C., Response Surface Methodology New York: John Wiley & Sons, Inc., 2002, pp. 393 Â– 395, 308. 84. Tang, B., Â“Orthogonal ArrayBased Latin Hypercubes,Â” Journal of the American Statistical Association Vol. 88, No. 424, 1993, pp. 1392Â–1397. 85. RaisRohani, M. and Singh, M., Â“Compari son of Global and Local Response Surface Techniques in ReliabilityBased Opti mization of Composite Structures,Â” Journal of Structural and Multidisciplinary Optimization Vol. 26, 2004, pp. 333Â–345.
PAGE 231
231 86. Sevant, N., Bloor, M., and Wilson, M., Â“A erodynamic Design of a Flying Wing Using Response Surface Methodology,Â” Journal of Aircraft Vol. 37, No. 4, 2000, pp. 562Â–569. 87. Sturdza, P., Manning, V., Kroo, I., and Tr acy, R., Â“Boundary Layer Calculations for Preliminary Design of Wings in Supe rsonic Flow,Â” AIAA Paper 19993104, 1999. 88. Kurtaran, H., Eskandarian, A., Marzougui, D ., and Bedewi, N., Â“Crashworthiness Design Optimization Using Successive Response Surface Approximations,Â” Computational Mechanics Vol. 29, 2002, pp. 409421. 89. Hosder, S., Watson, L., Grossman, B., Mason, W., and Kim, H., Â“Polynomial Response Surface Approximations for the Multidiscip linary Design Optimization of a High Speed Civil Transport,Â” Optimization and Engineering Vol. 2, 2001, pp. 431452. 90. Matheron, G., Â“Principles of Geostatistics,Â” Economic Geology Vol. 58, 1963, pp. 1246Â– 1266. 91. Koch, P., Simpson, T., Allen, J., and Mist ree, F., Â“Statistical Approximations for Multidisciplinary Design Optimization: The Problem of Size,Â” Journal of Aircraft Vol. 36, No. 1, 1999, pp. 275Â–286. 92. Mahadevan, S., Shi P., Salter J., Raghotha machar P., Â“Advanced Reliability Methods,Â” AIAA Paper 19991603, 1999. 93. Forsberg, J. and Nilsson, L., Â“On Polynomi al Response Surfaces and Kriging for Use in Structural Optimization of Crashworthiness,Â” Journal of Structural and Multidisciplinary Optimization Vol. 29, 2005, pp. 232Â–243, 2005 94. Jin, R., Chen, W., and Simpson, T., Â“Compa rative Studies of Metamodelling Techniques Under Multiple Modeling Criteria,Â” Journal of Structural and Multidisciplinary Vol. 23, 2001, pp. 113. 95. Rijpkema, J., Etman, L., and Schoofs, A., Â“Use of Design Sensitivity Information in Response Surface and Kriging Metamodels,Â” Optimization and Engineering Vol. 2, 2001, pp. 469Â–484, 2001. 96. Chung, H. and Alonso, J., Â“Comparison of A pproximation Models with Merit Functions for Design Optimization,Â” AIAA Paper 20004754, 2000. 97. Simpson, T., Mauery, T., Korte, J., and Mist ree, F., Â“Comparison of Response Surface and Kriging Models for Multid isciplinary Design Optimization,Â” AIAA Paper 19984755, 1998. 98. Kanazaki, M., Tanaka, K., Jeong, S., and Ya mamoto, K., Â”MultiObjective Aerodynamic Exploration of ElementsÂ’ Setting for Hi ghLift Airfoil Using Kriging Model,Â” Journal of Aircraft Vol. 44, No. 3, 2007, pp. 858Â–864.
