|
|
| Citation |
- Permanent Link:
- https://ufdc.ufl.edu/UFE0043944/00001
Material Information
- Title:
- Deterministic and Reliability Based Optimization of Integrated Thermal Protection System Composite Panel using Adaptive Sampling Techniques
- Creator:
- Ravishankar, Bharani P
- Place of Publication:
- [Gainesville, Fla.]
Florida
- Publisher:
- University of Florida
- Publication Date:
- 2012
- Language:
- english
- Physical Description:
- 1 online resource (130 p.)
Thesis/Dissertation Information
- Degree:
- Doctorate ( Ph.D.)
- Degree Grantor:
- University of Florida
- Degree Disciplines:
- Mechanical Engineering
Mechanical and Aerospace Engineering
- Committee Chair:
- Sankar, Bhavani V
- Committee Co-Chair:
- Haftka, Raphael T
- Committee Members:
- Kumar, Ashok V
Consolazio, Gary R
- Graduation Date:
- 5/5/2012
Subjects
- Subjects / Keywords:
- Boundary conditions ( jstor )
Buckling ( jstor )
Design optimization ( jstor )
Estimate reliability ( jstor )
Estimation methods ( jstor )
Finite element analysis ( jstor )
Modeling ( jstor )
Monte Carlo methods ( jstor )
Random variables ( jstor )
Stiffness ( jstor )
Mechanical and Aerospace Engineering -- Dissertations, Academic -- UF
deterministic -- homogenization -- optimization -- reliability-based
- Genre:
- bibliography ( marcgt )
theses ( marcgt )
government publication (state, provincial, terriorial, dependent) ( marcgt )
born-digital ( sobekcm )
Electronic Thesis or Dissertation
Mechanical Engineering thesis, Ph.D.
Notes
- Abstract:
- Conventional space vehicles have thermal protection systems (TPS) that provide protection to an underlying structure that carries the flight loads. In an attempt to save weight, there is interest in an integrated TPS (ITPS) that combines the structural function and the TPS function. This has weight saving potential, but complicates the design of the ITPS that now has both thermal and structural failure modes.
The main objectives of this dissertation was to optimally design the ITPS subjected to thermal and mechanical loads through deterministic and reliability based optimization.
The optimization of the ITPS structure requires computationally expensive finite element analyses of 3D ITPS (solid) model. To reduce the computational expenses involved in the structural analysis, finite element based homogenization method was employed, homogenizing the 3D ITPS model to a 2D orthotropic plate. However it was found that homogenization was applicable only for panels that are much larger than the characteristic dimensions of the repeating unit cell in the ITPS panel. Hence a single unit cell was used for the optimization process to reduce the computational cost. Deterministic and probabilistic optimization of the ITPS panel required evaluation of failure constraints at various design points. This further demands computationally expensive finite element analyses which was replaced by efficient, low fidelity surrogate models. In an optimization process, it is important to represent the constraints accurately to find the optimum design. Instead of building global surrogate models using large number of designs, the computational resources were directed towards target regions near constraint boundaries for accurate representation of constraints using adaptive sampling strategies.
Efficient Global Reliability Analyses (EGRA) facilitates sequentially sampling of design points around the region of interest in the design space. EGRA was applied to the response surface construction of the failure constraints in the deterministic and reliability based optimization of the ITPS panel.
It was shown that using adaptive sampling, the number of designs required to find the optimum were reduced drastically, while improving the accuracy.
System reliability of ITPS was estimated using Monte Carlo Simulation (MCS) based method. Separable Monte Carlo method was employed that allowed separable sampling of the random variables to predict the probability of failure accurately. The reliability analysis considered uncertainties in the geometry, material properties, loading conditions of the panel and error in finite element modeling. These uncertainties further increased the computational cost of MCS techniques which was also reduced by employing surrogate models. In order to estimate the error in the probability of failure estimate, bootstrapping method was applied.
This research work thus demonstrates optimization of the ITPS composite panel with multiple failure modes and large number of uncertainties using adaptive sampling techniques. ( en )
- General Note:
- In the series University of Florida Digital Collections.
- General Note:
- Includes vita.
- Bibliography:
- Includes bibliographical references.
- Source of Description:
- Description based on online resource; title from PDF title page.
- Source of Description:
- This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
- Thesis:
- Thesis (Ph.D.)--University of Florida, 2012.
- Local:
- Adviser: Sankar, Bhavani V.
- Local:
- Co-adviser: Haftka, Raphael T.
- Statement of Responsibility:
- by Bharani P Ravishankar.
Record Information
- Source Institution:
- UFRGP
- Rights Management:
- Copyright Ravishankar, Bharani P. Permission granted to the University of Florida to digitize, archive and distribute this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
- Classification:
- LD1780 2012 ( lcc )
|
| Downloads |
This item has the following downloads:
|
| Full Text |
PAGE 1
DETERMINISTICANDRELIABILITYBASEDOPTIMIZATIONOFINTEGRATEDTHERMALPROTECTIONSYSTEMCOMPOSITEPANELUSINGADAPTIVESAMPLINGTECHNIQUESByBHARANIRAVISHANKARADISSERTATIONPRESENTEDTOTHEGRADUATESCHOOLOFTHEUNIVERSITYOFFLORIDAINPARTIALFULFILLMENTOFTHEREQUIREMENTSFORTHEDEGREEOFDOCTOROFPHILOSOPHYUNIVERSITYOFFLORIDA2012
PAGE 2
c2012BharaniRavishankar 2
PAGE 3
Tomyparents,RavishankarandParameswariandmysister,Vidhya 3
PAGE 4
ACKNOWLEDGMENTSFirstandforemostIwouldliketothankmyadvisorsDr.BhavaniSankarandDr.RaphaelHaftkaattheDepartmentofMechanicalandAerospaceEngineering.Allofyourinsights,guidance,andpatiencehasbeengreatlyappreciated.Iwouldalsoliketothankthemembersofmyadvisorycommittee,Dr.AshokKumarattheDepartmentofMechanicalandAerospaceEngineeringandDr.GaryConsolazioattheDepartmentofCivilandCoastalEngineering.IwouldlikethankmyfriendsattheCenterforAdvancedCompositeslab,Anurag,Prasanna,MinSong,Marlana,TimandSayan.IwouldalsoliketothankthemembersoftheStructuralandMultidiscplinaryGroup,Ben,Felipe,Diane,AnirbanandTaikiforalltheirhelp,valuableinputsandsuggestions.IwouldalsoliketothankmyfriendsAnirudh,SriramandNarenforallthehelpandsupportthroughoutmyPhDprogram,makingthisexperiencememorable.Iwouldliketothankmyfamilyfortheirsupportandpatience. 4
PAGE 5
TABLEOFCONTENTS page ACKNOWLEDGMENTS .................................. 4 LISTOFTABLES ...................................... 8 LISTOFFIGURES ..................................... 11 ABSTRACT ......................................... 13 CHAPTER 1INTRODUCTION ................................... 15 1.1OutlineofDissertation ............................. 19 1.1.1MotivationandObjectives ....................... 19 1.1.2ThesisOrganization .......................... 20 2BACKGROUNDSTUDYANDLITERATUREREVIEW .............. 21 2.1OrbiterThermalProtectionSystem-30yearsLegacy ............ 21 2.1.1TileTPS ................................. 23 2.1.2Non-TileTPS ............................. 23 2.2FiniteElementbasedHomogenization .................... 24 2.3DeterministicandReliabilitybasedoptimization ............... 27 2.3.1DeterministicOptimization ....................... 28 2.3.2StructuralReliability .......................... 29 2.4UseofSurrogateModels ........................... 31 2.4.1ApplicationofSurrogatestoImproveConstraintBoundaries .... 32 2.5AdaptiveSampling ............................... 33 2.5.1EfcientGlobalReliabilityAnalysis .................. 34 2.6UncertaintyModeling ............................. 35 2.7EstimationofProbabilityofFailure ...................... 36 2.7.1MomentbasedMethods ........................ 37 2.7.2SamplingMethods ........................... 38 3DESIGNANDANALYSISOFINTEGRATEDTHERMALPROTECTIONSYSTEMPANEL ........................................ 41 3.1DesignofITPSCompositePanel ....................... 41 3.1.1KeyDimensionsoftheITPSstructure ................ 41 3.1.2ITPSMaterials ............................. 42 3.1.3BoundaryConditions .......................... 43 3.1.4Loads .................................. 44 3.2FiniteElementAnalyses ............................ 44 3.2.1TransientHeatTransferAnalysis ................... 45 3.2.2StressAnalysis ............................. 47 5
PAGE 6
3.2.3BucklingAnalysis ............................ 48 3.3Summary .................................... 50 4FINITEELEMENTBASEDHOMOGENIZATION ................. 51 4.1A,B,Dmatrices ................................. 51 4.2TransverseShearStiffness .......................... 53 4.3AccuracyoftheHomogenizationMethod .................. 57 4.3.1DeectionComparison ......................... 57 4.3.2StressComparison ........................... 57 4.4ResultsandDiscussion ............................ 58 4.4.1A,B,Dmatrices-Equivalentmaterialvs.Equivalentstructure ... 58 4.4.2TransverseShearStiffnessofITPS .................. 60 4.4.3AccuracyoftheHomogenizationMethod ............... 63 4.5Summary .................................... 66 5MONTECARLOSIMULATIONS-RELIABILITYESTIMATION ......... 67 5.1CrudeMonteCarloMethod(CMC) ...................... 67 5.2SeparableMonteCarlomethod(SMC) .................... 68 5.3ErrorintheProbabilityofFailureEstimate .................. 69 5.4SMCwithregroupingandseparablesamplingofthelimitstaterandomvariables .................................... 71 5.5ApplicationtoFailureAnalysisofCompositeLaminate ........... 72 5.6ResultsandDiscussion ............................ 75 5.6.1CrudeandSeparableMonteCarloMethod .............. 75 5.6.2Regroupingandseparablesamplingofthelimitstatevariablesforimprovingaccuracy ........................... 78 5.6.3Summary ................................ 80 6EFFICIENTGLOBALRELIABILITYANALYSIS(EGRA) ............. 82 6.1EGRAalgorithm ................................ 82 6.2IllustrationofEGRA .............................. 83 6.3Summary .................................... 88 7DETERMINISTICANDRELIABILITYBASEDOPTIMIZATION ......... 90 7.1DeterministicOptimization ........................... 90 7.2UncertaintyModeling ............................. 93 7.3EstimationofSystemReliability ........................ 94 7.4ReliabilityBasedOptimization ......................... 96 7.5ResultsandDiscussion ............................ 98 7.5.1DeterministicOptimization ....................... 98 7.5.2SensitivityAnalysis ........................... 104 7.5.3ReliabilityoftheDeterministicOptimum ............... 107 7.5.4Reliabilitybasedoptimization ..................... 110 6
PAGE 7
7.5.5Errorintheprobabilityoffailureestimate ............... 113 7.5.6Summary ................................ 115 8CONCLUSIONS ................................... 116 8.1Conclusion ................................... 116 8.2FutureWork ................................... 117 ATHERMALPROPERTIESOFITPSCOMPONENTS ............... 118 BSTRENGTHPROPERTIESOFITPSMATERIALS ................ 120 CDESIGNOFEXPERIMENTSADDEDBYEGRASAMPLINGTECHNIQUE .. 121 REFERENCES ....................................... 123 BIOGRAPHICALSKETCH ................................ 130 7
PAGE 8
LISTOFTABLES Table page 3-1KeyDimensionsoftheITPSpanel ......................... 41 3-2ComponentsofITPS,materialsandberorientation ............... 43 3-3MaterialpropertiesofthecomponentsofITPS .................. 43 3-4Heatuxloadstepsinthetransientheattransferanalysis ............ 46 4-1Periodicboundaryconditionsforthesixdeformations .............. 52 4-2VariationofA44andA55ofITPSpanelwithn ................... 61 4-3Tipdeectionratioalongwithcontributionofbendingandsheardeformationtowardsdeection .................................. 64 5-1Materialpropertiesanduncertaintyoftherandomvariables ........... 74 5-2EmpiricalandbootstrappingestimatesofprobabilityoffailureusingseparableandcrudeMonteCarlowithN=M=500andn=10,000repetitions ....... 76 5-3StandarddeviationandcoefcientofvariationofempiricalandbootstrappingpfestimatesusingseparableandcrudeMonteCarlowithN=M=500andn=10,000repetitionsfororiginallimitstate ...................... 77 5-4Relativecontributionsofresponse(stresses)andcapacity(strengths)towardstheuncertaintyinpfthroughbootstrappingandalsocomparedwithempiricalresults ......................................... 77 5-5StandarddeviationandcoefcientofvariationofCMCandSMCforincreasingsamplesizeofMandN=500 ............................ 79 5-6StandarddeviationandcoefcientofvariationofCMC,SMCandSMCregroupedforincreasingsamplesizeofMandN=500 .................... 79 7-1Failure,safetyfactorsappliedonconstraints .................... 91 7-2Lowerandupperboundsofdesignvariablesfordeterministicoptimization ... 92 7-3CoefcientofvariationofinputrandomvariablesincludedintheITPSdesign 93 7-4Nominalallowablevaluesofcapacityandcoefcientofvariation ........ 94 7-5OptimumdesignvariablesandminimizedstructuralmassthroughglobalandEGRADOEfordeterministicoptimization ..................... 101 7-6ComparisonbetweentheaccuracyofglobalsurrogateandEGRAsurrogateatthedeterministicoptimumusingcorrespondingsurrogates .......... 103 8
PAGE 9
7-7Comparisonofresponsesatallthedeterministicoptimausingthe60DOEsurrogate ....................................... 103 7-8Comparisonofresponsesatallthedeterministicoptimausingthe20DOEsurrogate ....................................... 104 7-9ComparisonofresponsesatallthedeterministicoptimausingtheEGRAupdatedsurrogate ....................................... 104 7-10Differenceinuncertaintyinresponsesbeforeandaftersensitivityanalysis ... 106 7-11Uncertaintyinresponseduetoinputuncertainty ................. 108 7-12Individualprobabilitiesoffailureofthedeterministicoptimum .......... 109 7-13Lowerandupperboundsofdesignvariablesforprobabilisticoptimization ... 110 7-14OptimumdesignvariablesandminimizedstructuralmassthroughEGRAupdatedDOEforreliabilitybasedoptimization ........................ 112 7-15Individualprobabilitiesoffailureoftheprobabilisticoptimum ........... 112 7-16Individualprobabilitiesoffailureofthedeterministicoptimumusingprobabilisticoptimumsurrogate .................................. 112 7-17Errorintheprobabilitiesoffailureofthedeterministicoptimumusingbootstrappingandrepetitions .................................... 114 7-18Errorintheprobabilitiesoffailureoftheprobabilisticoptimumusingbootstrappingandrepetitions .................................... 114 A-1DensityoftheITPSComponents .......................... 118 A-2ThermalPropertiesofthebottomfacesheet-GraphiteEpoxy ......... 118 A-3ThermalPropertiesofthetopfacesheetandWrap-SiC/SiC .......... 118 A-4ThermalConductivityoftheinsulationfoam-AETB ............... 119 A-5Specicheatoftheinsulationfoam-AETB .................... 119 B-1DensityoftheITPSComponents .......................... 120 C-1DesignpointsacquiredusingEGRAtoimprovetheaccuracyofTemperatureboundary ....................................... 121 C-2DesignpointsacquiredusingEGRAtoimprovetheaccuracyofstressboundary 121 C-3DesignpointsacquiredusingEGRAtoimprovetheaccuracyofbucklingloadboundary ....................................... 121 9
PAGE 10
C-4DesignpointsacquiredusingEGRAtoimprovetheaccuracyofreliabilityindexboundary ....................................... 122 10
PAGE 11
LISTOFFIGURES Figure page 2-1ThermalProtectionSystem(TPS)usedinspaceshuttles ............ 22 2-2Sinusoidalcoresandwichedbetweenlaminates .................. 26 2-3IntegratedThermalProtectionSystemwithcorrugatedcore ........... 26 3-1SchematicviewoftheITPSpanel ......................... 42 3-2DimensionsoftheITPSpanel ............................ 42 3-3BoundaryconditionsappliedontheITPSpanel .................. 44 3-4HeatuxproleincidentontheITPSpanel .................... 45 3-5Boundaryconditionsforthetransientheattransferanalysis ........... 46 3-6TemperaturevariationonTFSandBFSwithrespecttoreentrytime ...... 47 3-7BuckledITPSduetothermalloads ........................ 49 4-1SchematicviewoftheITPSunitcell(RVE) .................... 52 4-2Periodicboundaryconditionappliedontheunitcellforx=1 .......... 54 4-3Schematicviewoftheunitcelldeformationsduetostrainsandcurvatures ... 54 4-41-DITPSmodelwithunitcellsinx-direction ................... 55 4-51DmodeloftheITPSwithunitcellsinx-directionsubjectedtopressureloadandxedBCs ..................................... 58 4-6TransverseshearstiffnessA55oforthotropicpanel ................ 61 4-7VariationofA44andA55ofITPSpanelwithn ................... 62 4-8ConvergedA44andA55ofITPSpanelwithn ................... 62 4-9Ratiobetweentipdeectionofthehomogenizedmodelandthe3-DITPSmodelalongthexandy-direction ............................. 63 4-10Stress(11,22)ratiobetweenthehomogenizedmodelandthe3DITPSmodel 65 4-11Stress(11,22)atn=10comparisonbetweenthe3-DITPSmodelandthehomogenizedmodel ................................. 65 4-12Verticaldeectionalongx-axisoftheoriginalITPSandhomogenizedmodelduetopressureload ................................. 66 5-1IllustrationofcrudeandseparableMonteCarloMethodcomparisons ...... 69 11
PAGE 12
5-2Schematicrepresentationofbootstrappingwhenonlyresponseissampled. .. 70 5-3Illustrationofseparablesamplingwithunitloads ................. 72 5-4Compositepressurevesselwithinternalpressureof100kPaandstressesactinginasmallelementofthevessel. ....................... 73 5-5Distributionofstressandstrengthinthe2-direction(2)showingtheprobablefailureregion ..................................... 75 5-6StandardDeviationofCMC,SMCandregroupedlimitstateSMCwhereN=500(xed)andMisvaryingfor10,000repetitions ................... 80 6-1TruecontoursoftheBraninhoofunction ...................... 83 6-2TargetcontourapproximatedbytheinitialkrigingmodeloftheBraninFunction 84 6-3ExpectedFeasibilityFunctionevaluatedusingtheinitialkrigingapproximation 85 6-4Krigingmodelupdatedwiththenewdesignandexpectedfeasibilityfunction .. 86 6-5TargetcontourapproximatedbytheinitialkrigingmodeloftheBraninFunction 86 6-6TargetcontouraccuratelyapproximatedbytheEGRAmethodology ...... 87 7-1VariationofresponseswrtthicknessofthewraptW ............... 99 7-2VariationofcriticalresponseswrtthicknessofthebottomfacesheettB .... 100 7-3VariationofcriticalresponseswrtthicknessofthetopfacesheettT ....... 101 7-4Variationofcriticalresponseswrtheightofthefoamh .............. 102 7-5SensitivityanalysisofmaximumBFStemperature ................ 106 7-6SensitivityanalysisofWrapstress11 ....................... 107 7-7Sensitivityanalysisofbucklingloadfactor ..................... 108 12
PAGE 13
AbstractofDissertationPresentedtotheGraduateSchooloftheUniversityofFloridainPartialFulllmentoftheRequirementsfortheDegreeofDoctorofPhilosophyDETERMINISTICANDRELIABILITYBASEDOPTIMIZATIONOFINTEGRATEDTHERMALPROTECTIONSYSTEMCOMPOSITEPANELUSINGADAPTIVESAMPLINGTECHNIQUESByBharaniRavishankarMay2012Chair:BhavaniV.SankarCochair:RaphaelT.HaftkaMajor:MechanicalEngineering Conventionalspacevehicleshavethermalprotectionsystems(TPS)thatprovideprotectiontoanunderlyingstructurethatcarriestheightloads.Inanattempttosaveweight,thereisinterestinanintegratedTPS(ITPS)thatcombinesthestructuralfunctionandtheTPSfunction.Thishasweightsavingpotential,butcomplicatesthedesignoftheITPSthatnowhasboththermalandstructuralfailuremodes.ThemainobjectivesofthisdissertationwastooptimallydesigntheITPSsubjectedtothermalandmechanicalloadsthroughdeterministicandreliabilitybasedoptimization. TheoptimizationoftheITPSstructurerequirescomputationallyexpensiveniteelementanalysesof3DITPS(solid)model.Toreducethecomputationalexpensesinvolvedinthestructuralanalysis,niteelementbasedhomogenizationmethodwasemployed,homogenizingthe3DITPSmodeltoa2Dorthotropicplate.HoweveritwasfoundthathomogenizationwasapplicableonlyforpanelsthataremuchlargerthanthecharacteristicdimensionsoftherepeatingunitcellintheITPSpanel.Henceasingleunitcellwasusedfortheoptimizationprocesstoreducethecomputationalcost. DeterministicandprobabilisticoptimizationoftheITPSpanelrequiredevaluationoffailureconstraintsatvariousdesignpoints.Thisfurtherdemandscomputationallyexpensiveniteelementanalyseswhichwasreplacedbyefcient,lowdelitysurrogate 13