PAGE 232
232 99. Jouhaud, J. Â–C., Sagaut, P., Montagnac, M., and Laurenceau, J., Â“A SurrogateModel Based Multidisciplinary Shape Optimization Method with Applicati on to a 2D Subsonic Airfoil,Â” Computers and Fluids Vol. 36, No. 3, 2007, pp. 520Â–529. 100. Sacks, J., Schiller, S., and Welch, W. Â“Designs for Computer Experiments,Â” Technometrics Vol. 31, 1989, pp. 41Â–47. 101. Lophaven, S. N., Nielsen, H. B., a nd Sndergaard, Â“DACE: A MATLAB Kriging Toolbox, Version 2.0Â” IMMTR200212, T echnical University of Denmark, 2002. 102. Mahajan, R. L., Fichera, B. M., and Hors t T. W., Â“Mechanically Aspirated Radiation Shields: A CFD and Neural Network Design Analysis,Â” International Jour nal of Heat and Mass Transfer Vol. 48, 2005, pp. 2856Â–2867. 103. Charalambous, C., Charitou, A., Kaourou, F., Â“C omparative Analysis of Artificial Neural Network Models: Application in Bankruptcy Prediction,Â” Annals of Operations Research Vol. 99, 2001, pp. 403Â–425. 104. Hsken, M., Jin, Y., Sendhoff, B., Â“Structu re Optimization of Neural Networks for Evolutionary Design Optimization,Â” Soft Computing Vol. 9, 2005, pp. 21Â–28. 105. Shi, Q. and Hagiwara, I., Â“Structural Optim ization Based on Hologr aphic Neural Network and its Extrapolation,Â” AIAA Paper 19984975, 1998. 106. Pidaparti, R. and Palakal, M., Â“Fatique Crack Growth Predictions in Aging Aircraft Panels using Optimization Neural Network,Â” AIAA Journal Vol. 36, No. 7, 1998, pp. 1300Â–1304. 107. Chan, Y. and Zhu, Z., Â“Neural Network Modeling of Aerodynamics and Applications to Store Separation,Â” AIAA Paper 19990685, 1999. 108. Brigham, J. C., and Aquino, W., Â“Surrogate Model Accelerated Random Search Algorithm for Global Optimization with Applications to Inverse Material Identification,Â” Computer Methods in Applied Mechanics and Engineering 2007, in press, doi:10.1016/j.cma.2007.05.013. 109. Gmez, M. Â–D. and Tarantola, Sl, Â“Global Sensitivity Analysis, GIS and MultiCriteria Evaluation for a Sustainable Planning of a Hazardous Waste Disposal Site in Spain,Â” International Journal of Ge ographical Information Science Vol. 20, No. 4, 2006, pp. 449Â– 466. 110. Carrero, E., Queipo, N. V., Pintos, S., Zerpa, L. E., Â“Global Sensitivity Analysis of AlkaliSurfactantPolymer Enhanced Oil Recover Processes,Â” Journal of Petroleum Science and Engineering, Vol. 58, 2007, pp. 30Â–42. 111. 1 Sobol, I. M., Tarantola, S., Gatelli, D., Kucherenko, S. S ., and Mauntz, W., Â“Estimating the Approximate Error when Fixing Unessentia l Factors in Global Sensitivity Analysis,Â” Reliability Engineering and System Safety Vol. 92, 2007, pp. 957Â–960.
PAGE 233
233 112. Sobol, I. M., 1993, "Sensitivity Analysis for Nonlinear Mathematical Models," Mathematical Modeling & Computational Experiment Vol. 1, No. 4, 1993, pp. 407Â–414. 113. Deb K., Agrawal S., Pratap A., and Meyariva n T. Â“A Fast and Elitist MultiObjective Genetic Algorithm for MultiObjective Optimi zation: NSGAII,Â” Nature VI Conference, Paris, pp 849Â–858, 2000. 114. Goel T., Vaidyanathan R., Haftka R. T., Queipo N. V., Shyy W., and Tucker P. K., Â“Response Surface Approximation of Pareto Optimal Front in MultiObjective Optimization,Â” AIAA Paper 20044501, 2004. 115. Huang, D., Allen, T. T., Notz, W. I., and Zeng, N., Â“Global Optimization of Stochastic BlackBox Systems via Sequential Kriging MetaModels,Â” Journal of Global Optimization Vol. 34, No. 3, 2006. 116. Toropov, V., van Keulen, F., Markine, V., a nd de Doer, H., Â“Refinements in the MultiPoint Approximation Method to Reduce the Eff ects of Noisy Structur al Responses,Â” AIAA Paper 19964087, 1996. 117. Goel, T., Dorney, D., Haftka, R., and Shyy, W., Â“Improving the Hydrodynamic Performance of Diffuser Vanes via Sh ape Optimization,Â” AIAA Paper 20075551, 2007. 118. Jones, D. R., Schonlau, M. and Welch, W. J ., Â“Efficient Global Optimization of Expensive BlackBox Functions,Â” Journal of Global Optimization Vol. 13, No. 4, 1998, pp. 455Â–492. 119. Jeong, S. and Obayashi, S., Â“Efficient Gl obal Optimization (EGO) for MultiObjective Problem and Data Mining,Â” Evolutionary Computation Vol. 3, 2005, pp. 2138Â–2145. 120. Knowles, J. D. Â“ParEGO: A Hybrid Al gorithm with OnLine Landscape Approximation for Expensive Multiobjective Optimization Problems,Â” IEEE Transactions on Evolutionary Computation Vol 10., No. 1, 2006, pp. 50Â–66. 121. FarhangMehr, A. and Azarm, S., Â“Bayes ian MetaModelling of Engineering Design Simulations: A Sequential Approach with Ad aptation to Irregulariti es in the Response Behavior,Â” International Journal for Nume rical Methods in Engineering Vol. 62, 2005, pp. 2104Â–2126. 122. Sasena, M. J., Papalambros, P., and G oovaerts, P., Â“Exploration of Metamodeling Sampling Criteria for Constrained Global Optim ization,Â” Engineering Optimization, Vol. 34, No. 3, pp. 263278, 2002. 123. Jones, D. Â“A Taxonomy of Global Optimiza tion Methods Based on Response Surfaces,Â” Journal of Global Optimization Vol 21, 2001, pp. 345Â–383. 124. Friedman, J. H., Â“Multivariate Adaptive Regression SplinesÂ” (with discussion). Annals of Statistics Vol. 19, No. 1, 1991, pp. 1Â–67.
PAGE 234
234 125. BosqueSendra, J., Pescarolo, S., Cuadro sRodrigues, L., GarciaCampana, A., and AlmansaLopez, E., Â“Optimizing Analytical Methods Using Sequential Response Surface Methodology. Application to the Pararosa niline Determination of Formaldehyde,Â” Journal of Analytical Chemistry Vol. 369, 2001, pp. 715Â–718. 126. Immich H. and Caporicci, M., Â“FESTIP Technology Developments in Liquid Rocket Propulsion for Reusable Launch Ve hicles,Â” AIAA Paper 19963113, 1996. 127. Huber, F., Â“Turbine Aerodynamic Design Tool Development,Â” pres ented at the Space Transportation Fluids Workshop, Marshall Space Flight Center, AL, April, 2001. 128. Sondak, D. L. and Dorney, D. J., Â“General Equation Set Solver for Compressible and Incompressible Turbomachinery Flows, '' AIAA Paper 20034420, July 2003. 129. Venkateswaran, S. and Merkle, C. L., Â“Analy sis of Preconditioning Methods for the Euler and NavierStokes Equations,Â” Von Karman Institute Lecture Series March 812, 1999. 130. Baldwin, B. S., and Lomax, H., Â“Thin La yer Approximation and Algebraic Model for Separated Turbulent Flow,Â” AIAA Paper 1+78257, Jan. 1978 131. NIST Chemistry WebBook [online database], URL: http://webbook.nist.gov/chemistry/fluid [cited 19 July 2006] 132. Dorney, D. J., Sondak, D. L. and Marcu, B ., Â“Application of a Real Fluid Turbomachinery Analysis to Rocket Turbopump Geom etries,Â” AIAA Paper 051007, Jan. 2005. 