PAGE 14
models.Inanoptimizationprocess,itisimportanttorepresenttheconstraintsaccuratelytondtheoptimumdesign.Insteadofbuildingglobalsurrogatemodelsusinglargenumberofdesigns,thecomputationalresourcesweredirectedtowardstargetregionsnearconstraintboundariesforaccuraterepresentationofconstraintsusingadaptivesamplingstrategies. EfcientGlobalReliabilityAnalyses(EGRA)facilitatessequentiallysamplingofdesignpointsaroundtheregionofinterestinthedesignspace.EGRAwasappliedtotheresponsesurfaceconstructionofthefailureconstraintsinthedeterministicandreliabilitybasedoptimizationoftheITPSpanel.Itwasshownthatusingadaptivesampling,thenumberofdesignsrequiredtondtheoptimumwerereduceddrastically,whileimprovingtheaccuracy. SystemreliabilityofITPSwasestimatedusingMonteCarloSimulation(MCS)basedmethod.SeparableMonteCarlomethodwasemployedthatallowedseparablesamplingoftherandomvariablestopredicttheprobabilityoffailureaccurately.Thereliabilityanalysisconsidereduncertaintiesinthegeometry,materialproperties,loadingconditionsofthepanelanderrorinniteelementmodeling.TheseuncertaintiesfurtherincreasedthecomputationalcostofMCStechniqueswhichwasalsoreducedbyemployingsurrogatemodels.Inordertoestimatetheerrorintheprobabilityoffailureestimate,bootstrappingmethodwasapplied. ThisresearchworkthusdemonstratesoptimizationoftheITPScompositepanelwithmultiplefailuremodesandlargenumberofuncertaintiesusingadaptivesamplingtechniques. 14
PAGE 15
CHAPTER1INTRODUCTION Whenaspacevehiclereenterstheatmosphereathypersonicspeeds,thevehicle'sexteriorissubjectedtosevereaerodynamicheatingandpressure.Toprotectthevehiclefromsuchextremetemperatures,aThermalProtectionSystem(TPS),isaddedontothemainloadbearingstructure.But,theTPSisincompatiblewiththemainstructureduetothemismatchintheirthermalproperties.Thisincompatibility,alongwiththeactionofnumerousotherloadslikeaerodynamicpressure,impactloadsandotherin-planeinertialloadsmayleadtothestructure'sfailurewithcatastrophicconsequences.Moreover,theTPSencompassesasignicantportionofthevehicle'sexteriorconstitutingmajorpartofthelaunchweight.Thus,itisimperativethatapartfrommakingtheTPSsuitableforprotectionpurposes,itshouldbelightweightinordertoincreasepayloadsthatcanbecarriedaboard. Toaccommodatetheserequirements,amultifunctionalIntegratedThermalProtectionSystem(ITPS)couldbedesigned,inwhichtheloadbearingstructureandtheTPSareintegratedintoacompositesandwichpanel.ThegreatestchallengeinsuchadesignwouldbeincombiningthevariousconictingrequirementsoftheTPSandtheloadbearingmemberintoasinglestructure.Thestructuralrequirementsfavorusingmetals,butthiscangreatlyincreasetheconductionofheatthroughtheITPS.Ontheotherhandmostinsulationmaterialssuchasceramicsandfoamshavelowstrength.Materialselectionisimportantaswellasdimensioningthestructureproperly,transformingthisdesignintoamultidisciplinaryoptimizationproblem.Evenassumingthatthematerialpropertiesandtheirexactbehaviorareknownaccurately,thisproblemiscomputationallyintensiveinvolvinglargenumberofniteelementanalyses.IfcompositesareusedforITPS,itfurtherincreasesthenumberofuncertaintiesintheproblem.Whentheseuncertaintiesareincludedinthedesign,thecomputationalexpensebecomesprohibitivelylarge.Thiscouldbesolvedusingtraditionaldeterministic 15
PAGE 16
methods,whicharecomputationallyfeasiblebutlessaccurate.Thisproblemcouldbedenitivelyaddressedthroughprobabilisticoptimization,whichisaccuratebutcomputationallyintractable. TheoverallobjectiveofthisdissertationistoemployefcientandcosteffectivemethodstooptimizetheITPSpanel.Varioustechniqueshavebeenappliedalongthedifferentphasesofthisdissertationtoincreasethecomputationalefciencyandreducethecostandtimeinvolved.ToreducethecomputationalcostofstructuralanalysisofITPS,niteelementbasedhomogenizationisinvestigated.Homogenizationisaprocessofapproximatingthebehaviorofheterogeneousstructuresashomogeneousbydeterminingtheirequivalentstiffnessproperties.Pastresearcheffortshaveshownthathomogenizationofthecompositestructureasanequivalenttwo-dimensionalorthotropicplatewouldreducethecomputationalcostassociatedwithstructuralanalysistremendously[ 1 2 3 4 5 ].Sharmaetal.[ 6 7 ]appliedthehomogenizationtoanITPSpanelwithcorrugatedcore.Owingtonegligiblefoamdensity,thepanelwasmodeledusingshellelements.TheITPSpanelconsideredinthisdissertationhasdenserinsulationfoamrequiring3-Dbrickelements.Thisresearchthusaimstoextendthehomogenizationtechniqueto3-Dsandwichpanels. Inhomogenization,theequivalentstiffnesspropertiesofacompositesandwichpanelarethein-planestiffness[A],couplingstiffness[B],bendingstiffness[D]whicharecombinedandrepresentedasthe[ABD]stiffnessmatrixandtransverseshearstiffness(A44andA55).Theequivalentstiffnesspropertiesweredeterminedbytakingadvantageoftherepeatingstructureofthepanel.Therepetitiveblockofthepanelisreferredasunitcellorrepresentativevolumeelement(RVE).The[ABD]stiffnessmatrixwasdeterminedbysubjectingtheunitcelltounitstrainsandcurvatures.Thetransverseshearstiffnesswasdeterminedbyanalyzingaone-dimensionalplateoftheITPSmodelunderendloads.However,itwasfoundthathomogenizationofthecompositesandwichpanelwasaccurateonlywhenthepaneldimensionsaremuch 16
PAGE 17
largerthanthecharacteristicdimensionsoftheunitcell.HenceoneunitcellofITPSpanelisconsideredandanalyzedduringtheoptimizationprocess. Thesecondobjectiveofthisresearchistoperformdesignoptimizationthroughdeterministicandreliabilitybasedapproaches.OneofthemainpurposesofoptimizationistominimizetheweightoftheITPSwhilesatisfyingseveralconictingdesignconstraints.Designoptimizationthroughatraditional,safetyfactorapproachwithoutconsideringtheuncertaintiesinvolvedwouldprovidelessaccurateandconservativedesigns.Thecompositepanelunderconsiderationcontributestothemajorpartoftheweightontheexteriorofthespacevehicle.ThusitisimperativetoreducethestructuralweightofITPSwithoutresultinginconservativeandunsafedesigns.Hence,itisdesiredtooptimizethedesignoftheITPSusingreliabilitybasedapproachbyincludingalltheuncertaintiesinthedesign.Howeveritisdesiredtoperformaninitialdeterministicoptimizationasitwouldhelpinnarrowingtotheappropriatedesignspacetoperformreliabilitybasedoptimization. Likemostengineeringstructures,theITPScouldundergomultiplefailuremodes:temperature,stressesandbucklingloads.Fordeterministicoptimization,thedesignvariablesareoptimizedtominimizethestructuralmass.Theoptimizationprocessinvolvesevaluationoftheaforementionedfailureinducersatlargenumberofdesignpointswhichincreasesthecomputationalcost.Constructionofsurrogatemodels(interchangeablyreferredasresponsesurfaceapproximation)oftheconstraintsatlimiteddesignpointsserveasanefcientalternativetohighdelityniteelementanalyses.Reliabilitybasedoptimizationhasanadditionalreliabilityconstraintthatisafunctionofdesignvariablesandrandomvariables.ThedesignvariablesaredimensionsoftheITPSpanelandtherandomvariablesconsideredarematerialproperties,geometry,thermalandmechanicalloads.Materialpropertiesarecomplextomodelastheydependontemperature.Geometryisanothersourceofuncertaintyaffectingboththestructuralbehaviorandthermalloading.Henceitisimportantto 17
PAGE 18
analyzethefailureinducerssuchstresses,temperatureandbucklingduetotheabovementioneduncertainties.Withlargenumberofuncertaintiesinvolved,evenconstructionofqualitysurrogatemodelsrequiresniteelementanalyzesatextremelylargenumberofrandominputs. Failureconstraintsthatseparatethefeasibledesignsfromtheinfeasibleoneshavetorepresentedaccuratelyinordertodetermineanaccurateoptimum.Commonlytheconstraintsareevaluatedusingglobalsurrogatemodelsusingspacellingdesignofexperiments.Thiswouldresultinwastingalargenumberofhighdelityniteelementanalyses.Earlierresearcheffortshaveintroducedsamplingmethodstoimproveaccuracyofthetargetregionswithlessnumberofdesigns[ 8 9 10 11 12 13 14 ].ThesamplingmethoddevelopedbyBichonetal.[ 14 ],EfcientGlobalReliabilityAnalyses(EGRA)sequentiallysamplesdesignpointsinthevicinityofthetargetregionusingagaussianprocessresponsesurfaceconstructedinitiallywithfewerdesignpoints.Thiscoulddrasticallyreducethenumberofcomputationallyexpensiveniteelementsimulationswhileresultinginaccurateoptimumdesign.TheITPSoptimizationprobleminvolvinghighdimensionaldesignspaceseekspolynomialresponsesurface.ThisresearchworkextendsthecurrentEGRAmethodologytoadapttothecurrentITPSproblembyincorporatingpolynomialsurrogatesinthealgorithm. Thegoalofreliabilityanalysisistodeterminetheprobabilitythatasystemwillfailinservice,giventhatitsbehaviorisdependentonrandominputs.Tocalculatethesystemreliability(probabilityoffailure),MonteCarlosimulation(MCS),acommonlyusedmethodformultiplefailuremodesisemployed.Toefcientlyhandlenumerousuncertaintiesandmultiplefailuremodes,SeparableMonteCarlo(SMC)developedbySmarsloketal.[ 15 ]wasadopted.Whentheuncertainrandomvariablesarestatisticallyindependent,theconceptofSMCistogrouptheuncertaintiesinvolvedintheproblemandsamplethemappropriatelytohaveanaccuratefailurepredictionforagivencomputationalbudget.PreviousworkwithSMConlyexploredsimplelimitstates 18
PAGE 19
expressedasadifferencebetweenarandomresponseandarandomcapacity[ 16 ].Thisresearchworklooksatamoregenerallimitstatefunctionthatcombinessetsofrandomresponseandcapacitycomponents[ 17 ].Furthertheerrorintheprobabilityoffailureestimateforthesimplelimitstatewasderivedintermsofthenumberofsamplesoftheresponseandcapacity.Thisisnotapplicableforthegeneralizedlimitstate.Thisresearchaimsatestimatingtheerrorintheprobabilityoffailureforseparablecaseusingthebootstrappingmethod[ 18 ],aresamplingtechnique,whichinvolvestakingthesamplesofresponse(expensive)andresamplingthemwithreplacement[ 19 ]. 1.1OutlineofDissertation 1.1.1MotivationandObjectives Tosummarizetheabovediscussion,themainobjectivesofthisresearcharegivenbelow. OptimizationofITPSiscomputationallyexpensive,becauseof Structuralanalysisofhighdelity3Dniteelementmodel Probabilisticanalysisincludingalluncertainties-aleatory(materialproperties,geometryandloadingconditions)andepistemicuncertainties(errorinmodelingandsimulation). MultiplefailuremodesofITPSdemandsMonteCarlosimulationsforreliabilitycalculations. TosuccessfullyoptimizetheITPSstructurethroughdeterministicandreliabilitybasedoptimization,thisdissertationproposes TodevelopareducedbutefcientmodelofITPStolowerthecomputationcostinvolvedintheniteelementanalyses. Toinvestigatelowdelitysurrogatemodels(ResponseSurfaceApproximationoftheconstraints)toreplacehighdelityniteelementanalyses. ToadaptEfcientGlobalReliabilityAnalysisthatfacilitatessequentiallysamplingaroundtheconstraintsrequiringfewernumberofsamplestoimprovetargetregionandpredictaccurateoptimum. 19
PAGE 20
ToutilizeseparableMonteCarlomethodthatallowsseparablesamplingoftheexpensiveandinexpensivevariableswhichwillreducecomputationalcostandalsoprovideimprovedaccuracy. Thisdissertationessentiallydemonstratesthemethodologyfordeterministicandreliabilitybasedoptimizationofcomplexstructureswithlargenumberofuncertaintiesusingsurrogatemodels,EGRA,separableMonteCarlotechniqueandbootstrappingmethod. 1.1.2ThesisOrganization Chapter 2 reviewsthethermalprotectionsystemusedinspacevehiclessuchasChallenger,DiscoverandColumbia.Thischapterpresentsareviewontheniteelementbasedhomogenizationtechnique.Further,presentsareviewonstructuraloptimizationandthedifferentaspectsofdeterministicandreliabilitybasedoptimization.Chapter 3 presentsdetaileddescriptiononthegeometry,components,loadsandboundaryconditionsofthemulti-functionalIntegratedThermalProtectionSystemfollowedbyadiscussiononniteelementanalysesofITPSinvolvedinthedesignandtheoptimizationprocess.Chapter 4 presenttheniteelementbasedhomogenizationmethodandresultsobtainedfromthemethod.Chapter 5 presentsadetaileddiscussiononMonteCarlosimulationtechniquestoestimateprobabilityoffailurewhichincludescrudeMonteCarloandseparableMonteCarlomethod.bootstrappingmethodtoestimateerrorintheprobabilityoffailureapplyingittoapressurevesselproblem.Chapter 6 illustratestheworkingofEfcientGlobalReliabilityAnalysis(EGRA)usingatwovariableproblem.EGRAappliedtotheoptimizationoftheITPSpanelisdiscussedinChapter 7 .ThischapterpresentsthedeterministicandreliabilitybasedoptimizationoftheITPSpanelandotherimportantcomponents(uncertaintymodelingandpropagation,sensitivityanalyses)associatedwiththeoptimizationprocess.ThedissertationisconcludedwithsummaryofresultsandpossiblefutureworkinChapter 8 20
PAGE 21
CHAPTER2BACKGROUNDSTUDYANDLITERATUREREVIEW Thermalprotectionsystemshavebeenanareaofcontinuousresearchanddevelopmentinspacevehicledesign.Withatmosphericheatingduringreentrybeingthemajorcauseofconcernforaspacevehicle'ssafeoperation,severaldesignandanalysistechniqueshavebeenputforthbyresearches.ThischapterreviewsthedifferenttypesofThermalProtectionsystemusedsofarinspacevehicles.FurtheritalsointroducesandreviewsthevarioustechniquesthatareaimedatreducingthecomputationalcostassociatedwiththedeterministicandreliabilitybasedoptimizationofthestructurallyintegratedThermalProtectionSystem. 2.1OrbiterThermalProtectionSystem-30yearsLegacy Whenaspaceshuttleorbiterre-enterstheEarth'satmosphere,itistravelinginexcessof17,000mph(7600m/s).Whileslowingdowntolandingspeed,frictionwiththeatmosphereproducesexternalsurfacetemperaturesashighas1900K.Specialthermalshieldsontheexteriorsurfaceprotectthevehicleanditsoccupantsduringlaunchandreentry[ 20 21 22 ].Earliermannedspacecraft,suchasMercury,GeminiandApollo,wereprotectedduringre-entrybyaheatshieldconstructedofphenolicepoxyresinsinanickel-alloyhoneycombmatrix.Theheatshieldwascapableofwithstandingveryhighheatingrates.Duringthereentry,theheatshieldwouldablate,orcontrollablyburnwiththecharlayerprotectingthelayersbelow.Despitetheadvantages,ablativeheatshieldshadsomemajordrawbacks.Theywerebondeddirectlytothevehicle,theywereheavy,andtheywerenotreusable.Withadesignlifeof100missions,spaceshuttleorbiterrequiredalightweightreusablethermalprotectionsystem(TPS)notonlytoprotecttheorbiterfromthesearingheatofreentry,butalsotoprotecttheairframeandmajorsystems[ 23 24 ]. Thevehicle'scongurationandentrytrajectorydenesthetemperaturedistributiononthevehicle.Overthepast30years,theSpaceShuttle'sThermalProtectionsystem 21
PAGE 22
hasbeenbuiltwithmaterialswithahightemperaturecapabilityandunderlyingthermalinsulationtoinhibittheconductionofheattotheinteriorofthevehicle.Duringtheentiremission,differentlocationsontheorbitergetheatedtodifferenttemperatures.Theleadingedgesofwingsandthenosecaparethehighesttemperatureregions.DuetothewidevariationofthesetemperaturestheTPSselectedforspaceshuttlewascomposedofmanydifferentmaterials.Eachmaterial'stemperaturecapability,durabilityandweightdeterminedtheextentofitsapplicationonthevehicle.TherearebasicallytwocategoriesofTPSbeingused,thetileTPSandnon-tileTPS.AdetaileddescriptionofeachcategorywouldprovideabetterunderstandingontheTPSselectionfordifferentregionsonthevehicle[ 25 26 ]. Figure2-1.ThermalProtectionSystem(TPS)usedinspaceshuttles[ 26 ] 22
PAGE 23
2.1.1TileTPS ThedifferenttypesoftileTPSare High-temperaturereusablesurfaceinsulation(HRSI)tilesweredevelopedtoprovideprotectionagainsttemperaturesupto1533K.Theywereusedinareasontheupperforwardfuselage,verticalstabilizerleadingedge,andupperbodyapsurface.Thetileiscomposedofhighpuritysilicabers.Ablackcoating,reactioncuredglass(RCG)wasappliedtoallbutonesideofthetiletoprotecttheporoussilicaandtoincreasetheheatsinkproperties.HRSIwasprimarilydesignedtowithstandtransitionfromareasofextremelylowtemperature,about-270Ctothehightemperaturesofre-entrytypicallyaround1600Cthusmaximizingheatrejectionduringthehotphaseofreentry. Fibrousrefractorycompositeinsulationtiles(FRCI)weresimilarformoftheHRSItiles.TheFRCItileshadhigherstrengthderivedbyaddingalumina-borosilicateber,calledNextel,tothepuresilicatileslurry.Thoughdevelopedforthesamepurpose,FRCIandHRSIhaddifferentphysicalpropertiesbecauseof20%Nextelinit.FRCItileswerelighterthanthebasicHRSItiles.Furthermore,theFRCItilesalsohadtensilestrengththatwasatleastthreetimesgreaterthanthatoftheHRSItilesandcouldbeusedatatemperaturealmost100CofhigherthanthatofHRSItiles.Inanutshell,theyprovidedabetterstrength,durability,crackingresistance,andweightreduction. Tougheneduni-piecebrousinsulation(TUFI)isanimprovedlowdensityrigidceramiccomposite,withveryhighimpactresistance(20-100timesmorethanRCGcoating).TUFIwasusedinregionswheretemperaturesreachashighas1260degreesCelsius.TUFItileswerebuiltashightemperatureblackversionsforuseintheorbiter'sundersideprovidingsufcientheatinsulationfortheorbiter'sunderside.Andlowertemperaturewhiteversionsforuseontheupperbodyconductingmoreheatwhichlimitstheirusetotheorbiter'supperbodyapandmainenginearea. Low-temperaturereusablesurfaceinsulation(LRSI)werewhiteincolorpossessinghighthermalreectivity.Thesetileswereusedtoprotectareaswherereentrytemperaturesarebelow649C(1200F).Theyaregenerallyinstalledontheuppersurfaceofthevehicle,maximizingsolargainwhentheorbiterisontheilluminatedpartoftheorbit.Theywerealsousedtoprotecttheupperwingneartheleadingedgeandalsoareasoftheforward,mid,andaftfuselage,verticaltail. 2.1.2Non-TileTPS Thedifferenttypesofnon-tileTPSare ReinforcedCarbon-Carbon(RCC)laminatedcompositematerialwereprimarilyusedforcoveringthewingleadingedgesandnosecapwherethetemperatures 23
PAGE 24
reachamaximumof1510Cduringreentry.Thiscompositecoveringhadveryhighfatigueresistancewhichisessentialduringascentandreentry.GenerallyallTPScomponents(tilesandblankets)weremountedontostructuralmaterialsthatsupportthem,mainlythealuminumframeandskinoftheorbiter.RCCistheonlyTPSmaterialthatalsoservedasasupportforpartsoftheorbitersaerodynamicshape,wingleadingedgesandthenosecap. FlexibleInsulationBlankets(FIB)werewhitelow-densitybroussilicamaterial.TheseblanketsweredevelopedasreplacementforLRSItiles.TheyrequireverylessmaintenancethanLRSItilesyettheypossessedthesamethermalproperties. FeltReusableSurfaceInsulation(FRSI)aregenerallywhite,exiblefabricofferingprotectionupto371C(700F).FRSIcoveredtheOrbiter'swinguppersurface,theupperpayloadbaydoors,andaftfuselage. Gapllerswereplacedatdoorsandmovingsurfacestominimizetheheatcreatedopengapsintheheatprotectionsystem.Thesematerialswereusedaroundtheleadingedgeoftheforwardfuselagenosecaps,windshields,sidehatch,wing,verticalstabilizer,therudder,bodyap,andheatshieldoftheshuttle'smainengines.ThellermaterialsaremadeofeitherwhiteAB312bersorblackAB312clothcovers(whichcontainaluminabers). Forfurtherindepthdiscussiononthermalprotectionconceptsusedinspacevehicles,thereaderisreferredtoBlosser[ 27 ]andBapanapalli[ 28 ]. 2.2FiniteElementbasedHomogenization Micromechanicalanalyseshavebeentraditionallyusedtoestimatetheeffectivestiffnesspropertiesofcompositematerials.Someofthesemethodsarealsosuitabletoobtainhomogenizedpropertiesforplatelikestructureswithperiodicity.Thereareseveralapproachestohomogenization:mechanicsofmaterialsapproach,elasticityapproach,energymethodsandniteelementanalysis.Allmethodsassumethatthereisarepresentativevolumeelement(RVE)orunitcellthatrepeatsitselftoformthestructure.Theunitcellissubjectedtosixlinearlyindependentdeformationstodeterminetheequivalentstiffnesspropertiesofthepanel.Thedeformationsareappliedasperiodicdisplacementboundaryconditions(PBC)whichincludesthreemid-planestrainsandthreecurvatures. 24
PAGE 25