133. Pavli, A. L., Â“Design and Evaluation of High Performance Rocket Engine Injectors for use with Hydrocarbon Fuels,Â” NASA TM69319, 1979. 134. Calhoon, D., Ito, J., and Kors, D., Â“Investig ation of Gaseous Propellant Combustion and Associated Injector/Chamber Desi gn Guidelines,Â” NASA TM121234, 1973. 135. Hutt, J. J. and Cramer, J. M., Â“Advanced Rocket Injector Development at Marshall Space Flight Center,Â” AIAA Paper 964266, 1996. 136. Rupe, J. and Jaivin, G., Â“The Effects of In jection Mass Flux Distri butions and Resonant Combustion on Local Heat Transfer in a LiquidPropellant Rocket Engine,Â” NASA Technical Report 32648, 1964. 137. Farhangi, S., Yu, T., Rojas, L., and Sprouse, K., Â“GasGas Injector Technology for Full Flow Stage Combustion Cycle Appl ication,Â” AIAA Paper 19992757, 1999. 138. Branam, R. and Mayer, W., Â“Length Scales in Cryogenic Injecti on at Supercritical Pressure,Â” Experiments in Fluids Vol. 33, 2002, pp. 422Â–428. 139. Quentmeyer, R. J. and Roncace, R. A., Â“HotG asSide Heat Transfer Characteristics of Subscale, PlugNozzle Rocket Calorimeter Chamber,Â” NASATP3380, July 1993.
PAGE 235
235 140. Wanhainen, J. P., Feiler, C. E., and Morgan, C. J., Â“Effect of Chamber Pressure, Flow Per Element, and Contraction Ratio on Acous ticMode Instability in HydrogenOxygen Rockets,Â” NASA TN D4733, NASA, Washington, D.C., August 1968. 141. Mayer, W., Telaar, J., Branam, R., Schneider G., and Hussong, J., Â“Raman Measurements of Cryogenic Injection at Supercritical Pressure,Â” Heat and Mass Transfer Vol. 39, pp. 709 Â–719. 142. Moon, L. F., Â“Pressure and Velocity in a Developing Coaxial Jet,Â” AIAA Journal Vol. 14, No. 1, pp. 43Â–49. 143. Preclik, D., Wiedmann, D., O echslein, W., and Kretschm er, J., Â“Cryogenic Rocket Calorimeter Chamber Experiments and Heat Transfer Simulations,Â” AIAA Paper 19983440, 1998. 144. Conley, C.A., Â“HighPressure GO2/GH2 Combustion Chamber Dynamics,Â” MS thesis, University of Florida, 2006. 145. Reed, B. D., Â“Small Hydrogen/Oxygen Rock et Flowfield Behavi or from Heat Flux Measurements,Â” AIAA Paper 19932162, 1993. 146. Foust, M. J., Deshpande, M., Pal, S., Ni, T., Merkle, C. L., and Santoro, R. J., Â“Experimental and Analytical Characteri zation of a Shear Coaxial Combusting GO2/GH2 Flowfield,Â” AIAA Paper 19960646, 1996. 147. Morren, S. H., Myers, R. M., Benko, S. E ., Arrington, L. A., and Reed, B. R., Â“A Laboratory Model of a Hydrogen/Oxygen Engi ne for Combustion and Nozzle Studies,Â” NASA TM106281, 1993. 148. de Groot, W. A., McGuire, T. J., and Schneider, S. J., Â“Qualitative Fl ow Visualization of a 110N Hydrogen/Oxygen Laboratory Model Thruster,Â” AIAA Paper 19972847, 1997. 149. Strakey, P., Talley, D., Tseng, L., and Mine r, K., Â“The Effects of LOX Post Biasing on SSME Injector Wall Compatibility,Â” AIAA Paper 19992888, 1999. 150. Smith, T., Klem, M., Breisacher, K., Farh angi, S., and Sutton, R., Â“Experimental Evaluation of a Subscale Gase ous Hydrogen/Gaseous Oxygen Coaxial Rocket Injector,Â” NASA TM 2002211982, 2002. 151. Cheng, G. C., Anderson, P. G., and Farmer, R. C., Â“Development of CFD Model for Simulating Gas/Liquid Injectors in Rocket Engine Design, AIAA Paper 19973228, 1997. 152. Schley, C., Hagemann, G., Tucker, K. P., Venkateswaran, S., and Merkle, C. L., Â“Comparison of Computati onal Codes for Modeling Hydr ogenOxygen Injectors,Â” AIAA Paper 19973302, 1997. 153. Schley, C.A., Hagemann, G., Golovitchev, V., Â“Comparison of High Pressure H2/O2 Rocket Model Engine Reference Si mulations,Â” AIAA Paper 19952429, 1995.