Biancolini[ 1 ]usedanenergyequivalenceapproachtohomogenizeacorrugatedcorepaneltoanequivalentanisotropiclamina.Thehomogenizationprocedurewasappliedusingstaticcondensationinwhichtheinternalnodesofthemicrogeometryareremovedandexternalnodesattheboundaryofthemodelrepresenttheedgesoftheequivalentlamina.Theeffectivestiffnesspropertiesobtainedusingthismethodwasvalidatedwithaseriesofnumericaltests.Yuetal.[ 29 30 ]variationalasymptoticmethodforunitcellhomogenization(VAMUCH)topredicteffectivepropertiesofperiodicallyheterogeneousmaterials.Davalosetal.[ 2 ]evaluatedtheequivalentpropertiesofberreinforcedhoneycombsandwichpanels(Figure 2-2 ).Thepanelconsistedsinusoidalcoreintheplaneandextendedverticallybetweenthelaminates.Thehomogenizationtechniquewasacombinationofenergymethodandmechanicsofmaterialapproachtopredicttheequivalentpropertiesofthepanel.Forasimilarsinusoidalcorepanel,Buannicetal.[ 3 ]usedanasymptoticexpansionmethodforestimatingtheequivalentstiffnessproperties.WallachandGibson[ 4 ]usedtheunitcellapproachtocalculatethestiffnessproperties,compressivestrengthandshearstrengthofsandwichstructureshavingpyramidaltrusscores;employingtwomethods,atruss-analysisprogrambasedonmatrixmethodsandacommercialniteelementanalysis(FEA)programusingABAQUS.Theniteelementapproachconrmedresultsfromthetrussanalysesandallowedtoextendtheanalysistoincludenonlineareffectssuchasmaterialyieldandlargedeformation. PasteffortsonthehomogenizationoftheITPSpanel(Figure 2-3 )includeananalyticalapproachandniteelementbasedapproach.Martinezetal.[ 5 31 ]followedstrainenergyapproachandsheardeformableplatetheorytodevelopananalyticalmodelforthehomogenizationofacorrugatedsandwichpanel(Figure 2-3 )oftheITPS.Thoughanalyticalmodelsprovidereasonablygoodestimateofstiffnessproperties,theyinvolveseveralassumptionscompromisingtheaccuracyofthestructure.Moreovermostoftheanalyticalapproachesrequireaniteelementbasedvalidation.Sharmaetal. 25
PAGE 26
Figure2-2.Sinusoidalcoresandwichedbetweenlaminates[ 1 ] [ 32 6 ]employedaniteelementbasedhomogenizationtechniqueforhomogenizingcorrugatedcoresandwichpanelasanorthotropicplateandshowedthattheresultsagreedwellwiththethreedimensionalmodel.Henceaniteelementmethodbasedhomogenizationprocedureisadoptedheretoobtaintheequivalentplateproperties.Inniteelementbasedhomogenizationperiodicboundaryconditionsareimposedontherepresentativeunitcell(RVE)thatcorrespondstoagivenstateofmid-planestrainsandcurvaturesoftheequivalentplatetoobtaintheeffectivestiffnessproperties.TheniteelementbasedhomogenizationoftheITPSpanel(Figure 3-1 )consideredinthisstudywillbediscussedinChapter 4 Figure2-3.IntegratedThermalProtectionSystemwithcorrugatedcore[ 28 31 ] 26
PAGE 27
2.3DeterministicandReliabilitybasedoptimization Anefcientstructuraldesignplaysanimportantroleinvariousengineeringdisciplines(foreg.inautomobiledesign,aerospaceengineeringandinmarineapplications).Akeyareaofcontinuousresearchisinndinganoptimum,efcientstructure-makingastructurelightweightyethavingadequateloadcarryingproperties,isapopularexample.Recentdevelopmentinoptimizationtheoryandcomputationaltoolshavefacilitatedwaystondoptimalstructures[ 33 ].Further,designandoptimizationofcompositestructureshavegainedspecialattention[ 34 35 36 ]. Themethodologiesdiscussedinthisresearcharedeterministicoptimizationandreliabilitybasedoptimization.Deterministicoptimizationinvolvesminimizingtheweightwhileapplyingsafetyfactorsonfailureconstraints.Reliabilitybasedoptimizationincludesalltherandomnessinthedesignprocessandhasanadditionalconstraintreferredasreliabilityorprobabilityoffailureconstraint.Reliabilitybasedoptimizationhasgainedincreasingpopularitymainlyduetotheuseofcompositematerialsandthetypicaldesigndriversbeingminimumstructuralmassandhighreliability.Howeveritprovesbenecialtoperformdeterministicoptimizationinitiallyasitfacilitatesinchoosingappropriatedesignspaceforreliabilitybasedoptimization.Theoptimizationprocessinvolvesevaluationofthefailureconstraintsusingniteelementanalysesthatarecomputationallyexpensive.Thisresearchproposestoperformdeterministicandreliabilitybasedoptimizationoftheintegratedthermalprotectionsystemcompositepanelandaimsatreducingthecomputationalcostatdifferentphasesofoptimization,usingtechniquessuchasresponsesurfacemethodology,EfcientGlobalReliabilityAnalysis(EGRA),separableMonteCarlo(SMC)methodandbootstrappingmethod. Thefollowingsectionspresentssomebackgroundandreviewondeterministicoptimization,useofsurrogatemodelsforoptimization,typesofuncertaintiesincludedtoestimatereliability,methodstoevaluateprobabilityoffailure,methodstoestimateaccuracyofprobabilityoffailure.FurtheritalsodiscussesEfcientGlobalReliability 27
PAGE 28
Analysis(EGRA)foraccurateestimationoftheoptimizationconstraintswithlimitedcomputationalresources. 2.3.1DeterministicOptimization Deterministicoptimizationinvolvesoptimizingthedesignvariablestominimizetheobjectivefunctionwhilesatisfyingequalityorinequalityconstraintsthatfollowalinearornon-linearformulation.Thegeneralformulationoftheproblemis Minimizew=f(d)suchthatg(d)zh(d)=mdLB
PAGE 29
77Konthetemperatureconstraintand1.5onboththestressandbucklingconstraints.Sincethedeterministicoptimumwilllieoneitheroralloftheconstraints,itseparatesthefeasibledesignspacefromtheinfeasibleone.ThishelpsinnarrowingthedesignspacetothefeasibleonesforperformingprobabilisticoptimizationandthusavoidingexpensiveFEanalysisontheinfeasiblespace. 2.3.2StructuralReliability Reliabilitybasedoptimizationhasgrownratherquicklyduringthelastfewdecadesandstructuralreliabilitymethodshavedevelopedrapidlyandhavebeenwidelyappliedinthepracticaldesignofstructures[ 37 ].Theaimofreliabilityanalysisisthequanticationandtreatmentofuncertaintiesandthentheevaluationofameasureofsafetyorreliabilitytobeusedindesign[ 38 ].Thetermsreliabilityandprobabilityoffailurearecomplementary,inthatthemorereliablethedesign,thelowertheprobabilityoffailure.Instructuralapplications,reliabilitybasedoptimizationinvolvesminimizingweightofthesystemwhilesatisfyingareliability/probabilityoffailureconstraint.Thegeneralformulationoftheproblemis Minimizew=f(d)suchthatpf(g(d;X) Cr(2) wheretheobjectivefunctionfisafunctionofonlythedeterministicdesignvariablesd,buttheresponsefunctionintheprobabilityoffailureconstraintgisafunctionofdandX,avectorofrandomvariablesdenedbyknownprobabilitydistributions[ 39 40 ].pf;Cristhetargetprobabilityoffailure.Sinceprobabilitymeasuresarehighlynon-linear,thisconstraintisoftenreplacedbyreliabilityconstraintwherethereliabilityindexiscalculatedas 29
PAGE 30
=)]TJ /F9 11.955 Tf 9.3 0 Td[()]TJ /F4 7.97 Tf 6.59 0 Td[(1(pf)(2) whereisthecumulationdistributionfunctionofstandardnormaldistribution.Cristhetargetreliabilityindex. ReliabilityanalysescanbeperformedusingMoment-basedmethodssuchastherst-order-reliability-method(FORM)orsecond-order-reliability-method(SORM)orsamplingmethodssuchasMonteCarlosimulations(MCS).Researchershaveusedthesemethodstooptimizecompositestructuresaswell.LehetaandMansour[ 41 ]carriedoutlimitstateanalysis,rstorderreliabilityanalysisandreliability-basedstructuraloptimizationofshipstiffenedpanels.QuandHaftka[ 35 ]demonstratedreliabilitybasedoptimizationofcompositesatcryogenictemperaturesusingresponsesurfaceapproximationsandMonteCarloSimulations.ScuivaandLomario[ 42 ]performedacomparisonbetweenMonteCarlosimulationsandFORMsincalculatingthereliabilityofacompositestructure.AsignicantamountofresearchhasdealtwithoptimizationandreliabilityanalysesofthecorrugatedcoreITPSpanel(Figure 2-3 ).Kumaretal.[ 43 ]andVillanuevaetal.[ 44 ]studiedthedifferenceinriskallocationbetweenstructuralandthermalfailuresbydeterministicandprobabilisticoptimization.Sharmaetal.[ 45 ]performedamulti-delityanalysisoftheITPSpanelusingCorrectionResponseSurface.FurtherVillanuevaetal.[ 46 ]studiedtheeffectsofsinglefuturetestinthedeterministicandprobabilisticdesignoftheITPSpanel.TheyconsideredonlythethermalfailuremodeoftheITPSpanel.Matsmuraetal.[ 47 ]consideredmultiplefailuremodesoftheITPSpaneltostudytheeffectoffuturetestsinthereliabilityestimationoftheITPSpanel.ThisresearchconsidersadifferentconceptofITPSandaimsatdeterministicallyandprobabilisticallyoptimizingtheITPSpanelusingcosteffectiveadaptivesamplingtechniques. Momentbasedmethodscanbecost-efcientastheyinvolveapproximationofresponsefunctiontondthemostprobablepoint(MPP),however,theyareinaccurate. 30
PAGE 31
Samplingmethodsareaccurateifadequatesamplesareusedbutcomputationallyintensive.Inengineeringapplicationsreliabilityestimationmayinvolveindividualfailuremodesuchasstressorsystemlevelfailurewithmultiplefailuremodeswhichcouldbe,say,stressanddeection.Moment-basedmethodscouldbeusedforsystemwithsinglefailuremodes.Howeverwhentheprobleminvolveshigherdimensionswithmultiplefailuremodesthatareexpensivetoevaluate,theresponsefunctionscannotbeapproximatedasanalyticalderivatives[ 42 ].SuchsituationsareaddressedbynumericalapproximationofresponsesusingsurrogatemodelsandsamplingmethodssuchasMonteCarlosimulationstoestimatereliability. Asmentionedearlier,theITPSpanelconsideredhasmultiplefailuremodesandinvolvescalculationofsystemlevelreliability.Systemlevelfailureisdenedeitherasparallelfailureorseriesfailure.Whensystemisassumedtohavefailedifallthemodesfailitisreferredasaparallelsystemfailure,wheneitheroneofthemodesfail,itisreferredseriessystemfailure.IntheITPSproblem,ifeitherofthemodes,temperature,stressorbucklingmodesfail,thepanelisconsideredtohavefailed. 2.4UseofSurrogateModels Inordertoreducethecomputationcostassociatedwithniteelementsimulations,qualitysurrogatemodelsofthevariousconstraintsareconstructed.TheyarealternatelyreferredasResponsesurfaceapproximations(RSA)ormetamodels.Surrogatemodelsusuallyemployloworderpolynomialstothestructuralresponsewithalimitednumberofinputdesignpoints[ 48 49 ].Theinitialdesignpointsareobtainedusingdesignofexperimentssuchasfullfactorialdesign,Centralcompositedesign,Latinhypercube,A-optimalorD-optimaldesigns. Inthecaseofexpensiveresponseevaluations,surrogatemodelshavebeenusedforevaluatingconstraintsinthedeterministicandprobabilisticoptimizationprocess.Vianaetal.[ 50 51 ]havedemonstratedtheuseofsurrogatemodelsandsurrogatebaseddesignoptimization.FurthertheyalsodevelopedaSurrogateToolbox 31
PAGE 32
inMATLABthatmakestheoptimizationprocesseasier[ 52 ].ThisresearchmainlyusestheSurrogateToolboxforttingresponsesurfacemodelsandfmincon()aMATLABoptimizerfortheoptimizationprocess.Moreover,theprobabilitycalculatedfromMonteCarlosimulationoftenintroducesrandomerrorsduetolimitedsamplesizeresultinginunsafeoptimum[ 35 53 ].Theuseofresponsesurfaceapproximationreducessucherrors. Thoughsurrogatemodelsarebenecialinthecaseofcomputationallyexpensiveproblemssuchasthosedemandingcomplexniteelementanalyses(deterministicoptimization),withlargenumberofuncertaintiesandmultipleconstraints,theprobabilityoffailurecalculationsevenwithresponsesurfacemodelsaredifculttohandle.Traditionallytheresponsesurfacemodelsconstructedwereaglobalapproximationoftheresponsewithdesignofexperimentsthatareindependentoftheresponsefunction.Sincefailureisdenedusingaconstraints/limitstatesthatseparatethefeasibleandinfeasibledesigns,itisimportanttoconstructaccuratecontoursoftheconstraintswhileitisacceptabletohavelargeerrorsatotherregionsinthedesignspace.Accurateapproximationofconstraintswouldleadtoaccuratepredictionofoptimum. 2.4.1ApplicationofSurrogatestoImproveConstraintBoundaries Indeterministicandreliabilitybasedoptimization,itisimportantthattheconstraintboundariesthatseparatethefeasibledesignsfromtheinfeasibleonesareestimatedaccuratelywithminimalerror.Globalsurrogatemodelscanbeusedbutwouldrequirelargenumberofdesignpointsandwouldnotensureaccurateapproximationofconstraintboundaries.Tobuildsurrogatemodelsthatapproximatesconstraintboundaryaccurately,researchershaveintroducedvariousmethods.KuczeraandMourelatos[ 8 ]usedacombinationofglobalandlocalsurrogatemodelstorstdetectthecriticalregionsandthenobtainalocallyaccurateapproximation.Arenbecketal.[ 9 ]usedsupportvectormachineandadaptivesamplingtoapproximatefailureregions.TuandBarton[ 10 ]usedamodiedD-optimalstrategyforboundary-focused 32
PAGE 33
polynomialregression.VazquezandBect[ 12 ]proposedaniterativestrategyforaccuratecomputationofprobabilityoffailurebasedonkriging.ShanandWang[ 54 ]developedaroughsetbasedapproachtoidentifysub-regionsofthedesignspacethatareexpectedtohaveperformancevaluesequaltoatargetvalue.Forconstrainedoptimizationandreliabilityestimation,Pichenyetal.[ 13 ]developedtargetedIntegratedMeanSquarecriterion(IMSE)toconstructdesignofexperimentssuchthatthemetamodelaccuratelyapproximatethevicinityofaboundaryindesignspacewhichiseitherdenedbyatargetvalueorgaussiandistribution. 2.5AdaptiveSampling Whenoptimizationproblemsinvolvenon-linearandmultimodalobjectivefunctions,theyrequirelargenumberoffunctionsevaluationstondtheglobaloptimum.Andwhentheevaluationsarelimitedbyexpensivecomputations,ndinganaccurateglobaloptimumbecomesdifcult.Surrogatemodelscouldbeemployedbuttheiraccuracyisalsocompromisedwhencomplexfunctionsareconstructedwithfewerdesignpoints.Jonesetal.[ 55 ]developedEfcientGlobaloptimization(EGO),anunconstrainedoptimizerthatfocusesonaddingpointstothedesignspacetoaccuratelymodelcomplexfunctionsandhencendtheglobaloptimumaccurately. Inthismethod,aninitialGaussianprocessmodel[ 56 57 ]isbuiltasaglobalsurrogatefortheresponsefunction.EGOthenadaptivelyselectsadditionalsamplestobeaddedtothedesignspacetoformanewGaussianprocessmodelinsubsequentiterations.EGOusesaspecicformulationknownastheExpectedImprovementthatidentiesthenewdesignpointsbasedonhowmuchtheyareexpectedtoimprovethecurrentbestsolutiontotheoptimizationproblem.Whenthisexpectedimprovementisacceptablysmall,thegloballyoptimalsolutionhasbeenfound. EfcientGlobalOptimizationwasfurtherextendedtoimprovetargetregionsinthedesignspace.Ranjanetal.[ 11 58 ]developedasequentialdesignmethodologybasedon(EGO)thatexploresthedesignalongthecontourofinterest.Bichonetal.[ 59 ] 33
PAGE 34
alsodevelopedasimilarmethodbasedonEGOreferredasEfcientGlobalReliabilityAnalyses(EGRA).TheyillustratedtheapplicationofEGRAbyimprovinglimitstatesforaccuratereliabilityestimation.ThisresearchproposestoextendtheapplicationofEGRAtoimproveconstraintboundariesindeterministicoptimizationandreliabilitybasedoptimizationoftheITPSpanel.Theconceptofsamplingdesignsiteratively,reducesthecomputationalcostofexpensiveFEsimulationsandatthesametimeincreasestheaccuracyofthefailureboundary. 2.5.1EfcientGlobalReliabilityAnalysis Themethodhastwomainfeatures,aninitialgaussianprocessmodelandtheexpectedfeasibilityfunctiontoidentifytheadditionalsamplesiterativelyatthevicinityofthelimitstate.Theinitialsurrogatemodel^Gisconstructedusinglimitednumberofdesignsusingaknownsamplingmethod,inthiscase,Latinhypercubesampling.Theexpectedfeasibilityfunctionisoptimizedtondthenextdesignthatwouldimprovethetargetboundaryz.Thefunctionisgivenas EF(^G(X))=(G)]TJ /F9 11.955 Tf 12.67 0 Td[(z)2z)]TJ /F3 11.955 Tf 11.95 0 Td[(G G)]TJ /F9 11.955 Tf 11.95 0 Td[(z)]TJ /F2 11.955 Tf 9.75 -4.34 Td[()]TJ /F3 11.955 Tf 11.95 0 Td[(G G+z+)]TJ /F3 11.955 Tf 11.95 0 Td[(G G)]TJ /F3 11.955 Tf 9.3 0 Td[(G2z)]TJ /F3 11.955 Tf 11.96 0 Td[(G G)]TJ /F3 11.955 Tf 11.95 0 Td[(z)]TJ /F2 11.955 Tf 9.74 -4.34 Td[()]TJ /F3 11.955 Tf 11.96 0 Td[(G G+z+)]TJ /F3 11.955 Tf 11.95 0 Td[(G G+z+)]TJ /F3 11.955 Tf 11.96 0 Td[(G G)]TJ /F9 11.955 Tf 11.95 0 Td[(z+)]TJ /F3 11.955 Tf 11.95 0 Td[(G G(2) TheresponsepredictedbythegaussianmodelfollowsthedistributionN(G,G),=GistheerrorbandaroundthelimitstatewhichisafunctionofthestandarddeviationGand,afactorappliedonthestandarddeviation.z+=z+andz)]TJ /F1 11.955 Tf 10.41 -4.34 Td[(=z-.andarethestandardnormalpdf(probabilitydensityfunction)andcdf(cumulativedistributionfunction)respectively.WhenGisclosetothetargetcontour,thersttermdominatestheexpression.Thesepointsareclosetotheband.IfGisfarawayfromtarget,thesecondtermtendstodominate.Thistermfacilitatessamplinginregionsoftheinputspacewheretheestimatedresponseisoutsidetheband,buttheuncertainty 34
PAGE 35
ofpredictionishigh.Thethirdtermisrelatedtothevariabilityofthepredictedresponseintheneighborhoodof[ 11 58 ].Itsupportssamplinginregionsneartheestimatedcontourbutwherethepredictionvarianceisquitehigh. CurrentlyeithergaussianprocessorkrigingarethedefaultsurrogatemodelsforimplementingEGRA.SincetheITPSinvolvesalargenumberofrandomvariables,theuseofpolynomialresponsesurfaceisfavoredovergaussianprocessorkriging.HencetoadapttotheoptimizationoftheITPSpanelthisresearchproposestoimplementpolynomialresponsesurfaceinEGRAmethodology. 2.6UncertaintyModeling Modelingandanalysisofinputuncertaintiesaidinunderstandingthepropagationofuncertaintyintheoutputs.Uncertaintiesinanengineeringsystemcanbeclassiedasepistemicuncertaintyandaleatoryuncertaintyalsocalledvariability.Epistemicuncertaintygenerallyrepresentsalackofknowledgeofaquantityorprocessofasystemorenvironment.Thisuncertaintyisalsoreferredasreducibleuncertaintyasitcanbereduced(orincreased)fromincreasedunderstandingoftheuncertainvariableorfrommorerelevantexperimentaldata.Aleatoryuncertaintyisgenerallycharacterizedbyinherentrandomnessinthephysicalsystemorenvironmentwhichcannotbereducedbyfurtherdata[ 60 ]. Instructuralapplications,aleatoryvariabilitycanbeintroducedbymanufacturingimperfectionssuchasvariabilityinmaterialpropertiesandgeometricdimensions,variabilityinloadingandepistemicuncertaintiesbyerrorsinmodelingandsimulation.Uncertainvariablesareincludedinthedesignprocessbytheirrespectiveprobabilisticdistributions.Inadditiontothevariabilityintheinputparameters,errorinmodelingandsimulation(epistemicuncertainty)isalsointroducedintheoutputparameterstoestimatetheprobabilityoffailure.Theprobabilitydistributionoftherandomvariablescouldfollowuniform,normal,lognormalorWeibulldistribution[ 61 15 ].However,thisresearchmainlyassumesuniformandnormaldistributionfortheinputrandomvariables. 35
PAGE 36
Estimatessuchasvariance,standarddeviationorcoefcientofvariationoftheresponsesampleswouldprovidetheuncertaintyintheresponseduetoinputuncertainties. Theimportanceofuncertaintycharacterizationanduncertaintypropagationincompositeshasbeenillustratedbyvariousresearchers[ 62 63 64 65 66 ].AntonioandHoffbauer[ 62 ]studiedtheeffectsofdeviationsinmechanicalproperties,plyangles,plythicknessandappliedloadsforalaminatedshellcomposite.Theydemonstratedthatuncertaintyanalysisisveryusefulindesigninglaminatedcompositestructuresminimizingtheunavoidableeffectsofinputparameteruncertaintiesonstructuralreliability.OhandLibrescu[ 63 ]addressedtheproblemofalleviatingtheeffectsofuncertaintiesforfreevibrationofcompositecantileversunderuncertaintiessuchaslayerthickness,elasticconstantsandplyangle.Nooretal.[ 64 ]studiedthevariabilityofnon-linearresponseofstiffenedcompositepanelsduetovariationsingeometricandmaterialparametersusinghierarchicalsensitivityanalysisandfuzzysetanalysisapproach. FortheITPSoptimizationproblem,therandomvariablesaredimensions,mechanicalloads,thermalloads,thermo-mechanicalpropertiessuchasYoung'smodulus,shearmodulus,poisson'sratio,coefcientofthermalexpansion,thermalconductivityandspecicheat.Inadditiontothis,uncertaintyisconsideredintheallowablelimitssuchasmaterialstrength,temperature.Theerrorinniteelementmodelingandsimulations(transientheattransfer,staticstressandbucklinganalyses)isalsoincluded. 2.7EstimationofProbabilityofFailure TheprobabilityoffailureistheprobabilitythattherandomvariablesX=fx1,x2,...xigareinthefailureregionthatisdenedbylimitstatefunctionG(X)<0.IfthejointpdfofXisfx(X),theprobabilityoffailureisevaluatedwiththeintegral pf=PfG(X)>0g=ZG(X)<0fx(X)dx(2) 36
PAGE 37
Thereliabilityiscomputedby R=1)]TJ /F3 11.955 Tf 11.95 0 Td[(pf=PfG(X)>0g=1)]TJ /F10 11.955 Tf 11.96 16.28 Td[(ZG(X)<0fx(X)dx(2) Thedirectevaluationoftheprobabilityintegrationstateaboveisextremelydifcultwhen alargenumberofrandomvariablesareinvolved.Insuchcasestheprobabilityintegrationbecomesmultidimensionalwhichistypicalofengineeringapplications. theintegrandfx(X)isthejointpdfofXandisgenerallyanonlinearmultidimensionalfunction. thelimitstateboundaryG(X)=0isalsomultidimensional,nonlinearfunction.Instructuralapplications,G(X)isoftenablack-boxmodel(orniteelementsimulation),andtheevaluationofG(X)iscomputationallyexpensive. Hencemomentbasedmethodsandsamplingmethodscouldbeappliedtoevaluatetheprobabilityintegration. 2.7.1MomentbasedMethods MomentbasedmethodssuchasFirst-OrderReliabilityMethod(FORM)andSecond-OrderReliabilityMethod(SORM)easethecomputationaldifcultiesbysimplifyingtheintegrandfx(X)andapproximatingthelimitstatefunctionG(X)toestimateprobabilityoffailure(Equation 2 ).ThemethodsimpliesthejointdistributionfunctionbytransformingtheoriginalrandomvariablesfromX-spacetostandardnormalU-space.Thelimitstatefunctionisapproximatedusingrst-orderTaylorseriesexpansion.Inthestandardnormalspace,thepointonthelimitstatefunctionwhereG(U)=0attheminimumdistancefromtheoriginisthemostprobablepoint(MPP)offailure.ThereliabilityismeasuredasthedistancefromtheorigintotheMPP.Thismeasureisreferredasreliabilityindex().TheMPPisdeterminedas Minimize=p UTUsuchthatG(U)=0(2) 37
PAGE 38
whereUisthevectorofvariablesinstandardnormalspace.FORMisfairlyaccuratewhenthelimitstatefunctioncanbeapproximatedasalinearfunction.Secondordermethodscanbeusedwhenthelimitstatefunctionhasahigherorderofcurvature.Thismethodapproximatesthelimitstateasaquadratic,andprovidesamoreaccurateapproximationinsuchcases. 2.7.2SamplingMethods MonteCarlosimulation(MCS)isoneofthepowerfulandeasytoimplementmethodstopropagatetheuncertaintyininputrandomvariablestotheuncertaintyinfailure[ 67 ].Reliability-baseddesignofastructuralsystemisoftenaddressedusingMonteCarlosimulations[ 68 39 69 ],especiallywhenthesystemunderconsiderationfailsduetomultiplefailuremodes.ProbabilityoffailurepfisdeterminedthroughalimitstatefunctionGthatseparatesthefeasibledesignsfromtheinfeasibleones.Thelimitstateisgenerallyafunctionofrandomvariables,responseRandcapacityC.Whentheresponseexceedsthecapacity,itisconsideredfailure,example,maximumstressfailuretheorywheretheresponsewillbecalculatedstressandcapacitywouldbestrengthofthematerial.ThecapacityandresponseareassumedtobefunctionsofstatisticallyindependentrandomvariablesX1andX2,respectively.Equation 2 showstheseparablecasewherefailureoccurswhenasinglecomponentofresponseexceedsasinglecomponentofcapacity. G(X1;X2)=R(X1))]TJ /F3 11.955 Tf 11.95 0 Td[(C(X2)(2) FailureoccurswhenG0andthesystemissafewhenG<0.Inthemoregeneralcase,thecapacityandtheresponseinthelimitstatecannotbeexplicitlyseparated,andthelimitstatefunctionmayberepresentedas G(X1;X2)=G(R(X1);C(X2))(2) 38
PAGE 39
WhereRandCmaybescalarorvectorquantities.ThelimitstatecouldbeahigherorderpolynomialfunctionintermsofresponseandcapacitysuchasTsai-HillorTsai-Wucriteriongenerallyemployedtopredictfailureincompositestructures.Statisticalestimatessuchasvariance,standarddeviationandcoefcientofvariationofprobabilityoffailurewouldprovideaccuracyintheestimate. Thetraditional,crudeMonteCarlotechnique(CMC)issimple,butitlacksaccuracywhensamplingislimitedduetocomputationallyexpensivestructuralanalysis,suchasfromniteelementanalysis(FEA).TherearevarioustechniquestoimprovetheaccuracyorefciencyofCMC,includingtailmodeling,conditionalexpectation,importancesamplingandtheuseofsurrogates[ 70 71 ].However,anothertechniquereferredasseparableMonteCarlowasdevelopedbySmarsloketal.[ 72 16 ],whichisapplicableincombinationwiththeabovementionedmethodstofurtherimproveaccuracyorefciency.Whentheresponseandcapacityrandomvariablesarestatisticallyindependent,accuracycanbeimprovedbytheseparableMonteCarlomethod(SMC).Thisfacilitatesimprovedaccuracyofthecalculationoftheprobabilityoffailureforthesamecomputationalbudget. TheerrorinprobabilityoffailureestimateforcrudeMonteCarlocanbeobtainedfrombinomiallaw.ForSMCinvolvingsimplelimitstate(Equation 2 ),thevarianceestimatorwasderivedusingconditionalcalculusandvalidatedusingsimulationestimates[ 15 16 ].ThesimulationestimatesofvariancewereshowntobeofcomparableaccuracytothoseobtainedforCMC.Howeverthisisnotapplicableinthecaseofcomplexnon-separablelimitstate(Equation 2 ).ForSMCwithnon-separablelimitstates,bootstrappingtechniqueisproposed[ 18 ].Theerrorinthestandarddeviationestimateofthebootstrappedprobabilityoffailureprovidesameasureoftheaccuracyofbootstrapping.Furtheritwasalsodemonstratedthatthevariabilityestimateofcapacityandresponsecanhelpinchoosingthesamplesizeneededforgivenaccuracy[ 17 19 ].Chapter 5 onestimationofprobabilityoffailureprovidesadetaileddescriptionofthe 39
PAGE 40
CrudeMonteCarlo,SeparableMonteCarlomethod,bootstrappingtechniqueandregroupingofrandomvariablestoimproveaccuracyofpfbyapplyingtheseconceptstoacompositepressurevesselproblem. 40
PAGE 41
CHAPTER3DESIGNANDANALYSISOFINTEGRATEDTHERMALPROTECTIONSYSTEMPANEL 3.1DesignofITPSCompositePanel Theprimaryobjectiveofthisresearch,asstatedearlier,istodeterminetheoptimaldesignofanITPScompositepanel.ThischapterdiscussesthegeometryandmaterialselectioninvolvedinatypicalITPSconstruction.Further,theniteelementanalysesoftheITPSmodelisalsodiscussed. TheintegratedThermalProtectionSytempanelconsideredhereconsistsofstackedrigidAluminaEnhancedThermalBarrierfoam(AETB)insulationbarsthatarespirallywrappedwithaSiliconCarbide/SiliconCarbide(SiC/SiC)laminate.Thebarsarestackedorthogonallyintwolayersina0/90conguration.ThebarsarethensupportedbyatopfacesheetandabottomfacesheetmadeofSiC/SiCandGraphite/Epoxylaminate,respectivelyformingasandwichstructure(Figure 3-1 ).Thevariouscomponents,stackorientationandmaterialpropertiesaregiveninTables 3-2 and 3-3 [ 73 74 75 ].Theconceptnotonlyhasmaterialasymmetryaboutthemid-planebutalsogeometricasymmetryastherigidinsulationbarsarestackedorthogonally.Further,thedesigniscomplexbecauseofthelaminatewrappedaroundtheAETBinsulationbar.Itisimportanttonotethatinsuchadesign,transverseshearoftheinsulationandthewebs(wraps)wouldhaveapronouncedeffectonthestructure. 3.1.1KeyDimensionsoftheITPSstructure ThekeydimensionsoftheITPSarethewidthwf,andheighthoftheinsulationfoam,thicknessofthetopfacesheettT,bottomfacesheettBandwraptWandthenumberofbarsn(Figure 3-2 ).ThenominalvaluesofthedimensionsaregiveninTable 3-1 Table3-1.KeyDimensionsoftheITPSpanel wmmhmmtTmmtBmmtWmmn 20.020.02.02.00.512 41
PAGE 42
Figure3-1.SchematicviewoftheITPSpanel Figure3-2.DimensionsoftheITPSpanel 3.1.2ITPSMaterials ThecomponentsofITPSaremadeofceramicmatrixcomposites,polymermatrixcompositeandhighdensityinsulationfoam.ThetopfacesheetismadeofSiC/SiCplainwoventextilecompositelaminatewhichisknownforitshighmechanicalstrengthatelevatedtemperatures,highthermalstabilityandlowdensity[ 76 ].Thebottomfacesheetisgraphite/epoxylaminate(polymermatrixcomposite).Thenotablepropertiesofpolymermatrixcompositesarehightensilestrengthandstiffness,highfracturetoughnessandgoodcorrosionresistance.Moreovertheyarepopularduetotheir 42
PAGE 43
lowcostandsimplefabricationmethods.Theinsulation,madeofaluminaenhancedthermalbarrier(AETB)foamdemonstratehigherstrength,addeddurability,andhaveamaximumoperationaltemperatureof1700K[ 26 ].Thematerialpropertiesofthecomponents,theberorientationandthethicknessofeachcomponentofITPSarepresentedinTable 3-2 andTable 3-3 below. Table3-2.ComponentsofITPS,materialsandberorientation ComponentMaterialFiberorientationThickness TopfacesheetSiC/SiClaminate[(0/90)4]S16layers-0.125mmeachInsulationfoamAETB8-35mmBottomfacesheetGr/Eplaminate[(0/90)4]S16layers-0.125mmeachWrapSiC/SiClaminate[0/90]S4layers-0.125mmeach Table3-3.MaterialpropertiesofthecomponentsofITPS[ 75 74 73 ] PropertiesSiC/SiCGraphite/EpoxyAETBfoam E1GPa1701380.153E2=E3GPa12490.153v12=v130.260.30.25v230.200.3420.25G12=G13GPa566.90.062G23GPa603.580.0621x10)]TJ /F4 7.97 Tf 6.59 0 Td[(6/K40.00321.772x10)]TJ /F4 7.97 Tf 6.59 0 Td[(6/K40.1131.77 3.1.3BoundaryConditions InthepastresearchonITPSpanels,thecommonboundaryconditionsappliedontheITPSpanelweresimplysupportedboundaryconditionsandclampedboundaryconditions.Also,sincethepanelissymmetricinthexandydirections,onlyonequarterofthepanelisconsideredandsymmetryboundaryconditionswereapplied.Inthisresearchwork,simplysupportedboundaryconditionsandxedboundaryconditionsareconsideredforthehomogenizationofthepanel(Chapter4).HoweveritwasfoundthathomogenizationcouldnotbesatisfactorilyappliedtotheITPSpanel(detaildiscussioninChapter4).HenceanewsetofboundaryconditionswhichsimulatesastructurallyintegratedTPSpanelwasanalyzed.Itsimulatesthepanelmountedontheframeofthevehicleandallowedtoexpandequallyinthex)]TJ /F3 11.955 Tf 11.95 0 Td[(yplane. 43
PAGE 44
Figure3-3.BoundaryconditionsappliedontheITPSpanel 3.1.4Loads Fromtakeofftolanding,thespacevehiclesaresubjectedtovariousloadingconditionssuchasdrag,aerodynamicheating(thermalloads),in-planeinertialloads(mechanicalloads),pressureandforeignobjectimpactloads.Theseloadscouldactindividuallyatanytimeduringthemission,ortheycouldacttogetherleadingtomultiplecomponentfailure.Thisresearchworkprimarilyexploresandanalyzestheeffectofthermalloadscausedbytemperaturegradientandmismatchinthermo-mechanicalpropertiesofdifferentmaterialcombinations,andmechanicalloads(pressure)throughtransientheattransferanalyses,thermalstressanalyses,pressureanalysesandthermalandmechanicalbucklinganalyses. 3.2FiniteElementAnalyses Thissectionpresentsadescriptionoftheniteelementmodelsandanalysesforheattransfer,stressanalysisandbucklinganalysisoftheITPSpanels.Themodelconsideredinthisresearchisconstructedusingacommercialniteelementsoftware,ABAQUS(Figure 3-3 ).ABAQUS.Twenty-nodethree-dimensionalbrickelementswereusedtomodeltheITPS.The3Dsolidelementmodelhas3displacementdegreesoffreedomateachnode. 44
PAGE 45
3.2.1TransientHeatTransferAnalysis Itwasobservedthatthereentryheatingratesaremoreseverethantheheatingratesduringascent,thatis,theheatingratesincreasemoresteeplyandthetotalintegratedheatloadismuchlargerduringreentry.Thus,itcanbeinferredthatthereentryheatingrateswouldbemostinuentialintheITPSdesignand,therefore,theywereusedintheheattransferFEanalysisforthedesignprocess. AtypicalheatingrateusedfordesignisshowninFigure 3-4 .Theinitialtemperatureofthestructureisassumedas295K(72F)[ 28 ].Alargeportionofthisheatisradiatedouttotheambientbythetopsurface.TheremainingheatisconductedintotheITPS.Somepartofthisheatisconductedtothebottomfacesheetbytheinsulationmaterialandsomebythewebs. Figure3-4.HeatuxproleincidentontheITPSpanel LoadsandboundaryconditionsfortheheattransferproblemareschematicallyillustratedinFigure 3-5 .Thefourouterfacesandbottomsurfaceofthebottomfacesheetareassumedtobeperfectlyinsulated.Thisisaworstcasescenariowherethebottomfacesheettemperaturewouldrisetoamaximumasitcannotdissipatetheheat.Itisalsoassumedthatthereisnolateralheatowoutoftheunitcell.Theheatuxincidentonaunitcelliscompletelyabsorbedbythatunitcellonly,thoughinanactual 45
PAGE 46
ITPSpanelheatwouldowintothefacesheetsandwrapswhichcouldactasathermalmassandtherewouldbealateralowofheatinthepanelfromoneunitcelltoanother.TheloadstepsaretabulatedinTable 3-4 Figure3-5.Boundaryconditionsforthetransientheattransferanalysis Table3-4.Heatuxloadstepsinthetransientheattransferanalysis LoadTimeHeatFluxTimeStepAmbientStepPeriodInputSizeTemperature Step10-4500.034,069W=m230sec213KRamplinearlyStep2450-157534,06939,748W=m225sec243KRamplinearlyStep31575-217539,7480.0W=m230sec273KRamplinearlyStep42175-5175-50sec295K Therstthreeloadstepsconsideredintheheattransferanalysis,theheatuxisrampedlinearlyfrom0to34,069to39,748W=m2.Thestep4representsthetimeperiodaftertouchdownandtheFEanalysiscontinuesforanother50secondsinordertocapturethetemperatureriseofthebottomfacesheet.Duringthisperiod,alongwithradiativeheattransfer,convectiveheattransferboundaryconditionsareimposedon 46
PAGE 47
thetopsurfacetosimulatetheheattransfertothesurroundingswhilethevehicleisstandingontherunway.Themaximumbottomfacesheettemperatureisrecordedfromthisanalysis.Also,alongthereentrytimethetemperaturedistributionwithmaximumdifferencebetweenthetopfacesheetandbottomfacesheetwouldinducemaximumstressesinthepanel(Figure 3-6 ).Hencethisdistributionalongtheheightofthepanelisrecordedwhichactsasthethermalloadforthestressandbucklinganalyses. Figure3-6.TemperaturevariationonTFSandBFSwithrespecttoreentrytime 3.2.2StressAnalysis StressesdevelopedintheITPSduetothermalandpressureloadsareanalyzedindividuallybyperformingastaticstressanalysis. Inpreviousstudies[ 7 28 31 ]theITPSpanelwasassumedtobesimplysupportedalongitsedges.Itwasfoundthatthedeformationsduetothermalstresseswerenotrealisticastherewasnocompatibilityofdisplacementsbetweenadjacentpanels.InorderforanITPStobeeffectivetheyhavetobestructurallyintegrated(S-ITPS). 47
PAGE 48
Inthisconcepttheattachmentswillbedesignedsuchthatthethermalforceswillbetransferredfromonepaneltonextadjoiningpanel.ThusonecanimposeperiodicBCsbetweenpanelsaslongasthesurfaceheatingisuniform.TheproblemcanbefurthersimpliedbyanalyzingoneunitcellwithperiodicBCs.Thisisispossiblebecauseweassumeeachpaneliscomposedofrepeatingunitcells.TheboundaryconditionsofastructurallyintegratedTPSpanelisshowninFigure 3-3 .Thenodesononeofthefacesonthex)]TJ /F3 11.955 Tf 12.59 0 Td[(zplaneareallowedhaveequaldisplacementsinthey-directionandtheadjacentfaceonthey)]TJ /F3 11.955 Tf 12.83 0 Td[(zplaneisallowedtohaveequaldisplacementsinthex-direction.Thebottomedgesofthesetwoadjacentfaceswereconstrainedfrommovingintheverticaldirectionsimulatingthatthepaneledgesweremountedontheframeofthevehicle.TheothertwoadjacentedgesareassignedsymmetryBCs. Thetemperaturedistributionsareappliedalongtheheightofthepanel(z-coordinate)oneachnodeofthemodel.Thisimpliesthatthetopandbottomfacesheettemperaturesareuniformthroughoutthelengthandwidthofthepanel.Althoughthetemperaturevariesslightlyinthexandy-directions,thisvariationisverysmallandcanbeneglected.Thestressesdevelopedinthevariouscomponents(facesheets,foamandwrap)areextractedandinvestigatedforthefailureofthepanel.Thestressesinthewraparehighercomparedtotheothercomponentscontributingtowardsfailureofthepanel.Forpressureanalysis,aunitcellcannotbeanalyzedwithperiodicBCs.Hence,aquarterpanelwithsixunitcellsinthexandydirectionwasanalyzed.Apressureofoneatmosphere(1atm.)isappliedonthetopfacesheet.Itwasobservedthatstressesunderpressureloadswereatleasttwoorderslesscomparedtothethermallyinducedstresses.Sincepressureloadsdonotcausefailureofthepanel,theywillnotbeincludedintheoptimizationprocess. 3.2.3BucklingAnalysis ThecomplexdesignoftheITPSwithfacesheetsandwebsleadstolocalandglobalbucklingofthelaminatesunderthermalstressesandmechanicalstresses.Though 48
PAGE 49
thewebsinthepanelaresupportedbytheinsulationfoam,thebucklingofthewebscouldbeduetopoorbondingbetweenthefoamandthewebandalsothepossibilityofdamagedfoam(duetothermalloads)notbeingabletosupportthewebsfrombuckling.ThebucklingoftheITPScanbecarriedoutontheunitcellorasalarge3-Dshellmodelwithoutanyfoam(assumingdamagedfoam).Theseanalyseswouldprovideanestimateofthebucklingload.Whilelocalbuckling,byitself,maynotalwaysleadtocatastrophicfailure,itcouldcontributeindirectlyalongwithhightemperaturesandstresses. BucklingoftheITPSinABAQUSismodeledasaeigenvaluebucklingpredictionproblemwheretheeigenvaluecanbeconsideredastheloadfactoratwhichthestructurewouldfailduetobuckling.Theanalysisisverysimilartothestressanalysisunderthermalloadsexceptfortheloadstepwhichislinearperturbationprocedure.Theloadconsideredwouldbetemperaturedistributionifthereisthermalbucklingandpressureloading,ifthereisbucklingduetomechanicalloads. Figure3-7.BuckledITPSduetothermalloads InitiallyonlybucklingoftheITPSunitcellwithoutthefoamwasanalyzed(Figure 3-7 ).Thebucklingeigenvaluepredictedshouldbegreaterthan1.0.Eigenvalueslessthan1.0suggestthatthestructurefailsatloadslesserthantheappliedload.For 49
PAGE 50
instance,iftheeigenvalueisequalto0.7,itimpliesthatat70%oftheappliedloadthestructurewouldbuckleaccordingtothecorrespondingbucklingmode.Ifthesmallesteigenvalueisaboveunity,thenthestructurewillnotbuckleundertheappliedloads.BucklinganalysesoftheITPSunitcellwithoutfoamshowstopfacesheetbuckleswhichleadstothebucklingofthetopweb. 3.3Summary Thedifferentniteelementanalysesincorporatedintheoptimizationprocessispresentedinthischapter.Furtherformoredetailsonthevariousaspectsoftheniteelementmodelingandaforementionedstructuralanalysis,thereaderisreferredtoABAQUSdocumentation[ 77 78 ]. PastresearchincludedthestructuralanalysisoftheITPSpanelwithcorrugatedcore[ 28 ]usingshellelementsneglectingthefoam.Thisresearchconsidersacomplexdesignwithlaminateswrappedaroundadenserfoamthatcannotbeneglectedthusrequiring3-DbrickelementstomodeltheITPSpanel.Astheanalysisofthe3-Dniteelementmodeliscomputationallyintensive,thisresearchproposestohomogenizethe3-DITPSpanelasanequivalent2-Dorthotropicplate,discussedindetailinChapter 4 50
PAGE 51
CHAPTER4FINITEELEMENTBASEDHOMOGENIZATION Structuralresponses(temperature,stressandbucklingload)arecomputedusinghighdelityniteelementanalysesofthe3-DITPSpanelthatarecomputationallyexpensive.Thecomplexdesignofthepanelconsistingoffacesheets,wrapsandinsulationfoamsandthefactthatthesecomponentsaremadeofcompositematerialsmakethecomputationalexpensesprohibitivelylarge.Moreover,thepresenceoffoaminthepanelrequires3-Dniteelementstomodelthepanel.A3-Dniteelementmodel(12x12unitcells)oftheITPSconsistsof114,000nodesandastaticthermo-mechanicalanalysisofthemodeltakesapproximately40minutes.ToreducethecomputationalcostinvolvedintheanalysisofthecompositeITPSpanel,thisresearchproposeshomogenizationofa3-Dplatelikestructureintoa2-Dequivalentorthotropicplate. Thehomogenizationmethodisbasedonmicromechanicsapproach[ 79 ]wherearepresentativevolumeelement(RVE)ofthestructureissubjectedtosixlinearlyindependentdeformationstopredicttheequivalentplatestiffnessproperties.Aunitcellisusedtoestimatethein-planestiffnessproperties([ABD]matrix)anda1-DplatemodelisemployedtoestimatethetransverseshearstiffnessoftheITPSpanel.TheniteelementmodeloftheITPSismodeledusing20-nodebrickelements(C3D20R)inthecommercialniteelementsoftwareABAQUS.Theaccuracyofthemethodisinvestigatedbycomparingtheresponses(displacementsandstresses)ofthe3-DmodeloftheITPSpaneltotheequivalenthomogenizedplatemodel. 4.1A,B,Dmatrices Theunitcellissubjectedtosixlinearlyindependentdeformationstodeterminetheequivalentstiffnesspropertiesofthepanel.Thedeformationsareappliedasperiodicdisplacementboundaryconditions(Table 4-1 )whichincludesthreemid-planestrains("x0,"y0andxy0)andthreecurvatures(x,yandxy)[ 80 ]. 51
PAGE 52
Figure4-1.SchematicviewoftheITPSunitcell(RVE) Table4-1.Periodicboundaryconditionsforthesixdeformations [u(a,y)-[v(a,y)-[w(a,y)-[u(x,b)-[v(x,b)-[w(x,b)-u(0,y)]v(0,y)]w(0,y)]u(x,0)]v(x,0)]w(x,0)] "x=1a00000"y=10000b0xy=10a/20b/200x=1az0-a2/2000y=10000bz-b2/2xy=10az/2-ay/2bz/20-bx/2 Thesedeformationscreatein-planeforces(Nx,NyandNxy)andmoments(Mx,MyandMxy)intheunitcell.Theconstitutiverelationbetweenthein-planeforce,momentsandstrainsandcurvaturegivesthestiffnesspropertiesofthestructure[ 81 ].Theconstitutiverelationisgivenas 8>>>>>>>>>>>>>><>>>>>>>>>>>>>>:NxNyNxyMxMyMxy9>>>>>>>>>>>>>>=>>>>>>>>>>>>>>;=2666666666666664A11A120B11B120A12A220B12B22000A6600B66B11B120D11D120B12B220D12D22000B6600D6637777777777777758>>>>>>>>>>>>>><>>>>>>>>>>>>>>:"x0"y0xy0xyxy9>>>>>>>>>>>>>>=>>>>>>>>>>>>>>;(4) 52
PAGE 53
Forinstancewhentheunitcellissubjectedtoin-planestrain"x0inthex-direction,thefacey)]TJ /F3 11.955 Tf 10.43 0 Td[(zatx=aoftheunitcellisdisplacedbyadistanceainthex-directionrelativetothefacex=0.Oneofthecornernodesisxedtopreventrigidbodymotion.Fromthenodalforcesactingonthenodesonthefacesoftheunitcell,thein-planeforcesandmomentscanbecalculatedasfollows: Nx=1 bnnPi=1F(i)x(a;y;z)Ny=1 annPi=1F(i)y(x;b;z)Nxy=1 bnnPi=1F(i)y(a;y;z)Mx=1 bnnPi=1zF(i)x(a;y;z)My=1 annPi=1zF(i)y(x;b;z)Mxy=1 bnnPi=1zF(i)y(a;y;z)(4) Thein-planeforceswouldprovidetherstcolumnofthein-planestiffnessmatrix[A]andthemomentswouldgivetherstcolumnofthecouplingstiffness[B](Equation 4 ).Theaboveprocedureisrepeatedwiththeothermid-planestrainsandcurvaturestopopulatetheentire[ABD]matrix.Whenthisprocedureisimplementedcorrectly,thecalculated[ABD]matrixshouldbeasymmetricmatrix. Figure 4-3 showsthedeformationoftheunitcellforeachcaseofunitstrainapplied.Figure 4-3 ashowshowtheunitcellwoulddeformduetotheperiodicboundaryconditionsappliedinFigure 4-2 .Theseguresareintendedtoshowhowtheunitcellsdeformduetotheperiodicboundaryconditions. 4.2TransverseShearStiffness InthecurrentdesignoftheITPSpanel,theinsulationbarsactasthecoreofthesandwichpanelandthewrapsactastheshearwebsprovidingconsiderableshearstiffness.TheunitcellmodelwithperiodicBCscannotbeusedforcalculatingthetransverseshearstiffness.Whenatransverseshearforce,say,Qxispresent,itgives 53
PAGE 54
Figure4-2.Periodicboundaryconditionappliedontheunitcellforx=1 Figure4-3.Schematicviewoftheunitcelldeformationsduetostrainsandcurvatures risetothebendingmomentMxwhichvarieslinearlyalongthex-direction.Thustheconditionsforperiodicityareviolatedandonecannotcomeupwithboundaryconditions 54
PAGE 55
thatreectthisvariationexactly.Hence,analternatemethodwithan1-DITPSplatemodelisconsidered.Thismethodcombines3Dniteelementanalysisandbeam/platetheorytocalculatethetransverseshearstiffness.Whiletheformerisaccurate,itcarrieswithittheeffectsoftheboundaryconditionsonbothendsofthemodel.Whenthemodelislongtheendeffectswillbecomeinsignicant.Thebeamtheoryisgoodforlongpanels.Thesetwofactorstogetherrequirelongbeam/1DplateforthepresentITPSmodel. Forinstance,todeterminethetransverseshearstiffnessA55the1-Dplatemodelwithunitcellsinthex-directionisconsidered(Figure 4-4 ). Figure4-4.1-DITPSmodelwithunitcellsinx-direction Theboundaryconditionsonthe1-DITPSmodelareatx=0,u(0;y;z)=0,Bottomedgew(0;y;0)=0;atx=L,u(L;;y;z)=0,w(L;y;z))]TJ /F3 11.955 Tf 10.18 0 Td[(w(0;y;z)=C(C=Constant).Figure 4-4 showsthedisplacementboundaryconditionsandplanestrainboundaryconditionsappliedonthex)]TJ /F3 11.955 Tf 12.08 0 Td[(zplane[ 82 ].Theaimoftheanalysisistodeectthe1-Dplatemodelvertically(transversedirection)usingatipforceF.Thetotaldeectionduetobendingandsheardeformationcanbecalculatedanalyticallyas 55