PAGE 236
236 154. Ivancic, B., Mayer, W., Krlle, G., and Brggemann, D., Â“Experimental and Numerical Investigation of Time and Length Scales in LOX/GH2Rocket Combustors, AIAA Paper 19992211, 1999. 155. Lin, J., West, J., Williams, R., Tucker, K., and Chenoweth, J., Â“CFD Code Validation of Wall Heat Fluxes for a GO2/GH2 Single El ement Combustor,Â” AIAA Paper 20054524, 2005. 156. West, J., Westra, D., Lin, J., and Tucker, P. K., Â“Accuracy Quantification of the LociCHEM Code for Chamber Wall Heat Fluxes in a GO2/GH2 Single Element Injector Model Problem,Â” 3rd International Workshop on Rocket Combustion Modeling Paris, France, 2005. 157. Thakur, S. and Wright, J., Â“V alidation of a PressureBased Combustion Simulation Tool for a Single Element Injector Test Problem, 3rd International Workshop on Rocket Combustion Modeling Paris, France, March 1315, 2006. 158. Tucker, K., West, J., Williams, R., Lin, J., Rocker, M., Canabal, F., Robles, B., and Garcia, R., Â“Using CFD as a Rocket Injector De sign Tool: Recent Progress at Marshall Space Flight Center,Â” NASA Marshall Space Flight Center CASI ID 20050217148, 2005. 159. Meyer, L., Nichols, J., Jones, J. M., and Sabol, W., Â“Integrated Powerhead Demonstrator (Booster Hydrogen Oxygen Rocket E ngines),Â” AIAA Paper 19964264, 1996. 160. Ber, J. M. and Chigier, N. A., Combustion Aerodynamics Robert E. Krieger Publishing Company, Malabar, FL, 1983, p. 1. 161. Hill, P. and Peterson, C., Mechanics and Thermodynamics of Propulsion AddisonWesley, 1992, pp. 124Â–130. 162. Turns, S. R., An Introduction to Combustion McGraw Hill, 2000, p. 486. 163. Launder, B. and Spalding, D., Â“The Nume rical Computation of Turbulent Flows,Â” Computer Methods in Applie d Mechanics and Engineering Vol. 3, No. 2, 1974, pp. 269Â– 289. 164. Patel, V., Rodi, W., and Scheuerer, G., Â“Turbulence Models for NearWall and Low Reynolds Number Flows: A Review,Â” AIAA Journal Vol. 23, No. 9, 1985, pp. 1308Â–1319. 165. Avva, R., Smith, C., and Singhal, A., Â“Comparative Study of High and Low Reynolds Number Versions of kepsilon Models,Â” AIAA Paper 19900246, 1990. 166. Chien, K., Â“Predictions of Channel a nd BoundaryLayer Flows with Low ReynoldsNumber Turbulence Model,Â” AIAA Journal Vol. 20, No. 1, 1982, pp. 3338. 167. Pope, S. B., Turbulent Flows Cambridge University Press, New York, 2000, pp. 442Â–444.