PAGE 56
w(x)=Fx A55)]TJ /F3 11.955 Tf 20.82 8.09 Td[(F 2D011x3 3)]TJ /F3 11.955 Tf 13.16 8.09 Td[(Lx2 2(4) D011=D11)]TJ /F3 11.955 Tf 13.15 8.09 Td[(B211 A11(4) Thersttermissheardeformationandthesecondtermisbendingdeformation.Fisthenetverticalforceduetothedisplacementboundarycondition,Listhelengthofthe1-DpanelandD011isthereducedbendingstiffness[ 82 ].ThebendingdeectioncanbecalculatedanalyticallyusingtipforceFandthe[ABD]matrixthatisalreadydetermined.ThetotaldeectioncanbecalculatedusingniteelementanalysisandthustheshearstiffnesscanbecalculatedusingEquation 4 Transverseshearstiffnessisacrosssectionalpropertyandforahomogenousstructurewhichcanbeconsideredtobemadeofinnitenumberofunitcells,thetransverseshearstiffnessisindependentofthelength.However,theITPSpanelconsideredherehasnitenumberofunitcells.Theshearstiffnessisoverestimatedduetotheboundaryeffects.Becauseofthelimitedcomputationalresources,itisnotaffordabletondtheshearstiffnessusing1-Dniteelementplatemodelsconsistingoflargenumberofunitcells,hencethestiffnessvalueiscomputedforsmallerlengths,10,20and40unitcellsandthenaconvergencecriterionisappliedtondtheconvergedvalueofthestiffnessvalue.Theconvergenceratecanbeexpressedas k3)]TJ /F3 11.955 Tf 11.96 0 Td[(k2 k2)]TJ /F3 11.955 Tf 11.96 0 Td[(k1=N1 N2(4) wherek1,k2,k3arethevaluesofstiffnessforsayn=10,20and40respectively.N1=N2istheratioofthelengthofthepanel(ratioofnumberofunitcells,hereN1=N2=10/20=20/40=0.5).istheconvergencefactorwhichshouldbegreaterthanunityforconvergenceofthestiffness.Calculatingtheconvergencefactorfromn=10,20and40,thenusingEquation 4 ,thestiffnesscouldbepredictedforlongerpanelstoobtaina 56
PAGE 57
convergedvalueofthestiffnessproperty.SimilarprocedureiscarriedoutforestimationofA44.Theconvergedtransverseshearstiffnessalongwiththe[ABD]matrixisusedtoestimatetheaccuracyofthehomogenizedmodel. 4.3AccuracyoftheHomogenizationMethod 4.3.1DeectionComparison TheITPSpanelissubjectedtovariouscombinationsofloadswheninstalledontheexteriorofthespacevehicle.Pressure,temperature,in-planeinertialloadsandimpactloadsarethevariousloadsthatactonthepanelduringight.Inthischapter,pressureloadsareappliedtocomparetheresponse(displacementandstresses)ofthe3-DITPSpaneltothatofthehomogenizedplatetoevaluatetheaccuracyofhomogenizationprocedure.Sinceitiscomputationallyexpensivetoconstructaniteelementmodelofthe3-DITPSpanelofincreasedlengths(Figure 3-1 ),an1-Dmodelconsistingofseriesofunitcellsineither(xory)directionisconsidered.The3-DITPSmodelisxedatoneendandfreeattheotherendwithplanestrainboundaryconditionsparalleltothex-zplane.Itissubjectedtoapressureloadp=1000Pa(Figure( 4-5 )).Foranequivalent1Dorthotropicplatewithsimilarboundaryconditionsandload,aclosedformsolutionofthedeectionisavailableasshowninEquation 4 w(x)=)]TJ /F3 11.955 Tf 16.43 8.09 Td[(p A55Lx)]TJ /F3 11.955 Tf 13.15 8.09 Td[(x2 2)]TJ /F3 11.955 Tf 22.48 8.09 Td[(p 2D011L2x2 2+x4 12)]TJ /F3 11.955 Tf 13.15 8.09 Td[(Lx3 3(4) Thehomogenizedandthe3-DITPSmodel(Figure( 4-5 ))deectionsarecomparedtoevaluatetheaccuracyoftheequivalentstiffnesspropertiesestimated. 4.3.2StressComparison ThestressesarecomparedbetweentheequivalenthomogenizedplateandITPSpanelbyemployingthereversehomogenizationprocedure.TheoriginalITPSmodelissubjectedtopressureloadandstressesatvariouscrosssectionsaredetermined.Forthehomogenizedmodelsubjectedtopressureload,atanydistancex,mid-planestrainsandcurvaturescanbecalculatedas 57
PAGE 58
Figure4-5.1DmodeloftheITPSwithunitcellsinx-directionsubjectedtopressureloadandxedBCs Mx=p(L)]TJ /F3 11.955 Tf 11.96 0 Td[(x)2 2(4) x=Mx1 D011(4) "x0=)]TJ /F3 11.955 Tf 10.5 8.09 Td[(B11 A11x(4) Themid-planestrainsandcurvaturesofthehomogenizedplateareappliedtotheITPSunitcellusingtheperiodicboundaryconditions(Table 4-1 )andthestressesaredetermined.Thisisreferredasreversehomogenizingtheequivalenthomogenizedplatetothe3-DITPSmodel.ThestressesintheunitcellarecomparedtothatoftheITPSatvariouspointsalongitslength. 4.4ResultsandDiscussion 4.4.1A,B,Dmatrices-Equivalentmaterialvs.Equivalentstructure Someresearchershaveconsideredhomogenizingthesandwichpanelasaplatemadeofahomogeneousmaterial.Thiswillbeapplicablewhenthepanelissubjectedtoonlyin-planeforces.However,whentherearetransverseforcesduetosurfacepressure 58
PAGE 59
orvibrations,thebendingstiffnesshastobeconsideredandaswillbeshownbelowthe[D]matrixhastobeevaluateddirectly.Further,forthecaseofaplatemadeofhomogeneousmaterial,[B]0.Howeverbending-stretchingcouplingmaybepresentduetoasymmetryandthismayplayacrucialroleinbucklingofthepanel. Homogenizingtoequivalentmaterialbasicallysmearsthepropertiesthroughoutthepanelbytreatingitasahomogenousorthotropicmaterial.Thein-planestiffnessmatrix[A]canbecalculatedfromelasticconstantsandusingtheconstitutiverelationinEquation 4 .Fromtheclassicallaminationplatetheory,the[D]matrix(bendingstiffness)isequalto(h2/12)[A]andforisotropichomogenousmaterial,[B][0].The[A]and[D]matrixcalculatedassumingthepanelisanequivalentmaterialisgiveninEquation 4 andEquation 4 [A]eqv:matl:=2666649141061431060143106913106000241106377775(4) [D]eqv:matl:=2666641611032510302510316010300043103377775(4) However,theITPSconstructionhasdifferentmaterialsforthetopfacesheet(TFS)andbottomfacesheet(BFS)andsothe[B]matrixcannotbezero.ItisinterestingtocomparetheITPSpropertiesasequivalentstructure(Equation 4 )versussmearingthepropertiestoformanequivalentmaterial(Equation 4 )and(Equation 4 ).Ifhomogenizedasanequivalentstructure,anon-zero[B]matrixisobtainedandthebendingstiffnessseemstobetwicethatofequivalentmaterial.ThusitisimportanttonotethathomogenizingtheITPSpanelusingequivalentstructureapproachcapturesthebehaviorofsandwichstructuresaccurately(Equation 4 )whencomparedtoequivalentmaterialapproach. 59
PAGE 60
[ABD]eqv:str:=266666666666666491410614310602:11061:4106014310691310601:41065106000241106002:21062:11061:410603101035010301:410651060501033101030002:210600861033777777777777775(4) Moreover,classicallaminationtheoryassumesinniteshearstiffness(A44=A55=innity)whichwouldmaketheITPSastiff,rigidstructure,butitwillbeshowninthenextsectionthatthepanelisasheardeformablesandwichstructure.Thenextsub-sectiondiscussesthesignicanceoftransverseshearstiffnessA44andA55ofthepanel. 4.4.2TransverseShearStiffnessofITPS ThetransverseshearstiffnessA44andA55arecrosssectionalpropertiesandhencearesupposedtobeaconstantforagivenmicrostructure.Theanalysisofthe1-Dplatemodeltodeterminetransverseshearstiffnesswasvalidatedconsideringamodelmadeentirelyoforthotropicmaterialwhosetransverseshearstiffnesscouldbecalculatedanalyticallyas A55=5 6Gh(4) WhereGistheshearmodulusinthex)]TJ /F3 11.955 Tf 12.33 0 Td[(zplaneandhistheheightofthemodel.Theorthotropicmaterialconsideredwasgraphite/epoxycomposite,G=6:9GPa(Table 3-3 ).ThelengthofthemodelwasL=420mm(n=20),widthw=21mmandheighth=46mm.ThetransverseshearstiffnessoftheorthotropicITPSpanelwasestimatedandcomparedwiththeanalyticalstiffnessvalue(Equation 4 ).Ideallyforsuchapanelmadeofinnitenumberofunitcells,thestiffnessvalueisindependentofthelength,howeverduetotheboundaryconditionsappliedonthepanelends,it 60
PAGE 61
approximatelytakes0.2mtoconvergeandagreewiththeanalyticalsolution.AstheITPSpanelconsistsofnitenumberofunitcells,thiseffectwouldbemorepronouncedandwouldrequirelargenumberofunitcellsforthestiffnessvaluetostabilizetoaconvergedvalue. Figure4-6.TransverseshearstiffnessA55oforthotropicpanel NowtheabovemethodisappliedtotheITPSmodelwithvaryingmaterialconguration.Owingtothecomputationalexpense,modelconsistingofunitcellsineitherxorydirectionisconsidered.Unitcellsarrangedinthex(A11,B11andD11stiffnessesconsidered)andy-direction(A22,B22andD22stiffnessesconsidered)providesA55andA44respectively. Table4-2.VariationofA44andA55ofITPSpanelwithn nA55N/mA44N/m 1016.910620.61062010.110611.8106408.11068.8106 61
PAGE 62
Figure4-7.VariationofA44andA55ofITPSpanelwithn Figure 4-7 showsthevariationofthetransverseshearstiffnessasafunctionofnumberofunitcells.Itcanbeseenthatthetransverseshearstiffnessishighwhenfewerunitcellsareconsidered,itdecreasesasthenumberofunitcellsincreases.Sincetheniteelementanalysesoflargeplatemodelswaslimitedbytheavailablecomputationalresources,theaforementionedconvergencecriterionisusedtoestimatetheshearstiffness.UsingthestiffnessvaluesfromTable 4-2 andEquation 4 theconvergencefactorisestimated.Asmentionedearliershouldbegreaterthanunityforconvergence.Thevaluesofwerefoundtobe2.2and1.9forA55andA44respectively,whichsuggestsconvergenceofthestiffnessvalue. Figure4-8.ConvergedA44andA55ofITPSpanelwithn 62
PAGE 63
Thetransverseshearstiffnesscanbepredictedforlargerpanelsusingtheabovevalueof.Consideringk1andk2correspondingtoN=10and20respectively,thevalueofk3(N=40)canbecalculatedusingEquation 4 (hereNi=Ni+1isalwaysequalto0.5).Thususingsuccessivecalculationsstiffnessforverylargenumberofunitcellscanbeestimated.Foraverylargenumberofunitcells(500unitcells),thetransverseshearstiffnessconvergedto7.4106(A55)and7.1106(A44)alsoshowninFigure 4-7 .Itisimportanttonotethatitrequireslargenumberofunitcellsfortheshearstiffnesstoconvergetoaconstantvalue.ThissuggeststhatdeterminedshearstiffnesscannotbeusedtohomogenizesmallerpanelssuchastheITPSpanelconsistingof12x12unitcells.However,theconvergedtransverseshearstiffnessandthecalculated[ABD]matrixcouldbeusedtopredicttheminimumlength(numberofunitcells)requiredforhomogenizationmethodtoyieldacceptableresults. 4.4.3AccuracyoftheHomogenizationMethod ThetipdeectionfromobtainedniteelementanalysisoftheITPSmodeliscomparedwiththeanalyticaldeectionofthehomogenizedplate(Equation 4 )forvariouslengthsofthecantileverbeammodel.Whenthetipdeectionratiobetweenthe2-Dmodeland3-Dmodelisclosetounity,thecorrespondinglengthwouldbetheminimumsizerequiredforhomogenizationtobeapplicable. Figure4-9.Ratiobetweentipdeectionofthehomogenizedmodelandthe3-DITPSmodelalongthexandy-direction 63
PAGE 64
Table4-3.Tipdeectionratioalongwithcontributionofbendingandsheardeformationtowardsdeection L/aTipdeectionratioalongShear%Bending%x-direction(w2D=w3D) 121.367228201.124852401.051981601.01991701.00793 UsingEquation( 4 )thecontributionofsheardeformationtowardstotaldeectionwasdetermined.FortheITPSpaneltypicallyconsistingof12unitcells,thecontributionofshearis72%thusconrmingthepanelishighlysheardeformable.Thetablealsoprovidesthecontributionofsheardeformationforlongerpanels.Figure 4-9 showsthetipdeectionratioforvariouslengthsoftheITPSmodel.Thex-axisoftheplotisL=aratio(a-widthofunitcell)denotingthenumberofunitcells.ItisevidentfromFigure 4-9 thattheminimumlengthofthepanelmustbe1.5m(70unitcells)tohomogenizetheITPSstructuretoanequivalentorthotropicplate.Evenforsuchalargepanel,thereis7%sheardeformationandneglectingtransverseshearstiffnesswouldyieldin7%discrepancyinthehomogenizationofthepanelfordeections. Nowthepanelwithlargerlength(1.5m)isanalyzedunderpressureloadforstresscomparison.Thestressesinthehomogenizedmodelaredeterminedbyapplyingmid-planestrainsandcurvatures(Equation 4 and 4 )onaunitcell.Thestressratiobetweenhomogenizedandoriginal3-Dmodeliscomparedatcrosssectionsatvariousdistancesalongthelengthofthepanel(Figure 4-10 ).Itcanbenotedthatthestressratioclosetofreeendisgreaterthanunity.ThiscanbeattributedtotheSt.Venant'seffectsuggestingcertaindistanceintothemodelisrequiredforthefreeedgeeffectstodissipate.Thestressratiodecreasesandconvergestounity(approximately)atL=a=10. ThestressvaluesatL=a=10alongthewidthofthepanelisshowninFigure( 4-11 ).Furthertoshowthathomogenizationcannotbeappliedforsmallpanels,the3-DITPSandhomogenizedplatemodelsconsistingof12x12unitcellsare 64
PAGE 65
Figure4-10.Stress(11,22)ratiobetweenthehomogenizedmodelandthe3DITPSmodel Figure4-11.Stress(11,22)atn=10comparisonbetweenthe3-DITPSmodelandthehomogenizedmodel analysedunderpressureloadandxedboundaryconditions.Aquarteroftheplateisconsideredwithsymmetryboundaryconditions.Figure 4-12 showsalargediscrepancyinthedeectionsbetweenthe2-Dhomogenizedplatemodeland3-DITPSpanel(w2D=w3D=3.8). Itcanbeconcludedthatacompositesandwichpanelhastobesufcientlylargetohomogenizeittoanorthotropicplate.ForthecurrentITPSpanelwithafairlycomplexdesignandmadeofcomposites,aminimumof70unitcellsarerequired. 65
PAGE 66
Figure4-12.Verticaldeectionalongx-axisoftheoriginalITPSandhomogenizedmodelduetopressureload 4.5Summary Arepresentativevolumeelement/unitcelloftheITPSpanelwasanalyzedtodeterminetheequivalentstiffnesspropertiesofthe2Dplate.Theunitcellwassubjectedtosixlinearlyindependentdeformationsandtheequivalentstiffnesspropertiesofthecompositestructurewereestimated.Analysisofthe1-DITPSplatemodelthatcombinesniteelementsandsheardeformableplatetheorywasproposedtoestimatethetransverseshearstiffnessoftheequivalenthomogenizedplatemodel.Toevaluatetheaccuracyofthehomogenization,thehomogenized2-Dplatemodelwasanalyzedunderpressureloadandcomparedwiththe3-DITPSpanel.Thevaluableinsightsofthisresearchworkare:a)TransverseshearstiffnessofITPSsandwichpanelsareimportantandcannotbeneglected.b)Homogenizationcannotbeappliedtothesmallpanelswithfewerunitcells.HenceasingleunitcelloftheITPSpanelwillbeusedforoptimizationprocess. Thehomogenizationmethodisapplicabletopanelsoflengthmuchlargerthantheunitcelldimensions.ItwasestimatedthattohomogenizethecurrentdesignofITPSpaneltoanequivalentorthotropicplate,thepanelmusthaveaminimumof70unitcells. 66
PAGE 67
CHAPTER5MONTECARLOSIMULATIONS-RELIABILITYESTIMATION ThischapterprovidesadetaileddescriptionofthecrudeMonteCarloandseparableMonteCarlomethod,there-allocationofuncertaintytopredictaccurateestimateofprobabilityoffailureandbootstrappingtechniquetoestimatetheaccuracyofthepfestimate. ThemeritsofusingseparableMonteCarloovercrudeMonteCarlomethodisillustratedusinganon-separablelimitstatefunction(theresponseandcapacitycomponentsareintegratedinthelimitstate).TheprobleminvolvesestimatingtheprobabilityoffailureofcompositepressurevesselusingTsai-Wufailurecriterion(non-separablelimitstate).TheaccuracyofthemethodsiscomparedfordifferentgroupingsoftherandomvariablesfortheseparableMonteCarlomethodfromtheestimateofcoefcientofvariation/standarddeviation.ItisdemonstratedthatusingSMCwiththeregroupedlimitstatereducestheerrorassociatedwiththeprobabilityoffailureestimate.Thisisaccomplishedbyreallocatingtherandomvariablesasinexpensivevariableswithnolimitonsamplesizeandexpensivevariableswithlimitedsamples.Furthertheexpensivevariablesarebootstrappedandtheaccuracyofthebootstrappedprobabilityoffailureestimateiscomparedwiththatoftheempiricalestimate. 5.1CrudeMonteCarloMethod(CMC) Acommonsampling-basedmethodforcalculatingtheprobabilityoffailurepf,istraditional,crudeMonteCarlo.Theprobabilityoffailureisestimatedbycomparingpairsofrandomlygeneratedresponseandcapacitysamples,asshowninEquation 5 ^pcmc=1 NNXi=1I[G(Ri;Ci)0](5) WhereIistheindicatorfunction,whichequals1iftheconditionistrueand0iftheconditionisfalse.Thus,thisessentiallysumsthenumberoffailuresforNcomparisons. 67
PAGE 68
Therootmeansquare(RMS)errorintheestimatemaybemeasuredbythestandarddeviationforcrudeMCgivenas stdev(^pcmc)=r pf(1)]TJ /F3 11.955 Tf 11.96 0 Td[(pf) N(5) CV(^pcmc)=stdev(^pcmc) N=s (1)]TJ /F3 11.955 Tf 11.95 0 Td[(pf) pfNr 1 pfN(5) Forexample,foraprobabilityoffailureofoneinamillion,100millionsimulationsareneededfor10%error.Thecoefcientofvariationcalculatedfromstandarddeviationandmeanprovidesameasureoftherelativerootmeansquare(RMS)errorintheprobabilityoffailureestimateaboutitsmeanvalueEquation 5 .Notethatbecausewehaveonlyanestimateoftheprobabilityoffailure,wegetonlyanestimateoftheerror.ThecostofresponsecalculationisoftenthelimitingfactorinthenumberofsamplesN,becauseresponsecalculationsoftenrequireexpensiveniteelementsimulations. 5.2SeparableMonteCarlomethod(SMC) Whenthecapacityandresponsearestatisticallyindependentinthelimitstate,thentheycanbesampledseparatelyusingamethodcalledseparableMonteCarlo(SMC).SeparableMChasalreadybeeninvestigatedforthesimplelimitstateasshowninEquation 2 .ThisstudylooksatSMCforthegenerallimitstateinEquation 2 ,theseparableMCmethodisshowninEquation 5 ^pcmc=1 NNXi=1I[G(Ri;Ci)0](5) WhereNisthenumberofresponsesamplesandMisthenumberofcapacitysamples.Sincetheresponseandcapacityaresampledseparately,allofthepossiblecombinationscanbeconsideredtoestimatepf.Figure 5-1 illustratestheresultingdifferencebetweenCMCandSMC. 68
PAGE 69
ACrudeMonteCarlo BSeparableMonteCarlo Figure5-1.IllustrationofcrudeandseparableMonteCarloMethodcomparisonsA.CrudeMonteCarlomethodB.SeparableMonteCarlomethodwhereeverysampledresponseiscomparedwitheverysampledcapacity InFigure 5-1 a,thedirectone-to-onecomparisonsofcrudeMCareshownforNsamples.Whereas,Figure 5-1 bshowsthatseparableMClooksatallofthepossiblecombinationsofrandomsamples,whichmakesitinherentlymoreaccuratethanCMC.Inaddition,differentsamplesizescanbeusedtoenhancetheaccuracy,dependingontherelativecomputationalexpenseoftheresponseandcapacitywhichwasdemonstratedusinganon-separablelimitstatetopredicttheprobabilityoffailureofacompositepressurevessel(Section 5.5 ). 5.3ErrorintheProbabilityofFailureEstimate ForCMC,theRMSerrorintheprobabilityestimateisprovidedbyEquation 5 .ForseparableMCwiththesimplelimitstateasinEquation 2 ,Smarsloketal.[ 15 ]derivedanalyticalestimatesofthestandarddeviationviaexpectationcalculus.ForthemoregeneralcaseofEquation 2 ,weproposebootstrappingthecomponentsofthelimitstate[ 18 ].Forestimatingtheerrorintheprobabilityoffailureestimate,weusebootstrapping,are-samplingtechnique,whichinvolvestakingthesamplesofresponse(expensive)andre-samplingthemwithreplacement(sothatthesamplesmaycontainduplicates).Whenweperformthere-samplingbtimes,weobtainbprobabilityoffailure 69
PAGE 70
Figure5-2.Schematicrepresentationofbootstrappingwhenonlyresponseissampled. estimates^pboot.Withbestimatesof^pboot,wecanobtainthestandarddeviationdenotingtheerrorintheestimate.Aswillbeshowninthelatterpart,thebootstrappingerrorestimatesappearstobecomparabletotheCMCestimateofEquation 5 Furthermore,wecanobtain^pbootestimatesbybootstrappingcapacityatmeanvaluesofresponseandviceversa.Theknowledgeoftheindividualcontributionsoftheresponseandcapacitytowardstheuncertaintyaidsinchoosingtheappropriatesamplesizeforresponse(N)andcapacity(M)whichwouldprovideanaccurateestimateofthevariationinthepfestimate.Whenwere-sampleboththeresponseandcapacity,weobtainthetotaluncertaintystdev(^psmc;boot).Butfortheindividualcontributionsoftheresponseandcapacity,theresponsehastobootstrappedatmeancapacityandviceversatoobtainstdevR(^psmc;boot)andstdevC(^psmc;boot). Inthenumericalresults,thebootstrappingvaluesarecomparedwiththeempiricalvalues,stdevR(^psmc;boot)andstdevC(^psmc;boot)todemonstratetheaccuracyofthebootstrappingmethod.Inordertomeasuretheerrorinthebootstrappedestimate,theuncertaintyinthestandarddeviationofthebootstrappedprobabilityoffailureestimate(stdev(stdev(^psmc;boot)))isalsocalculated.Thebootstrappingestimateswouldallowus 70
PAGE 71
tojudgewhetherthenumberofexpensivesimulationsissufcientforadesiredlevelofaccuracy. 5.4SMCwithregroupingandseparablesamplingofthelimitstaterandomvariables Whenthebootstrappingestimatesshowthattheaccuracyoftheprobabilityoffailureisnotgoodenough,andthattheculpritistoofewsamplesoftheexpensivesimulation,wemayhaveapainfulchoicebetweenveryhighcomputationalcostandaccuracy.Oftenthereisanalternativetoincreasingthenumberofexpensivesamplesbyreformulatingthelimitstate,whichisdescribedinthissection.Thenumberofsamplesrequiredforaccuratemodelingoftheresponseandcapacitydependsontherelativecontributionsoftherandomcomponentsinthelimitstatefunction.Wherethelargertheuncertaintycontribution,moresamplesarerequiredforaccuraterepresentationofthedistribution.Assumingacomputationallimitonthenumberofsamplesoftheexpensiveresponse(N),itisdesirabletoreducetheuncertaintyintheresponse(i.e.obtainnarrowerdistributionoftheresponse)toachieveimprovedaccuracy. Inmoststructuralproblems,failureofthesystemdependsonthestrengthofthematerialS(e.g.capacity),andstressesthestructuresustains(e.g.response).SothelimitstateinEquation 2 maybecomeG(;S).Inlinearproblems,stresses,arealinearfunctionoftheloadp,asinEquation 5 =Up(5) WhereUarestressesperunitload.Therandomnessintheloadp,isoftenindependentoftherandomvariablesthataffectU(geometryandmaterialproperties),butpaddsuncertaintytothecomputationallyexpensivestresscalculation.SoitwouldbeadvantageoustoseparatetheloadsfromthestressesanddeterminestressesperunitloadU.Thisalsoenablesalargersamplesizeoftheload.Evenwithalimited 71