PAGE 237
237 168. Tennekes, H. and Lumley, J., A First Course in Turbulence MIT Press, Cambridge, MA, 1972, p. 156. 169. Libby, P. A. and Williams, F. A., Turbulent Reacting Flows Academic Press, New York, 1994. 170. Thakur, J., Wright, J., and Shyy, W., Â“A Pr essureBased Algorithm for Reacting Flows with Finite Rate Chemistry,Â” AIAA Paper 20042334, 2004. 171. Weske, D. R. and Sturov, G. YE., Â“Experim ental Study of Turbulent Swirled Flows in a Cylindrical Tube,Â” Fluid Mechanics Â– Soviet Research Vol. 3, No. 1, 1974. 172. Yajnik, K. S. and Subbaiah, M. V., Â“Experiments on Swirling Turbulent Flows. Part 1. Similarity in swirling flows,Â” Journal of Fluid Mechanics Vol. 60, No. 4, 1973, pp. 665Â– 687. 173. Talbot, L., Â“Laminar Swirling Pipe Flow,Â” Journal of Applied Mechanics Vol 21., No. 1, 1954, pp. 1Â–7. 174. Morton, S. A., Steenman, M. B., Cummings, R. M., and Forsythe, J. R., Â“DES Grid Resolution Issues for Vortical Flows on a Delta Wing and an F18C,Â” AIAA Paper 20031103, 2003. 175. Burman, J., Papila, N., Shyy, W., and Geba rt, B. R., Â“Assessment of Response SurfaceBased Optimization Techniques for Unstea dy Flow Around Bluff B odies,Â” AIAA Paper 20025596, 2002 176. Thakur, S. and Wright J., Â“STREAM A Computational Fluid Dynamics and Heat Transfer Code NavierStokes Solver: Theo ry & Applications,Â” Streamline Numerics, Gainesville, Florida, 2002. 177. ICEM CFD version 5.0, IC EM CFD Engineering, 2004. 178. Goel, T., Mack, Y., Shyy, W., Haftka, R. T ., Queipo, N.V., Â“Numerical and Surrogate Model Uncertainty Assessment for Bluff Body Facilitated Mixing,Â” AIAA Paper 20050125, 2005. 179. Incropera, F. P. and DeWitt, D. P., Fundamentals of Heat and Mass Transfer John Wiley & Sons, Inc., New York, 1996, pp. 863. 180. Vaidyanathan, A. and Segal, C., Â“Quantitat ive Measurements of OH Concentration in GH2/GO2 combustion,Â” AIAA Paper 20075591, 2007. 181. Wang, B. P., Â“A New Method for Dual Response Surface Optimization,Â” AIAA Paper 19964041, 1996.
PAGE 238
238 BIOGRAPHICAL SKETCH Yolanda Mack was born in Willingboro, New Jersey, on January 2, 1980. Born to parents who served in the U.S. Air Force, she has lived around the country and in Japan. In 2002, she received her Bachelor of Science degree with honors in mechanical engineering from the University of Florida. She also received her Master of Science degree in 2004 from the University of Florida specializing in computa tional fluid dynamics(CFD)based optimization. In 2004, she was awarded a fellowship by the South Ea st Alliance for Graduate Education and the Professoriate that seeks to incr ease diversity in the professoriat e in the science, technology, engineering and mathematics fields. She is also a 2006 recipient of the Amelia Earhart Fellowship awarded by Zonta Intern ational to outstanding women Ph.D. students specializing in aerospacerelated sciences and engineering. She pursued her doctoral studies in CFDbased optimization at the University of Florida in aerospace engineering under the guidance of Drs. Wei Shyy and Raphael Haftka. Yolanda currently lives in Gainesville and stays active by spending time with her son, Trevor. After completing her doctoral studies, she will join Raytheon in Tucson, AZ as a Senior Multidisciplined Engineer and plans on later pursuing an academic career.