PAGE 72
numberofsamplesofstressesperunitload,theprobabilityoffailurecanbemoreaccuratelyestimatedasboththestrengthandtheloadcanbecheaplysampled.ThentheexpensiveunitloadresponseUissampledandcombinedwiththeloadinEquation 5 .Finally,thelimitstateisreformulatedas G(;S)=G(U;p;S)(5) TheprobabilityoffailurecanbedeterminedfromalargesampleofloadsandstrengthscomparedtoalimitedsampleofstresseswhichisillustratedinFigure 5-3 Awithstresses Bwithunitloadstresses Figure5-3.IllustrationofseparablesamplingwithunitloadsA.withstressesB.withunit-loadstresses TheseparableMonteCarloformulathatcorrespondstoFigure 5-1 bisshowninEquation 5 ^pUsmc=1 M1 NMXj=1NXi=1I[GU(Ui;Pj;Sj)0](5) AsimilarformofEquation 5 couldbewrittenforFigure 5-3 ,butwithdifferentindices. 5.5ApplicationtoFailureAnalysisofCompositeLaminate ThecrudeMonteCarloandseparablesamplingsimulationmethodsandtheirerrorestimatesarecomparedandillustratedbyapplyingthemtoanon-separablelimitstateproblem(Equation 2 ).Forcomplexstructures,thestressiscalculated 72
PAGE 73
Figure5-4.Compositepressurevesselwithinternalpressureof100kPaandstressesactinginasmallelementofthevessel. throughniteelementanalysisanditmaybecomputationallyexpensive.InordertoallowustoperformthousandsofMonteCarlosimulationsneededtovalidatethemethod,weselectedanexamplethatrequiresthecalculationofstressesatasinglepointusingClassicalLaminationTheory(CLT).TheprobleminvolvespredictionoffailureofcompositepressurevesselaccordingtotheTsai-Wufailurecriterion. Acompositelaminatepressurevessel(Figure 5-4 )ismadeofagraphite/epoxy[+25/-25]ssymmetriclaminatewitheachlayerbeing125mthickandissubjectedtoaninternalpressureof100kPa.ThematerialpropertiesofthecompositeareshowninTable 5-1 .Inthispaper,alloftheinputrandomvariablesareassumedtobenormallydistributed.However,theperformanceofSMCdependsonthedistributionoftheresponseandcapacity,whichisnotnecessarilynormal.Previousresearchhasusedotherdistributions,suchaslognormalanduniform,withseparableMonteCarlomethod[ 15 61 ]. Thestressesgeneratedareafunctionoftheinternalpressurepandmaterialpropertiesofthelaminatewhichareindependentofeachother. 8>>>><>>>>:12129>>>>=>>>>;=[T][Q][A])]TJ /F4 7.97 Tf 6.59 0 Td[(1=8>>>><>>>>:pd=2pd=409>>>>=>>>>;=[T][Q][A])]TJ /F4 7.97 Tf 6.59 0 Td[(18>>>><>>>>:d=2d=409>>>>=>>>>;p=8>>>><>>>>:U1U2U129>>>>=>>>>;p(5) 73
PAGE 74
Thestresses1,2(normal)and12(shear)actingineachplyofthelaminatearefunctionofin-planestiffnessmatrixofthelaminate[A],reducedstiffnessmatrixofeachlamina[Q],transformationmatrixofeachlamina[T],thepressureloadpandthediameterd(1m)ofthevesselasinEquation 5 .Failureofcompositelaminateispredictedfromthestrengthofthecompositeandstressesgeneratedusingasuitablefailurecriterion.ThemostwidelyusedcriterionforcompositesistheTsai-Wucriterion.ThecriterionisafunctionofthestrengthsS(showninTable 5-1 alongwiththeiruncertainties),normalstresses,intheberandtransversedirection(e.g.1and2direction,respectively)andshearstress,. Table5-1.Materialpropertiesanduncertaintyoftherandomvariables(normallydistributed)[ 83 ] PropertiesMeanCV%StrengthMeanCV% E1GPa159.15%S1T231210%E2GPa8.35%S1C180910%G12GPa3.35%S2T39.210%v12GPa0.2535%S2C97.210%PkPa10015%S1233.210% Accordingtothecriterion,alayerofthelaminateisassumedtohavefailedwhenthelimitstateinEquation 5 isgreaterthanorequaltozero. G(;S)=F1121+F2222+F66212+F11+F22+F1212)]TJ /F9 11.955 Tf 11.96 0 Td[(1(5) F11=1 S1TS1CF1=1 S1T)]TJ /F9 11.955 Tf 19.49 8.09 Td[(1 S1CF22=1 S2TS2CF2=1 S2T)]TJ /F9 11.955 Tf 19.49 8.09 Td[(1 S2CF12=1 S212F12=p F11F22 2(5) UncertaintiesintheTsai-Wucoefcients(F11;F22;F66;F1;F2;F12)areduetorandomnessintheunidirectionaltensile,compressive,andshearstrengthsSofthecomposite.Uncertaintyinthestressesisduetorandomnessinmaterialproperties(orU)andpressureloadp.Thein-planenormalandshearstresses(1,2,12T),are 74
PAGE 75
Figure5-5.Distributionofstressandstrengthinthe2-direction(2)showingtheprobablefailureregion uniqueineachlayerofthelaminate.Aspreviouslymentioned,thelaminateismadeof[+25/-25]spliesandtheanalysisshowsthattheinner(-25)pliesfailandthenouterpliesfail.Amongthenormalandshearstressesactingontheply,thestressinthetransversedirection2causesthefailureofthelaminate(theoverlapofthestressesandstrengths)whichcanbeseeninFigure 5-5 ThestressesareafunctionofmaterialpropertiesandinternalpressurepasinEquation 5 .Forunitpressureload(p=1),stressesareequaltoU.Therefore,theoriginallimitstatefunction(G(;S))canbereorganizedasindicatedbyEquations 5 5 and 5 5.6ResultsandDiscussion 5.6.1CrudeandSeparableMonteCarloMethod TheprobabilityoffailureofacompositepressurevesselwascalculatedusingcrudeMonteCarloandseparableMonteCarlomethod.Itwasassumedthatourcomputationalbudgetonlypermitted500stresscalculations.Therefore,forcrudeMonteCarlo,anequalnumberofrandomresponseandcapacity(S)variables(N=500)weresampledforcomparison.InthecaseofseparableMonteCarlo,theresponsesamples(N)werecomparedagainstallthecapacitysamples(M=500)resultingin250,000evaluationsofthelimitstate.Theactualprobabilityoffailureispf=0.0121(Asthisisasimpleproblem,theactualprobabilityoffailurewasestimatedbyCrudeMonteCarlomethodby 75
PAGE 76
generating5millionsamplesofstressandstrengthvalues).TherelativeerrorincrudeMonteCarlowasmeasuredbycalculatingthestandarddeviationfromEquation 5 andhencethecoefcientofvariation.Thisvalueprovidesameasureofhowaccurateistheprobabilityoffailureestimate.Forasimplelimitstate(asinEquation 2 ),theaccuracyofseparableMonteCarlocanbeestimatedasderivedbySmarsloketal.[ 16 ].Sincethisproblemisdenedbyagenerallimitstate(e.g.Tsai-Wu),bootstrappingwasperformedtoassesstheaccuracyoftheprobabilityoffailureestimateofSMCwithab=1,000bootstraprepetitions.Sincethestressesarecomputationallyexpensiveanditischeapertosamplethestrengths,randomsamplesofstresseswerebootstrapped,butthestrengthsweresampledanewratherthanbootstrapped.Thatis,ineachofthe1,000bootstraprepetitions,thestresseswerere-sampledfromthesame500samples,whilethestrengthshadafreshsampleeverytime. Inthisstudy(wheresimpleCLTcalculationswereused)theaccuracyoftheseparableMonteCarlomethodwasalsoassessedbyanempiricalcoefcientofvariationobtainedbyperformingn=10,000repetitions.TheprobabilitiesoffailureestimatesarelistedinTableandtheestimatesoftheerrorintheprobabilityoffailurearetabulatedinTable 5-3 .Itshowsthatthecoefcientofvariationisreducedfrom40%to21.0%bySMC.OnaveragethebootstrappingerrorestimateoftheSMCprobability(asmeasuredbythestandarddeviation)isabout20%low(0.002comparedto0.0025),withastandarddeviationwhichisvetimeslowerthanthataverage.Thusinthelargemajorityofcasestheerrorestimateiswithin50%oftheempiricalerror. Table5-2.EmpiricalandbootstrappingestimatesofprobabilityoffailureusingseparableandcrudeMonteCarlowithN=M=500andn=10,000repetitions CMCSMCoriginallimitstateSMCRegroupedEmpiricalBootstrappingEmpiricalBootstrappingmean(^pcmc)mean(^psmc)mean(^psmc;boot)mean(^pusmc)mean)]TJ /F9 11.955 Tf 6.46 -9.69 Td[(^pusmc;boot 0.01210.01210.01210.01200.0122 Nextwedemonstrateobtainingtheindividualcontributionoftheresponseandcapacitytotheuncertaintyintheprobabilityoffailureestimateobtainedby 76
PAGE 77
Table5-3.StandarddeviationandcoefcientofvariationofempiricalandbootstrappingpfestimatesusingseparableandcrudeMonteCarlowithN=M=500andn=10,000repetitionsfororiginallimitstate.Standarddeviationofthebootstrappingerrorisalsoshown CMCSMCoriginallimitstateSMCRegroupedEmpiricalBootstrappingstdev(^pcmc)CV(^pcmc)stdev(^psmc)CV(^psmc)mean(stdev(stdev(stdev( ^pusmc;boot)^pusmc;boot))0.004840%0.002521%0.00220.0004 bootstrappingtheresponseatmeanvaluesofthecapacity,andviceversa.Theindividualcontributionswouldhelpwhentheoverallerrorestimateislargeandneedstobereducedbyincreasedsamplesize.Thevaluesofmean(stdevR)]TJ /F9 11.955 Tf 6.46 -9.69 Td[(^pusmc;boot)andmean(stdevC)]TJ /F9 11.955 Tf 6.47 -9.68 Td[(^pusmc;boot)inTable 5-3 providetheuncertaintyinpfestimateduetothestressandstrengthrespectively.Thesevaluesarealsocomparedwithrelativecontributionsofthestressandstrengthobtainedempiricallytoillustratetheaccuracyofthebootstrappingmethod. Table5-4.Relativecontributionsofresponse(stresses)andcapacity(strengths)towardstheuncertaintyinpfthroughbootstrappingandalsocomparedwithempiricalresults EmpiricalBootstrapping stdevR(^psmc)0.0017mean(stdevR(^psmc;boot))0.0019CVR(^psmc)15.4%stdev(stdevR(^psmc;boot))0.0004stdevC(^psmc)0.0012mean(stdevC(^psmc;boot))0.0014CVC(^psmc)9.8%stdev(stdevC(^psmc;boot))0.0002 FromTable 5-4 wecanseethatthecontributionoftheresponsetotheuncertaintyinthepfestimateishigherthanthecontributionofthecapacity.Itispossibletoreducetheresponseuncertaintybyusingalargersample.However,sinceresponsecalculationisusuallyexpensive,welookinsteadtoreducetheuncertaintyintheresponsecontributionbyothermeans.Theresponsecontainstheloadwithitslargeuncertainty(CV(p)=15%).CalculatingstressperunitloadU,isolatestheexpensiveCLTcalculation(orFEA)fromthelargeuncertaintyintheload.Thenextsection 77
PAGE 78
exploreshowtherandomcomponentsinthelimitstatecanbereformulatedbyusingunitstressestoreducetheerrorinthepfestimate. 5.6.2Regroupingandseparablesamplingofthelimitstatevariablesforimprov-ingaccuracy Intheoriginallimitstate(G(,S)),thestresscalculationcontainsthelargeuncertaintyfromtheload.Therefore,rearrangingtheresponsetoobtainstressperunitloadU,andloadppermitscalculatingresponsethatdoesnotincludetheuncertaintyintheload.ThisarrangementwillenableseparablesamplingoftheindependentvariablesofthelimitstateGu(U,p,S),similartothatshowninEquation 5 .Thisregroupingshiftsthelargeuncertaintyintheloadawayfromtheexpensivestresscalculation)]TJ /F3 11.955 Tf 5.48 -9.68 Td[(U.ForN=M=500(N=numberofstressperunitloadsamples,M=numberofsamplesofstrengthandload),theuncertainty(mean(stdev()]TJ /F9 11.955 Tf 6.47 -9.68 Td[(^pusmc;boot))inthebootstrappedestimateforthereformulatedlimitstateduetotheresponse(unitloadstresses)reducestofrom0.0019to8.6x10)]TJ /F4 7.97 Tf 6.59 0 Td[(5.Ontheotherhand,thecapacity/loaduncertaintyincreasestofrom0.0014to0.0044.Itwouldappearthatwemadethesituationworse,butnowwecanreducetheerrorinpfestimatebyincreasingM,whichisnormallycheap.ThevalueofMwasvariedfrom500-10,000samplesandtheuncertaintyintheestimateforreformulatedlimitstateisshowninTable 5-5 .Itisclearlyseenthattheregroupingallowsustokeepasmallnumberofresponsecalculationsandreducetheuncertaintybyhavingaverylargenumberofinexpensivecapacity(load)calculations.Thestandarddeviationoftheprobabilityoffailurefortheregroupedlimitstatestdev)]TJ /F9 11.955 Tf 6.47 -9.68 Td[(^pUsmcwasalsoestimatedempiricallyandshowninTable 5-6 .ThestandarddeviationsobtainedareplottedinFigure 5-6 Figure 5-6 clearlyillustratestheeffectofregroupingoftheinexpensiverandomvariablesofthelimitstate.InthecrudeMonteCarlomethod,theprobabilityoffailureiscalculatedusinganequalnumberofresponseandcapacitysamples.Inthiscase,500 78
PAGE 79
capacitysamplesand500responsesampleswereused,whichcorrespondstoasinglevalueontheplotinFigure 5-6 Table5-5.StandarddeviationandcoefcientofvariationofCMCandSMCforincreasingsamplesizeofMandN=500.Bootstrappedandempiricalestimatesareshown. CMCSMC(Empirical)SMC(Bootstrapping)Mstdev(^pcmc)CV(^pcmc)stdev(^psmc)CV(^pcmc)mean(stdev(stdev(stdev(^psmc;boot)^psmc;boot) 5000.004840%0.002521%0.0020.000450000.002117.6%-*-100000.002016.8%-Table5-6.StandarddeviationandcoefcientofvariationofCMC,SMCandSMCregroupedforincreasingsamplesizeofMandN=500.Bootstrappedandempiricalestimatesareshown. CMCSMC(Empirical)SMC(Bootstrapping)Mstdev(^pcmc)CV(^pcmc)stdev)]TJ /F9 11.955 Tf 6.46 -9.69 Td[(^pUsmcCV)]TJ /F9 11.955 Tf 6.47 -9.69 Td[(^pUsmcmean(stdev(stdev(stdev(^pUsmc;boot)^pUsmc;boot) 5000.004840%0.004537.2%0.00460.000150000.001512.6%--100000.00097.9%-Incontrast,theseparableMonteCarlomethodcanusedifferentsamplesizes(M)fortherandomvariables.ObservethatthestandarddeviationfromtheoriginallimitstateofSMClevelsofftoanearlyconstantvalueof0.002forMsamplesgreaterthan5000.Ontheotherhand,thestandarddeviationfortheregroupedlimitstatecontinuallydecreaseswiththenumberofMsamplesallthewaytostdev)]TJ /F9 11.955 Tf 6.47 -9.68 Td[(^pUsmc0.0009(or7.9%CV).Inotherwords,inCMCtheerrorestimateofthefailureprobabilityis40%,buttheerrorassociatedinSMCisonly16.8%withtheoriginallimitstateor7.9%withtheregroupedlimitstate.Thatis,fornearlythesamecomputationalcost,separableMCwithregroupingcanestimatethefailureprobabilitymoreaccuratelythancrudeMonteCarlo. Figure 5-6 showsthatthemagnitudeoftheuncertaintyreduceswithincreaseinnumberofMsamples.Thusbyincreasingthesamplesizeoftheinexpensive 79
PAGE 80
Figure5-6.StandardDeviationofCMC,SMCandregroupedlimitstateSMCwhereN=500(xed)andMisvaryingfor10,000repetitions components(strengthandtheload),wecouldreducetheuncertaintyinthepfestimate.ForverylargeM,itwouldreachanasymptoticvalueduetothenitevalueofN.Figure 5-6 showsthatbytransferringsomeoftheuncertaintyfromtheresponsetothecapacity,wecantakeadvantageofincreasedMtofurtherreducetheerrorintheprobabilityestimate. 5.6.3Summary TheseparableMonteCarlo(SMC)methodcanprovidesubstantialimprovementsinaccuracyoverthecrudeMonteCarlomethodwhenresponseandcapacityaregovernedbyindependentrandomvariables.ObtainingestimatesoftheaccuracyofSMCiscriticaltotakingfulladvantageofthemethod.HereweproposedusingbootstrappinginordertoobtainestimatesoftheerrorintheSMCestimates,aswellasthecontributionsofthecapacityandtheresponsetothaterror.TheapproachwasdemonstratedthroughanexampleproblemoffailureanalysisofacompositepressurevesselusingTsai-Wufailurecriterion.Becauseofthelowcomputationalcostoftheexample,itwaspossibletoconductmultiplesimulationsandassesstheaccuracyofthebootstrappedestimateempirically.SeparableMonteCarloledtosubstantialaccuracyimprovementindeterminingtheprobabilityoffailurecomparedtocrudeMonteCarlo 80
PAGE 81
method.BootstrappingprovidedreasonableestimatesoftheuncertaintyintheSMCprobabilityoffailureestimates.Bootstrappingalsoallowedestimatingtheindividualcontributionsoftheresponseandcapacitytowardstheuncertaintyintheprobabilityoffailureestimate,thussuggestingadditionalsamplesofinexpensivecapacitywouldproveadvantageous.Furthersubstantialimprovementinaccuracywasachievedbytransferringuncertaintyawayfromexpensivecalculationsbyusingunitloadstressesandgeneratinglargesamplesofloadandstrengths. 81
PAGE 82
CHAPTER6EFFICIENTGLOBALRELIABILITYANALYSIS(EGRA) ThenameofthemethodincludesReliabilityAnalysis,howeverEGRAisasequentialsamplingmethodthataimstoimprovesurrogatesinvolvedinconstrainedoptimizationandreliabilityanalysis.Thismethodcanbeusedimprovethetargetboundariesinsurrogatemodelsfacilitatingestimationaccurateoptimum(orreliability).TheEGRAalgorithmcodedinthesurrogatetoolboxinMATLABwasusedtoillustratetheadvantagesofusingthismethod. TheITPSoptimizationprobleminvolvesconstructionofsurrogatesinfourormoredimensions,wherevisualizingtheworkingofEGRAisnotpossible.ThischapterillustratesthemethodologyofEGRAusingBranin-Hoofunctionintwodimensionalspacefortheeaseofvisualization.Asmentionedinsection 2.5.1 EGRAhastwoimportantfeatures-aninitialresponsesurfaceconstructedwithfewerdesignpointsandtheExpectedFeasibilityfunctioncriterion.Withagiventargetlevel,thefeasibilityfunctionusestheexistingresponsesurfacetoidentifyanewdesignpointthatwouldimprovethecontourofthetargetboundary. 6.1EGRAalgorithm Bichonetal.[ 59 14 ]hasprovidedadetaildescriptionoftheEGRAalgorithm. 1. Generateasmallnumberofrandomsamplesfromthetrueresponsefunction.(a)Thisinitialselectionisarbitrary.However,oneoftheoptionsareconsideringnumberofsamplesrequiredtodeneaquadraticpolynomialisusedhereasaconvenientruleofthumb((n+1)(n+2)/2sampleswherenisthenumberofrandomvariables).(b)Latinhypercubesampling(LHS)isusedtogeneratethesamplesinthedesignspace. 2. BuildaninitialGaussianprocessmodel(interchangeablyreferredaskrigingmodel)fromthegeneratedsamples. 3. Findthepointwithmaximumexpectedfeasibility.(a)Theexpectedfeasibilityfunctionisbuiltwith=2G.(b)Tolocatethemaximumexpectedfeasibilityanappropriateoptimizerisused.ThesurrogatetoolboxusesDifferentialEvolutionaryOptimizer[ 84 ]. 4. Evaluatethetrueresponsefunctionatthispoint 82
PAGE 83
5. Addthisnewsampletotheprevioussetoftrainingdataandbuildanewgaussianprocessmodel.Gotostep3. 6. Thisupdatedsurrogatemodelisthenusedanapproximationtoevaluateconstraintsinconstrainedoptimization 6.2IllustrationofEGRA TheBranin-Hoofunctionisastandardtestfunctionusedinoptimizationproblems.Thefunctionhastwovariableswithbounds-5x110and0x215andadequatelynon-linear.Thefunctionisdenedas g(x)=x2)]TJ /F9 11.955 Tf 13.15 8.08 Td[(5:1x21 42+5x1 )]TJ /F9 11.955 Tf 11.95 0 Td[(62+101)]TJ /F9 11.955 Tf 16.68 8.08 Td[(1 8cos(x1)+1050(6) ContoursofthetruefunctionisshowninFigure 6-1 A.Thecontourg(x)=50ischosenasthetargetboundary(Figure 6-1 B).ThegoalistorepresentthisboundaryasaccuratelyaspossibleusingtheEGRAmethodology.ThetruecontoursoftheBranin-HoofunctionareshowninFigure 6-1 A.Alsothetargetboundaryg(x)=50isshowninFigure 6-1 B. ATrueFunction Bg(x)=50ofTruefunction Figure6-1.TruecontoursoftheBraninhoofunctionA.allcontoursB.targetcontourg(x)=50 83
PAGE 84
Aninitialsurrogatemodelisbuilt.Thesurrogatemodelavailablebydefaultiseitherkrigingorgaussianprocessmodel.Theinitialsurrogatemodelwastted10LHSdesignofexperimentsasshowninFigure 6-2 A[ 85 ].Usingtheresponsesurfaceapproximationandthepredictionuncertainty,theexpectedfeasibilityimprovementfunction(Equation 2 )ismaximizedusingtheDifferentialEvolutionaryoptimizerinthesurrogatetoolbox.Figure 6-3 showsexpectedfeasibilityfunctioncomputedandplottedinthedesignspaceforthepurposeofvisualizingtheprocess.Itcanbeseenthatthefunctionhasamaximumatupperleftcornermarkedbyacircle. Figure6-2.TargetcontourapproximatedbytheinitialkrigingmodeloftheBraninFunction.TotalDOE=10points Theoptimizerfoundthemaximumtobeattheupperleftcorner(0,15)andthisdesignisevaluatedusingthetrueresponsefunctionandincludedintheDOEtoconstructthenewresponsesurface.ThenewdesignandthetargetcontourofthenewthekrigingmodelisshowninFigure 6-4 Theexpectedfeasibilityfunctionofthenewsurrogatemodelshowsthattheoptimizerwouldsuggestapointatlowerrightcorner(markedbyacircle)wherethe 84
PAGE 85
Figure6-3.ExpectedFeasibilityFunctionevaluatedusingtheinitialkrigingapproximation functionismaximum.ThenewdesignpointacquiredattheendofthisEGRAcycleis(10,0).TheexistingsurrogatemodelisupdatedwiththisdesignandthenewmodelisshowninFigure 6-5 Thisprocessisrepeatedforanother18cyclestoimprovethetargetboundaryg(x)=50.Thedifferentialevolutionaryoptimizer[ 84 ]maximizesthefeasibilityfunctiontondtheadditionaldesigns,onepercycle.ItshouldbenotedthatLHSsampling,oftentimesdonotplacepointsattheboundariesofthedesignspace.Thiscauseshighuncertaintyinthepredictionmodel.Theexpectedfeasibilityfunctioninitiallyaddspointsatregionswherepredictionuncertaintydominates.Thisisreferredtoasexplorationofthedesignspace.Whenthepredictionuncertaintyinthemodelisreducedduetotheadditionalpoints,theexpectedfeasibilityfunctionimprovesthecontourbyplacingpointsnearthetargetboundary(Figure 6-6 ).Thisisreferredasexploitation.Thusbalancingtheexplorationandexploitationthissamplingmethodsuccessfullyapproximates 85
PAGE 86
AKrigingmodelupdatedbyEGRADOE BExpectedFeasibilityFunction Figure6-4.Krigingmodelupdatedwiththenewdesignandexpectedfeasibilityfunction.DesignupdatedbyEGRADOEisshowninRed.TotalDOE=11points Figure6-5.TargetcontourapproximatedbytheinitialkrigingmodeloftheBraninFunction.TotalDOE=12points 86
PAGE 87
Figure6-6.TargetcontouraccuratelyapproximatedbytheEGRAmethodology.InitialDOE=10,EGRAupdatedDOE=20 Figure6-7.Krigingsurrogatemodelconstructedusingglobaldesignofexperiments.TotalDOE=30 87
PAGE 88
thetargetboundarieswithfewernumberofdesignpointsthanrequiredbyaglobalsurrogatewiththesameaccuracy. Thenalkrigingmodelwith10initialdesignsand20designsupdatedbyEGRAinshowninFigure 6-6 .Itcanbeseenthereismuchlessdiscrepancybetweenthetrueresponseandtheupdatedsurrogateresponsethusdemonstratingtheefciencyofthemethodology. Thisupdatedsurrogateisfurthercomparedwithasurrogateconstructedfrom30designofexperimentsdistributedglobally.ThedesignsarealsocreatedusingLHSformulation.TheglobalsurrogateshowninFigure 6-7 isgrosslyinaccurateincomparisonwithtrueresponsefunction.WiththesamenumberofpointsEGRAefcientlyprovidedanaccurateapproximationoftheresponse. ThismethodwasincorporatedtoestimatethedeterministicoptimumandprobabilisticoptimumoftheITPSpanel.TheITPSoptimizationproblemseekspolynomialsurrogatesasitinvolveshighdimensions.ThisresearchworkincludedpolynomialresponseapproximationintheexistingEGRAalgorithm. 6.3Summary BoundariesofcomplexfunctionscanbeapproximatedefcientlyusingEfcientGlobalReliabilityAnalysis.ItwasshownthatatargetboundaryoftheBranin-Hoofunctionwasaccuratelyapproximatedusing10initialLHSdesignsand20additionalpointsacquiredusingEGRA.Theimprovementcriterionbalancesbetweenexploitingregionsofthedesignspace(wheregoodsolutionshavebeendiscovered)andexploringregionsthathavenotbeenwellsampled(andthushavegreateruncertainty)tosuccessfullyapproximatetargetregions.EGRAreducesthecomputationalcostwhendealingwithevaluationofresponsefunctionthroughexpensiveniteelementanalyses.Itefcientlyfacilitatesaccurateoptimumbyaccuratelyapproximatingthetargetboundariesofconstraintswhileminimizingthecomputationalcostinvolved. 88
PAGE 89
FormoreexamplesanddetaileddescriptionofvariousapplicationsoftheEGRAmethodology,thereadersarereferredtoBichonetal.[ 59 86 14 ]. 89
PAGE 90
CHAPTER7DETERMINISTICANDRELIABILITYBASEDOPTIMIZATION ThischapterpresentsthedeterministicandprobabilisticoptimizationprocessoftheITPSpanel.Thecomputationalexpensesassociatedwiththehighdelityniteelementanalyseswasreplacedbyefcientsurrogatemodels.FurthertheaccuracyofthemodelsattargetregionswereimprovedusingEfcientGlobalReliabilityAnalysis. Asurrogateofthefailureresponsewasconstructedwithfewdesignpoints(eg.numberofdesignsequaltonumberofcoefcientswhenthesurrogatewasapolynomialapproximation).Withatargetlevelgiven,EfcientGlobalReliabilityAnalysis(EGRA)usedtheexistingresponseapproximationandtheexpectedfeasibilityfunctiontoidentifythenewdesignpointthatimprovedtheconstraintboundary(temperature,stressandbucklingloadindeterministicoptimizationandreliabilityindexconstraintintheprobabilisticoptimization).Further,eachdesigngivenbyEGRAwereinputtotheniteelementanalysesandtheresponseswereevaluated.Thiswasaddedtotheexistingdesignofexperimentstoconstructanupdatedsurrogate.Thenewdesignswereaddedsequentiallyandthesurrogatemodelwasupdatedlikewise.Theupdatedsurrogatemodelimprovedtheboundaryofthetargetconstraint.Thetotalnumberofnewdesignsrequiredtoapproximateatargetboundarygenerallydependsonthecomplexityofthefunctionandtheinitialresponsesurface.Thismethodologyimprovedtargetboundariesoftheconstraintsbothindeterministicandprobabilisticoptimization. FurtherthischapterpresentssystemreliabilityofITPSfoundthroughseparableMonteCarlomethodandtheerrorintheindividualprobabilitiesoffailureestimatedusingthebootstrappingmethod. 7.1DeterministicOptimization AsarststepinoptimizingthedesignoftheITPS,adeterministicoptimizationofthepanelisdesired.Thisresultedinanappropriatenarroweddowndesignspaceandprovedtobeausefulpointofreferencefortheprobabilisticoptimization.Furtheritwould 90
PAGE 91
beinterestingtocomparetheminimizedstructuralmassbetweenthedeterministicandtheprobabilisticoptimum.Thedesignvariables-thegeometryparameters,tW,tB,tTandh(Figure 3-2 )wereoptimizedforminimummassoftheITPSunitcell.Theformulationoftheoptimizationproblemis Minimizem=f(d)=f(tW;tB;tT;h)suchthatTmax;BFSallowLBdUB(7) wheremisthestructuralmass,Tmax;BFSisthemaximumbottomfacesheettemperature,=1,2,12arethethermalstressesintheITPScomponentsandisthethermalbucklingload.Thefailurethresholds,theallowablevaluesandthesafetyfactorsforeachconstraintisgiveninTable 7-1 .Thesafetyfactorsandmarginscanbeconsideredastheriskallocatedforeachmodeoffailure. Table7-1.Failure,safetyfactorsappliedonconstraints ConstraintsFailureSafetyFactorAllowabletargetconstraints Tallow550K77K(safetymargin)473KWrapstrengthS1T323MPa1.5215.3MPaBucklingloadallow1.01.51.5 Asmentionedinsection 3.2.2 ,sincepressureloadsdonotcausefailure,theITPSisanalyzedonlyfortemperatureloads.Itwasfoundthatofallthecomponents,thewrapstresseswerecriticaland1;Wofthewrapcausedfailureofthepanel.Henceonlythewrapstressesareincludedasconstraintsintheoptimizationofthepanel.TheboundsofthegeometricdesignvariablesfordeterministicoptimizationaregiveninTable 7-2 91
PAGE 92
Table7-2.Lowerandupperboundsofdesignvariablesfordeterministicoptimization DesignVariableLowerBound(LB)UpperBound(UB) Wrapthickness,tWmm0.41.0Bottomfacesheetthickness,tBmm2.06.0Topfacesheetthickness,tTmm1.03.0Heightofthefoam,hmm25.040.0 Withtheresponseapproximationsconstructedforeachfailureconstraint,theseapproximationswereusedalongwithMATLABoptimizerfmincon()toevaluatetheconstraintsatvariousiterationsfordeterminingtheoptimaldesignoftheITPS. Traditionallylargenumberofdesignswouldbeusedforconstructingglobalresponsesurfaceoftheconstraints.Thenumberofdesignswouldgenerallydependonthenumberofdesignvariables,thedesiredpolynomialapproximationandthenumberofsimulationsaffordable.Forexample,consideringasecondorderpolynomialapproximation,fora4designvariableproblem,thereare15polynomialcoefcientsandaminimumof30designswouldberequired.Generallytoensureaccuracyofthesurrogate,morethan30simulationscouldbeconsidered. Theoptimumislikelytofallonsomeoftheconstraintboundaries.Insteadofusingalargenumberofdesignofexperiments(DOE)toconstructglobalresponseapproximations,theavailablecomputationalresourcesweretargetedtowardsregionsinthedesignspacetoapproximatetheconstraintboundariesaccurately.Thiswasaccomplishedbyupdatingthesurrogatesusingadaptivesamplingtechnique,EfcientGlobalReliabilityAnalysis. EGRAfacilitatedestimationofanaccurateoptimumwhilereducingthecomputationalcostassociatedwiththeoptimizationprocess.SincethepolynomialsurrogatesweremoresuitableforthehighdimensionalITPSproblem,polynomialresponsesurfaceswereincorporatedintheEGRAalgorithminthesurrogatetoolbox.TheoptimumdeterminedusingtheEGRAupdatedsurrogatesiscomparedwiththeoptimumdeterminedusingglobalsurrogates.ItwasfoundthatusingEGRAupdatedsurrogates,theaccurateoptimumcouldbeobtainedwithone-thirdnumberofdesignpointsrequired 92
PAGE 93
tobuildglobalsurrogates.Furthermoreitwasfoundthatevenwithlargenumberofdesigns,theoptimumevaluatedusingglobalsurrogateswasinaccurate. Further,itwasdesiredtopredictthesystemreliabilityofthedeterministicoptimum.Thiswouldserveasaconstraintfortheprobabilisticoptimization.Topredictthesystemreliability,uncertaintymodelinganduncertaintyanalysiswascarriedout.Theuncertaintyintheresponsesduetotheuncertaintyininputrandomvariableswouldaidinselectingappropriatesamplesizefortherandomvariablesleadingtoaccuratereliabilityestimation. 7.2UncertaintyModeling ThealeatoryvariabilityconsideredintheITPSdesignaregeometry,materialproperties(Young'smodulus,Poisson'sratio,shearmodulus,thermalexpansioncoefcient,thermalconductivityandspecicheat)andloads.Theepistemicuncertaintiesincludedaretheerrorinmodelingandsimulationoftheniteelementanalyses. EachcomponentoftheITPSismadeofcompositematerial(Table 3-2 )increasingthenumberuncertaintiesdrastically.ForthecompositeITPSpanel,atotalof33uncertainmaterialpropertiesvariablesareconsidered.ThevariabilityintheinputparametersandthecapacityvariablesareshowninTables 7-3 and 7-4 .Inadditiontothevariabilityintheinputparameters,errorinmodelingandsimulationisalsointroducedintheresponsestoestimatetheprobabilityoffailure.Theerrorinheattransfereth,staticstressestrandthermalbucklinganalysisebkwasassumedtobe3.0%,uniformlydistributed. Table7-3.CoefcientofvariationofinputrandomvariablesincludedintheITPSdesign RandomNo.ofUncertaintyProbabilityVariablesvariables%distribution WrapthicknesstW16.25NormalBFSthicknesstB16.25NormalTFSthicknesstT16.25NormalHeightoffoamh13NormalMaterialProperties335NormalLoad,heatuxq1;q2215Normal 93
PAGE 94
Table7-4.Nominalallowablevaluesofcapacityandcoefcientofvariation CapacityNominalvalueCoefcientofTypeofvaluevariation%distribution WrapStrengthS1T323MPa3%NormalAllowableTmax;BFS55015KUniformAllowable1.0-ThetotalnumberofinputuncertainvariablesintheITPSdesignis39.Thevariablesareheatuxloads:q1,q2,geometry:tW;tB;tT;h,thermalpropertiesofthecomponents:thermalconductivitykandspecicheatCpofWrap,materialpropertiesofBFS:E1;E2;v12;nu23;G12;G23;1;2,foam:E;v;,TFS:E1;E3;v12;nu13;G12;G13;andWrap:E1;E3;v12;nu13;G12;G13;.Althoughthereare39uncertaintiesinthedesignofITPS,theresponseswouldbenotsensitivetoalltheuncertainvariables.Henceasimplesensitivityanalyseswasperformedusinglinearpolynomialsurrogates.ApolynomialsurrogateformaximumBFStemperature,stressandbucklingwasconstructedasfunctionofthe39variablestondthemostsensitivevariables.FurtherthemostsensitivevariableswereusedtoestimatesystemreliabilityoftheITPSpanelusinglinearsurrogatesofthecorrespondingresponses. 7.3EstimationofSystemReliability Thesystemreliabilityisthecombinedprobabilityoffailureduetotheaforementionedfailureconstraints(temperature,stressandbuckling).Itisaknownfactthatlargesamplesofresponsesandcapacityarerequiredtoestimatetheprobabilityoffailureaccurately.ThiswasaccomplishedbyconstructingresponsesurfaceswithlimitednumberofFEsimulations.Thuslargenumberofresponsesampleswereevaluatedwithminimumcomputationalexpense.Failureisdenedbythelimitstatefunctionthatseparatesthefeasibledesignsfromtheinfeasibleones.Limitstatefunctionforstresses,thermalandbucklingfailureareshowninEquation 7 .Thelimitstatesinthisproblemaredenedasdifferencebetweentherandomresponseandrandomcapacity(Equation 2 ). 94
PAGE 95
Gth=Tmax;BFS)]TJ /F3 11.955 Tf 11.96 0 Td[(TallowGstr=11;W)]TJ /F3 11.955 Tf 11.96 0 Td[(S1T;WGbk=allow)]TJ /F3 11.955 Tf 11.95 0 Td[((7) Whenevertheresponseexceedsthecapacitythepanelisassumedtohavefailed.Systemreliabilityofthepaneliscomputedwhenthepanelfailsineitherofthemodes.Sincethecapacitiesareindependentoftheresponses,andfollowagivenstatisticaldistribution,thecapacitiesandresponsescouldhavedifferentsamplesizeandallpossiblecombinationscouldbeconsidered.ConsideringNsamplesofresponseandMsamplesofcapacity,theprobabilityoffailurepfcouldbeestimatedusingseparableMonteCarlomethodas ^pth=1 NMMXj=1NXi=1I[Gth(Cj;Ri)>0]^pstr=1 NMMXj=1NXi=1I[Gstr(Cj;Ri)>0]^pbk=1 NNXi=1I[Gstr(C;Ri)>0](7) Thecapacitiesforthermalandstressfailuresfollowadistribution(Table 7-4 ),resultinginNMcomparisonstoestimateprobabilityoffailure.Thecapacityforbucklingloadfactorisdeterministic(allow=1.0)andhenceresultedinNcomparisons.Thissuggestedthatthebucklingprobabilityoffailure^pbkshouldbecalculatedwithlargesamplesizesofbucklingloadresponses. Furthertheniteelementmodelingerrorwasintegratedtotheresponseforeachcaseas Gth=Tmax;BFS(1+eth))]TJ /F3 11.955 Tf 11.95 0 Td[(TallowGstr=(1+estr))]TJ /F3 11.955 Tf 11.95 0 Td[(SGbk=allow)]TJ /F3 11.955 Tf 11.95 0 Td[((1+ebk)(7) 95
PAGE 96
Sincetheerrorsarealsoindependentoftheresponse,theycouldbeassignedadifferentsamplesize(Q).Foreacherror,theNsamplesoftheresponseandMsamplesofthecapacityaresampledtoestimateasinglevalueofprobabilityoffailure.QerrorvalueswouldprovideQprobabilityoffailureestimates.ThemeanoftheQestimatesisconsideredasthemeanprobabilityoffailureforeachfailuremode.Thesumoftheprobabilitiesoffailureofthethreemodesisconsideredasthesystemreliabilityofthepanel. Furthertheerrorintheindividualprobabilitesoffailurewasestimatedusingthebootstrappingtechnique,describedindetailinSection 5.3 .Furthertheerrorsestimatedwouldbecomparedwiththeerrorsestimatedfromnumerousrepetitionstovalidatethebootstrappingmethod. 7.4ReliabilityBasedOptimization Thereliabilitybasedoptimizationinvolvesoneadditionalconstraint,theprobabilityoffailureconstraint.Theprobabilityoffailureofthedeterministicoptimumactsastheconstraintfortheprobabilisticoptimization.TheoptimizationproblemoftheITPSpanelisformulatedas Minimizem=f(d)=f(tW;tB;tT;h)suchthat(g(d;X)Cr(7) whereiscalculatedusingEquation 2 .Itdependsonthefailureconstraintsg(d;X)evaluatedasafunctionofdesignandrandomvariables.Cristhedesiredlevelofreliabilityfortheoptimum.ThisvaluewasassignedthedeterministicreliabilityindextoanalyzehowthefailureconstraintsdifferfromthedeterministicoptimizationandwhetherthishasaidedinfurtherreducingorincreasingthemassoftheITPSpanel.Furtherthisanalysisontheriskallocationmightsuggestwhatsafetyfactorscouldbesetforthedeterministicoptimizationthatwouldfurtherminimizethemassleadingtononconservative,safeoptimumdesign. 96
PAGE 97
Thereliabilitybasedoptimizationrequiredcalculationofreliabilityindexinthedesignvariablespace.Hencearesponsesurfaceofthereliabilityindexasafunctiondesignvariableswasconstructed.Inthedesignvariablespace,ateachdesignsetd,probabilityoffailureandhencethereliabilityindexwascomputed.Thisinvolvedconstructionoffailureresponseapproximationsateachdesignpoint.Usingtheresponsesurfaces,therandomvariablesweresampledaccordingtotheirrespectiveprobabilitydistributiontoestimatetheprobabilityoffailure.Henceateachdesignpoint,adesignofexperimentswascreatedintherandomvariablespace,failureresponsesurfaceswereconstructedandsamplesoftherandomvariablesweregeneratedtoestimatetheprobabilityoffailure.Thisprocesswascarriedoutatdesignpointsinthedesignvariablespacetoobtainareliabilityindexresponsesurfacemakingthisprocesscomputationallyintractable. Sincetheprobabilisticoptimizationinvolvessatisfyingthereliabilityconstraint,EfcientGlobalReliabilityAnalysiswasusedtoimprovetheaccuracyofthesurrogateatthedesiredreliabilitylevelwithfewernumberofdesignpoints.Initiallyasecondorderpolynomialsurfaceofthereliabilityindexwasconstructed.HoweverafteraddingadditionaldesignpointsusingEGRA,thetargetboundarywasinaccurate.Itwasfoundthatthereliabilityindexresponsewashighlynon-linearandsecondorderpolynomialwasapoorapproximationoftheresponse.Theorderofthepolynomialwaslimitedbytheavailablecomputationalresources.ThusthereliabilityindexwasttedusingkrigingsurrogatemodelwhichapproximatedtheresponseandthenEGRAwasusedtoimprovetheboundaryofdesiredreliabilitylevel.ThispreventedwastingoflargeofFEsimulationsatunimportantregionsandfocusthecomputationalresourcesinapproximatingthedesiredreliabilitylevelaccuratelytodeterminetheaccurateprobabilisticoptimum. 97
PAGE 98
7.5ResultsandDiscussion 7.5.1DeterministicOptimization ThedesignvariablesforthedeterministicoptimizationwerethedimensionsoftheITPS.Thepanelwasoptimizedforminimummasssatisfyingthermal,structuralandbucklingconstraints.ThemaximumBFStemperatureandthermalbucklingloadfactorwerethecriticalresponsesconsideredasfailureconstraints.Bucklinganalysesoftheunitcellshowedthattopfacesheetwasthecomponentthatbuckledleadingtobucklingofthewebs.ThermalstressanalysisoftheITPSunitcellshowedthatthewrapstresseswerecriticalandwrapstress1;Wwasalsooneoftheactiveconstraintsintheoptimizationprocess. Astudyonthevariationoftheresponseswithrespecttothedesignvariableswasperformedtondtheorderofthepolynomialthatwouldcloselyapproximatetheresponsewiththeresultsfromtheniteelementanalyses.Finiteelementsimulationswereperformedwhereoneofthedesignvariableswasvariedbetweentheirrespectiveranges(Table 7-2 )andotherswereheldconstant(tW=0.4tB=2.5,tT=2.0andh=32.5mm).Thevariationofthebucklingloadfactorincreasestremendouslywiththeincreaseinthethicknessoftopfacesheetwhilethevariationisnotpronouncedwiththeincreaseinwrapthickness.Thisconrmsthatbucklingofthetopfacesheet(Figure 7-3 )leadstobucklingofthetopweb(Figure 3-7 ). WiththevariationoftheresponsesshowninFigures 7-1 7-2 7-3 and 7-4 ,secondorderpolynomialresponsesurfacesoftheconstraintswereconstructed. Thedesignpointsfortheresponsesurfacegenerationwerecreatedusingthelatinhypercubesamplingwithinthebounds.Asstatedearlier,foraquadraticresponsewith4designvariables,thenumberofpolynomialcoefcientsare15.Asaruleofthumb,designpointsmorethantwicethenumberofcoefcientsareconsidered.Inthisproblematotalof60designpointswereinputtotheniteelementanalyses,togeneratetheresponsesurfaces.Thisrequired180niteelementanalysesfor3responses(transient 98
PAGE 99
Figure7-1.VariationofresponseswrtthicknessofthewraptW.(tB,tTandhwereheldconstant) heattransfer,thermalstressandthermalbucklinganalyses).ThetimetakentoperformallthethreeFEsimulationsandextractallthedesiredresponseatasingledesignpointis8.5minutessummingto8.5hoursfor60designpoints.Thisisassumedastheavailablecomputationalbudgettondthedeterministicoptimum. Insteadofusing60designstobuildaglobalresponsesurface,thecomputationalresourceswereusedtoapproximatetheconstraintboundariesaccuratelytoestimateanaccurateoptimum.Itwasfoundthatallthefollowingconstraints,temperature,stress1andbucklingloadwereactiveintheoptimization.Hencetoimprovetheaccuracyofeachactiveconstraint,EfcientGlobalReliabilityAnalyseswasapplied.EGRAbalancesbothexplorationandexploitationofthedesignspacetoadddesignpointsthatimprovedtheaccuracyofthetargetboundary.Aninitialresponsesurfacewasconstructedusing16LHSdesignpointsandsixadditionalpointswereacquiredforeachconstraintusingEGRA.Sincetemperaturegradientactsastheloadsforthermalstressandbuckling 99
PAGE 100
Figure7-2.VariationofcriticalresponseswrtthicknessofthebottomfacesheettB.(tW,tTandhwereheldconstant) analyses,foreachstressorbucklinganalysis,atransientheattransferanalysishadtobeperformed.ThedesignsupdatedbyEGRAandthecorrespondingresponsesareprovidedinAppendix C .Thetrueresponsesattheadditionaldesignsareveryclosetothevalueofthetargetconstraint. Adesignpointonthetargetboundarywasevaluatedwiththeupdatedsurrogate.Whenthedifferencebetweenthesurrogatepredictionandthetrueresponsewas10)]TJ /F4 7.97 Tf 6.59 0 Td[(3,EGRAsamplingwasstopped.Whenthisprodedurewasfollowed,Itwasfoundthat4pointsupdatedbyEGRA,togetherwiththeinitial16pointsapproximatedthetargetboundaryaccuratelyyieldingaccurateoptimum.Thusthesimulationsrequiredwereonly68initialFEsimulationscomparedto180FEsimulations(globalsurrogate)andthecomputationaltimewas3hoursand15minutes.Thecomputationaltimewasreducedto40%whencomparedtothatofglobalsurrogates. 100
PAGE 101
Figure7-3.VariationofcriticalresponseswrtthicknessofthetopfacesheettT.(tW,tBandhwereheldconstant) Furthera20DOEglobalresponsewasusedforoptimizationtoprovideareasonablecomparisonwiththesamenumberofDOE.Theoptimumdesignvariablesd,theminimizedmassestimatedusing60DOE,20DOEand16+4EGRADOEsurrogatesaretabulatedinTable 7-5 Table7-5.OptimumdesignvariablesandminimizedstructuralmassthroughglobalandEGRADOEfordeterministicoptimization DOEtWmmtBmmtTmmhmmMasskg/m2 60DOE0.42.841.4132.321.420DOE0.42.871.3832.721.816+4EGRA0.42.851.3731.721.3 Theoptimizationthroughthesurrogatesyieldedoptimawithalltheconstraintsbeingactive.Further,theconstraintsevaluatedusingthesurrogatesarecomparedwithniteelementanalysestoassesstheaccuracyofthesurrogateattheirrespectiveoptimuminTable 7-6 .The60DOEoptimumwasapparentlynottheoptimumasthe 101
PAGE 102
Figure7-4.Variationofcriticalresponseswrtheightofthefoamh.(tW,tTandtBwereheldconstant) constraintsdeterminedbythesurrogatedifferfromtheniteelementanalyses.Theniteelementanalysesareconsideredastruevalueshere.Thus,itcanbeseenthatthoughtheglobalDOEhas60designpoints,itdoesnotapproximatetheconstraintboundaryaccuratelyleadingtoerroneousresults.Forexample,thoughtheTmax;BFSglobalsurrogatepredicted473Katthe(60DOE)optimum,thetrueresponseis469.6K.Evenwithlargenumberofdesigns,thediscrepancyisabout4K.HoweverthesurrogateconstructedusingthesequentialsamplingofEGRAhasimprovedthetargetboundary(472.9K)topredictaccurateoptimum.TheconstraintvaluespredictedbytheEGRAimprovedsurrogateagreeswellwiththetruevaluesoftheniteelementanalyseswithlessthanapercenterror.Theoptimumdeterminedfromthe20DOEsurrogateprovedthatasurrogatewith20globaldesignsisgoingtobegrosslyinaccuratewhereasthe16+4=20DOEconstructedusingEGRAwashighlyaccurate.Thisshowsthat 102
PAGE 103
whenthecomputationalresourcesareefcientlydirectedtowardstargetregions,thecomputationalcostcanbereducedto40%oftheavailablecomputationalbudget. Table7-6.ComparisonbetweentheaccuracyofglobalsurrogateandEGRAsurrogateatthedeterministicoptimumusingcorrespondingsurrogates.Thepredictedstandarddeviationisgivenin() SurrogateFiniteElementAnalysesTmax;BFSBucklingWrapStressTmax;BFSBucklingWrapStressKload1MPaKload1MPa 60DOE473(0.63)1.5(0.05)215.3(0.65)469.61.54211.320DOE473(3.4)1.5(0.78)215.3(1.6)467.51.77217.316+4EGRA473(9.1)1.5(2.5)215.3(2.4)472.91.51214.6 Threeoptimaweredeterminedusing60DOE,20DOEandEGRAupdatedsurrogates.Theresponsesatthethreeoptimawereevaluatedusingallthesurrogatestoanalyzehowthesurrogatesperformedattheotheroptima.Table 7-7 showstheresponsesofalltheoptimaevaluatedusingthe60DOEglobalsurrogate.Amaximumdifferenceof8%wasobservedinthebucklingloadofthe20DOEoptimum.Itwasfoundthattheerrorbetweentheresponseswerelessthan8%,howeverthecomputationaltimerequiredfortheanalyseswas60%morethanthatofEGRA. Table7-7.Comparisonofresponsesatallthedeterministicoptimausingthe60DOEsurrogate OptimaSurrogateFiniteElementAnalysesTmax;BFSBucklingWrapStressTmax;BFSBucklingWrapStressKload1MPaKload1MPa 60DOE4731.5215.3469.61.54211.320DOE470.91.62214.8467.51.77217.316+4EGRA475.71.45212.7472.91.51214.5 Inasimilarfashion,thethreeoptimawereevaluatedusingthe20DOEsurrogateandtabulatedinTable 7-8 .Thebucklingloadsofthe60DOEandEGRAoptimahadanerrorof13%whencomparedtothetrueresponse.ThevalueofTmax;BFSoftheEGRAoptimumdifferedby5K.Althoughbypercentitisalittleover1%error,thisisthelargestdifferencebetweenpredictionandtrueresponseusinganysurrogate.Tosumup,Itwasfoundthat20DOEsurrogatewasinaccurateattheotheroptimaaswell. 103
PAGE 104
Table7-8.Comparisonofresponsesatallthedeterministicoptimausingthe20DOEsurrogate OptimaSurrogateFiniteElementAnalysesTmax;BFSBucklingWrapStressTmax;BFSBucklingWrapStressKload1MPaKload1MPa 60DOE474.21.35214.7469.61.54211.320DOE4731.5215.3467.51.77217.316+4EGRA478.11.31210.3472.91.51214.5 TheresponsesofthethreeoptimaevaluatedusingtheEGRAupdatedsurrogatealongwiththetrueresponsesisshowninTable 7-9 .ItwasfoundthatEGRAupdatedsurrogatepredictedresponsesat60DOEwithlessthan4%error.Howeveratthe20DOEoptima,therewas16%errorinthebucklingloadresponses.ThiscouldattributedtothesignicantdifferenceintheheightofthefoambetweentheEGRAand20DOEoptimawhichisawayfromthetargetboundary.EGRAupdatedsurrogateseemstohaveminimalerroratalltheoptimawhencomparedtoothersurrogatesexceptforthebucklingresponseofthe20DOEoptimum.Alsothewrapstresseswerepredictedwithmaximumof3%errorbyallthesurrogates.Thus,whenanalysedonthebasisoncomputationaltimerequiredandaccuracyofthesurrogate,EGRAupdatedsurrogatesseemstoperformbetterthan60and20DOEglobalsurrogates. Table7-9.ComparisonofresponsesatallthedeterministicoptimausingtheEGRAupdatedsurrogate OptimaSurrogateFiniteElementAnalysesTmax;BFSBucklingWrapStressTmax;BFSBucklingWrapStressKload1MPaKload1MPa 60DOE470.61.56213.4469.61.54211.320DOE465.91.46215.6467.51.77217.316+4EGRA4731.5215.3472.91.51214.5 Toestimatethereliabilityoftheoptimumdesign,uncertaintymodelingandanalysisofthefailureresponsesisrequiredwhichisdiscussedinthenextsection. 7.5.2SensitivityAnalysis Thetotalnumberofuncertaintiesconsideredinthedesignwere39inputrandomvariables(Table 7-3 ).Tostudytheuncertaintypropagationandestimatereliability, 104
PAGE 105
responsesurfaceapproximationoftheoutputparametersasafunctionoftheinput39uncertaintieswasconstructed.Aminimumof40designsarerequiredtoconstructalinearresponsesurface(numberofcoefcients=40).Sinceallthevariableswouldnotbesensitivetotheresponses,sensitivityanalysiswasdesiredtostudythesensitivityoftherandomvariablestowardsthefailureresponses.Thiswouldaidinndingthemostsensitivevariablesandthusreducingthenumberofniteelementsimulationsrequiredtoconstructresponsesurfaces.Designpointstwicethenumberofcoefcientswerecreatedtoanalyzethesensitivevariables.Thecoefcientsofthelinearresponsewereusedasweightstondthemostsensitivevariablestotheresponses,maximumBFStemperature,WrapStressandbucklingloadfactor. AstheBFStemperaturedoesnotdependonthethermo-mechanicalpropertiesofthecomponents,therandomvariablesconsideredareheatuxloads,geometryandthermalpropertiesofthevariouscomponentsoftheITPS.Alinearpolynomialwastwith30points(15coefcients)wasconstructedasfunctionofthe14randomvariables.Figure 7-5 showsthelinearcoefcientsoftherandomvariablesasabarplot.Thehigherthemagnitudeofthecoefcient,thehigherthesensitivity.TheBFStemperaturewashighlysensitivetoq1andq2,tW,handkandCpofthewrap.Inanefforttoreducethenumberofvariablesfurtherq2wascalculatedasConstant+q2.Thetotalof5inputrandomvariablesweresensitivetoBFStemperature(q1,tW,h,kandCp). Inthecaseofwrapstressandbucklingload,inadditiontotheloads,geometryandthermo-mechanicalproperties,thethermalconductivityandspecicheatofthewrapwasalsoconsideredastemperaturegradientsactasloadforthethermalstressandbucklinganalysis.Withatotalof33uncertainvariables,66LHSdesignsweregeneratedandlinearresponseforwrapstress11andbucklingloadwereconstructed. Figure 7-6 and 7-7 showstheproductofuncertaintyandcoefcientofeachrandomvariableasbarplot.Fromthemagnitudeofthecoefcients,themostsensitivevariableswerechosenforeachresponse.Thesensitivevariablesofwrapstresswereq1,tW,tB, 105
PAGE 106
Figure7-5.SensitivityanalysisofmaximumBFStemperature,OrderofthevariableskandCpofWrap,BFS,TFS,Foam tT,E1,ofTFS,E1,ofthewrap.Furtherthesensitivevariablesofbucklingloadwereq1,tW,tT,h,ofTFS,ofwrap. Inordertomakesurethatsomeofthesensitivevariablesarenotneglectedintheprocess,coefcientofvariationoftheresponseasfunctionsoftheinitialnumberofuncertaintiesandreducednumberofuncertaintiesarecalculatedfromtheconstructedpolynomialequation.Iffistheresponse,theuncertaintyinresponseduetouncertaintyininputvariablesx1;x2;::;xiiscalculatedas f=s X@f @xi22xi(7) Table7-10.Differenceinuncertaintyinresponsesbeforeandaftersensitivityanalysis ResponseDifferenceinuncertainty Tmax;BFS4%Wrapstress117%Bucklingload1% 106
PAGE 107
Figure7-6.SensitivityanalysisofWrapstress11.Orderofthevariables,heatuxloads:q1,q2,geometry:tW;tB;tT;h,materialpropertiesofBFS:E1;E2;v12;v23;G12;G23;1;2,foam:E;v;,TFS:E1;E3;v12;v13;G12;G13;andWrap:E1;E3;v12;v13;G12;G13;;k;Cp ThedifferenceinuncertaintyintheresponseduetoinitialnumberofuncertainvariablesandreducednumberisprovidedinTable 7-10 .Themaximumdifferenceis7%variationinstressresponse.ThisisanacceptabledifferenceasasignicantnumberofexpensiveFEsimulationsaresavedasresultofthesensitivityanalysis.Thusthetotalnumberofsensitivevariablesreducedfrom39to11cumulatively. 7.5.3ReliabilityoftheDeterministicOptimum Thedesignspaceformedbytheuncertainvariablesaresmallandthevariationoftheresponsesisfairlylinearinthisrandomvariablespace.Hencelinearsurrogatesoftheresponseswereconstructedasfunctionoftherandomvariables.TheuncertaintyinthesevariablesisshowninTable 7-3 .Foratotalof11randomvariables,13LHSdesignsareconsidered.Heretheresponsesare:bottomfacesheettemperature,stressesandbucklingload.Thecorrespondingcapacitiesare:allowableBFStemperature, 107
PAGE 108
Figure7-7.Sensitivityanalysisofbucklingloadfactor.Orderofthevariables,heatuxloads:q1,q2,geometry:tW;tB;tT;h,materialpropertiesofBFS:E1;E2;v12;v23;G12;G23;1;2,foam:E;v;,TFS:E1;E3;v12;v13;G12;G13;andWrap:E1;E3;v12;v13;G12;G13;;k;Cp strengthvaluesandallowablebucklingload.Thesesurrogateswerealsousedtodeterminetheuncertaintypropagationfromtheinputrandomvariablestotheoutputresponses.Theuncertaintyintheresponseprovidedinsightontheappropriatesamplesizestoestimateaccurateprobabilityoffailure. Table7-11.Uncertaintyinresponseduetoinputuncertainty ResponseCoefcientofVariation% Tmax;BFS3.5Wrapstress119.1Bucklingload15.1 Theuncertaintyintheresponseisestimatedfromthestandarddeviationorcoefcientofvariation(CV)of1000samples.Thisprocesswassimulatedfornumerousrepetitions(1000)tondnominalcoefcientofvariationsforeachresponse.ThemeanoftheCVestimatesisgiveninTable 7-11 .ItcanbeseenthattheuncertaintyinthebucklingloadfactorishighcomparedtotheBFStemperatureandwrapstress.This 108
PAGE 109
kindofsituationwasaddressedbyrearrangingtherandomvariablesinthelimitstateandassigningappropriatesamplesize.However,capacityisdeterministicandtheerrorebkwasonlyvariablethatcouldbemovedtothecapacitytoreducetheuncertainty.Howeverthisdidnothelpisreducingtheuncertainty,hencethebucklingprobabilityoffailurewascalculatedusinglargenumberofsamples(Nbk=5000). Forthedeterministicoptimum(Table 7-5 ),theprobabilityoffailurewasestimatedforthethreefailuremodesusingthecorrespondinglimitstates(Equation 7 )Randomsamplesofthematerialproperties,loadandgeometryaregeneratedusingMonteCarlosimulations.Withthegeneratedsamplesofinputrandomvariables,theoutputresponsessuchascomponentstresses,maximumBFStemperatureandbucklingloadswereevaluatedfromthecorrespondinglinearresponseapproximations.TheprobabilityoffailureforeachfailuremodeofwascalculatedusingseparableMonteCarlomethod 7 .Itwasassumedthatourcomputationalbudgetonlypermitted500samplesfortheresponse,capacityanderror(N=M=Q=500)withanexceptiononthesamplessizeforbucklingload(Nbk=5000).InseparableMonteCarlomethod,theresponsesamples(N)werecomparedagainstallthecapacitysamples(M=500)resultingin250,000evaluationsofthelimitstate.Henceforeacherrorsample,250,000evaluationsresultedinasingleprobabilityoffailure.Themeanofthe500estimatesofprobabilityoffailureisconsideredtheprobabilityoffailureduetoerrorandvariability. Table7-12.Individualprobabilitiesoffailureofthedeterministicoptimum.Thetotalprobabilityoffailure=2.7%. FailuremodeTmax;BFSWrapstress11Bucklingload probabilityoffailure^p0.00180.00260.022Response472.9K214.5MPa1.51SafetyFactor77K1.51.5 TheindividualprobabilitiesoffailurearegiveninTable 7-12 .Itisassumedthatthethreefailuremodesareapproximatelyindependentandthetotalprobabilityoffailure(systemreliability)iscalculatedasthesumofalltheprobabilitiesoffailure.For 109
PAGE 110
thedeterministicoptimumthesystemprobabilityoffailurewas2.7%amountingtoareliabilityindexof1.93.Themassofthedeterministicoptimumwas21.3kg=m2. Oftentimesdeterministicoptimizationyieldconservativedesignsastheyarebasedonsafetyfactors.ThusreliabilitybasedoptimizationoftheITPSpanelwasperformed. 7.5.4Reliabilitybasedoptimization Deterministicoptimizationinvolvesapplyingsafetyfactorsonfailureresponses.Thesesafetyfactorscanbeconsideredasriskallocatedtoeachconstraint.Optimizationusingsafetyfactorsoftenleadstoconservativedesigns.Reliabilitybasedoptimizationdoesnotinvolvesafetyfactors,howeverincludesareliabilityconstraintthatisafunctionoftheresponses.Sincetherenohardconstraintsontheresponses,theoptimizationprocessmightyieldnon-conservativedesigns. Reliabilitybasedoptimizationinvolvesminimizingthestructuralmassofthepanelbyconsideringalltheuncertaintiesinthedesign.Inthisproblemthepanelwasoptimizedtohaveareliabilityequaltoorgreaterthanthereliabilityofthedeterministicoptimum.Itwasalsodesiredtoanalyzetheriskallocationtothefailureinducingresponsesincomparisontothedeterministicoptimum. Table7-13.Lowerandupperboundsofdesignvariablesforprobabilisticoptimization DesignVariableLowerBound(LB)UpperBound(UB) Wrapthickness,tWmm0.41.0Bottomfacesheetthickness,tBmm2.04.0Topfacesheetthickness,tTmm1.03.0Heightofthefoam,hmm29.036.0 TheoptimizationfortheprobabilisticoptimumwasalsoperformedusingtheMATLABoptimizerfmincon()withconstraintonthedeterministicreliabilityindex1.93.Thedeterministicoptimizationandtheknowledgeoftheresponsesatthedeterministicaidedinshrinkingthedesignvariablespace(Table 7-13 ).Thispreventedsamplingatregionswithlargeprobabilitiesoffailure. Similartotheresponsesurfaceapproximationofthedeterministicconstraints,30LHSdesignofexperimentswerecreatedtobuildaasecondorderpolynomialforthe 110
PAGE 111
reliabilityindex.Ateachdesignpoint,therandomnessintheuncertainvariableswereintroducedanddesignpointswerecreatedintherandomvariablespace.Therandomvariablespacewassmallmakingthevariationoftheresponseslinear.Foratotalof11uncertainvariablesconsidered,linearresponsesurfacesconstructedwith13DOEwereaccurate.ThesepointswereinputintheniteelementanalysestocomputethemaximumBFStemperature,wrapstressandbucklingload.Theseresponseswereusedtoconstructalinearapproximationoftheresponses.Further,largenumberofsamplesofuncertaintiesweregeneratedandtheresponseswereevaluatedusingtheaforementionedlinearapproximation.Usingthegeneratedsamplesofresponse,capacityanderrorsamples,theprobabilityoffailurewasestimatedusingtheseparableMonteCarlomethod.Thisprocesswasrepeatedforeverypointcreatedinthedesignvariablespace.Theniteelementanalysesat13designsapproximatelytook1hourand50minutes.FurtherthetimetoestimatetheprobabilityoffailurethroughseparableMonteCarlomethodwasapproximately25minutes.Thetotalcomputationaltimerequiredtoestimateprobabilityoffailureatasingleinthedesignvariablespacewas2hourand15minutes. FurtherEGRAwasemployedtoimprovethetargetboundaryequalto1.93(2.7%pf).Foreveryadditionalpointacquired,theentireprocessmentionedabovewasrepeatedtondtheprobabilityoffailurewhichwasthenincludedtothedesignofexperimentstoupdatethesurrogate.Itwasfoundthatevenafteradding25newdesignpoints,theapproximationofthetargetboundarywasinaccuratesuggestingthatasecondorderpolynomialcouldbeapoorapproximationtobeginwith.Toincreasetheorderofthepolynomial,largernumberofinitialdesignswererequired.Forexample,toconstructa3rdorderpolynomialasfunctionof4designvariables,thenumberofcoefcientsequals35andaminimumof70designpointsarerequired.Consideringthetimerequiredtoestimatetheprobabilityoffailureofonedesign,alargerDOEwascomputationallyintractable. 111
PAGE 112
Thuskrigingmodelwasconsideredforthereliabilityindexresponsesurfaceasitdoesnotrequireminimumnumberofdesignstoconstructaresponsesurface.ToaffordenoughadditionaldesignsthroughEGRA,a20DOEwascreatedtobuildtheinitialresponsesurface.Furtheranadditional27designswereacquiredtoimprovethedesiredlevelofreliabilityusingEGRA.Sowithatotalof47designpoints,thecomputationaltimerequiredwasapproximately105hoursand45minutes.GlobalsurrogatesofthereliabilityindexcouldbecreatedtocomparewiththeEGRAupdatedsurrogateduetolimitedcomputationalresources.Theprobabilisticoptimizationwasperformedwiththeupdatedsurrogate.TheoptimizeddesignvariablesandtheminimizedmassoftheprobabilisticoptimumisgiveninTable 7-14 Table7-14.OptimumdesignvariablesandminimizedstructuralmassthroughEGRAupdatedDOEforreliabilitybasedoptimization DOEtWmmtBmmtTmmhmmMasskg/m2 20+27EGRA0.42.01.331.119.4 Table7-15.Individualprobabilitiesoffailureoftheprobabilisticoptimum.Thetotalprobabilityoffailurepf=2.6%. FailuremodeTmax;BFSWrapstress11Bucklingload probabilityoffailure^p0.00340.0040.019Response492K245.3MPa1.7SafetyFactor55K1.241.7 Table7-16.Individualprobabilitiesoffailureofthedeterministicoptimumusingprobabilisticoptimumsurrogate.Thetotalprobabilityoffailurepf=2.8%. FailuremodeTmax;BFSWrapstress11Bucklingload probabilityoffailure^p0.00270.0030.023Response483.2K224.6MPa1.6 Atotalof13designswerecreatedthroughLatinHypercubesamplingaroundtheoptimizeddesignandwasinputtoFiniteElementanalyses.Linearsurrogatesoftheresponseswereconstructedandtheprobabilityoffailurewasestimated.Thereliabilityindexwasestimatedas1.95whichamountedto2.6%probabilityoffailure.TheprobabilisticoptimumobtainedusingtheEGRAupdatedresponsesurfacewas 112
PAGE 113
accuratewithanerrorof2%inreliabilityindex.Thisamounttoa3%differenceintheprobabilityoffailure.Furtherusingtheprobabilisticoptimumresponsesurrogates(temperature,stressandbuckling),theprobabilityoffailureofthedeterministicoptimumwasestimateforthepurposeofvalidcomparison.Thetotalprobabilityoffailurewas2.8%.ThisincreaseinthepfestimatewasduetothediscrepancyintheresponseevaluationasthesomeofthedeterministictBandhwereoutoftheprobabilisticrandomvariablespace. Withthereliabilityindexsurrogate,thereliabilityindexofthedeterministicoptimumwasfoundtobe1.97withastandarddeviationof0.07.Thisamountstoaprobabilityoffailureof2.4%whilethedeterministicprobabilityoffailureestimatedthroughseparableMonteCarlowas2.7%(=1.93). Themassminimizedthroughdeterministicoptimizationwas21.3kg/m2.ThesafetyfactorsforthevariousconstraintsisshowninTable 7-12 .Fromthevalueoftheresponsesattheprobabilisticoptimum,thesafetyfactorswerecalculated,showninTable 7-16 .Theprobabilisticoptimizationprocesshasallocatedriskdifferentlythatreducesthemassfrom21.3to19.4kg/m2.Theoptimumdesignseemstohavehigherriskallocatedtotemperatureandstresswhencomparedtodeterministicsafetyfactor.However,theincreaseinthesafetyfactorofthebucklingloadfactorhasyieldedreducedmasswiththesameprobabilityoffailure.Thisdemonstratesthatreliabilitybasedoptimizationyieldsnon-conservativebutasafedesign. Further,itwasdesiredtondtheaccuracyoftheprobabilitiesoffailure.Thiswasaccomplishedbyemployingbootstrappingmethod.Thismethodwasappliedearlierforcomplexlimitstateandwasvalidatedbyestimatingtheerrorthroughlargenumberofrepetitionsinprobabilityoffailureestimationprocess. 7.5.5Errorintheprobabilityoffailureestimate Forasimplelimitstate(asinEquation 2 ),theaccuracyofseparableMonteCarlocanbeestimatedasderivedbySmarsloketal.[ 16 ].Inthecaseagenerallimitstate 113
PAGE 114
(Equation 2 ),bootstrappingwasproposedtoassesstheaccuracyoftheprobabilityoffailure. Bootstrappingmethodismoresuitablewhencomputationoftheresponseinthelimitstateisexpensive.Moreover,thismethodcanbeemployedincaseoflowprobabilitiesoffailureoftheorder10)]TJ /F4 7.97 Tf 6.59 0 Td[(6-10)]TJ /F4 7.97 Tf 6.59 0 Td[(8.SincetheITPSprobleminvolvessimplelimitstates,surrogatesareusedtoevaluateresponsesfortheprobabilityoffailurecalculation,andthefacttheprobabilityoffailureisashighas3%,bootstrappingisnotrequiredtoestimatetheaccuracyoftheprobabilityoffailure.However,themethodisemployedinthisproblemtoshowthatthemethodworksforsuchsituationstoo. Theaccuracyoftheprobabilityoffailureestimatewascalculatedwithab=1,000bootstraprepetitions.Randomsamplesofresponseswerebootstrapped,butthecapacityanderrorsweresampledforeverybootstraprepetition.Thatis,ineachofthe1,000bootstraprepetitions,theresponseswerere-sampledfromthesame500samples,whilethecapacityanderrorshadafreshsampleeverytime. Table7-17.Errorintheprobabilitiesoffailureofthedeterministicoptimumusingbootstrappingandrepetitions.Bootstrapsamples=1000,Repetitions=1000 RepetitionsBootstrappingProbabilityoffailuremean(pf)std(pf)mean(^pboot)std(^pboot) ^pth0.0018510)]TJ /F4 7.97 Tf 6.59 0 Td[(40.0019610)]TJ /F4 7.97 Tf 6.59 0 Td[(4^pstr0.0026110)]TJ /F4 7.97 Tf 6.59 0 Td[(30.0026910)]TJ /F4 7.97 Tf 6.59 0 Td[(4^pbk0.022710)]TJ /F4 7.97 Tf 6.59 0 Td[(30.023810)]TJ /F4 7.97 Tf 6.59 0 Td[(3 Table7-18.Errorintheprobabilitiesoffailureoftheprobabilisticoptimumusingbootstrappingandrepetitions.Bootstrapsamples=1000,Repetitions=1000 RepetitionsBootstrappingProbabilityoffailuremean(pf)std(pf)mean(^pboot)std(^pboot) ^pth0.0034810)]TJ /F4 7.97 Tf 6.59 0 Td[(40.003710)]TJ /F4 7.97 Tf 6.59 0 Td[(4^pstr0.004110)]TJ /F4 7.97 Tf 6.59 0 Td[(30.003110)]TJ /F4 7.97 Tf 6.59 0 Td[(3^pbk0.019710)]TJ /F4 7.97 Tf 6.59 0 Td[(30.021910)]TJ /F4 7.97 Tf 6.59 0 Td[(3 TheaccuracyoftheseparableMonteCarlomethodwasalsoassessedbyanempiricalstandarddeviationobtainedbyperformingn=1000repetitions.Themeanandthestandarddeviationoftheprobabilitiesoffailureestimatedthroughrepetitionsand 114
PAGE 115
bootstrappingaretabulatedinTable 7-17 forthedeterministicoptimumandTable 7-18 fortheprobabilisticoptimum.Theerrorintheprobabilityoffailureisapproximately3timeslessthanthemeanestimates.Themeanprobabilitiesoffailureoftheprobabilisticoptimumdifferby7%forbucklingloadwhileotherpfestimatesagreewell.Thebootstrappingmethodiscomparablewiththeempiricalestimatesdeterminedthroughrepetitionswithmaximumdiscrepancyinthebucklingprobabilitiesoffailure.Themajordifferenceisintheerrorestimatesfor^pbkwhichwas14%and21%fordeterministicandprobabilisticoptimumrespectively.Thiscanbeattributedtohighuncertaintyinthebucklingloads(15%). 7.5.6Summary ThreecriticalresponsescausingfailureoftheITPSpanelwasincludedasconstraintsintheoptimizationprocesstodeterminethedeterministicandprobabilisticoptimumoftheITPSpanel.Surrogatemodelsefcientlyreplacedcomputationallyexpensiveniteelementanalyses.Inadditiontothis,theevaluationofdeterministicconstraintsusingEGRAupdatedsurrogatesreducedthecomputationalrequiredto40%oftheavailablecomputationalbudget.Itwasalsofoundthatglobalsurrogateswithlargenumberofdesignsdonotalwaysprovidetheaccurateoptimum.EGRAwasalsosuccessfullyusedtoupdatereliabilityindexsurrogatesforreliabilitybasedoptimization. ThesystemreliabilityofthepanelwasestimatedusingMonteCarlosimulations.ThestatisticalindependenceoftherandomvariablesallowedSeparableMonteCarlomethodtobeemployedtoestimatetheprobabilityoffailure.Thesafetyfactorbaseddeterministicoptimumyieldedanoptimaldesignwithstructuralmassof21.3kg/m2with2.7%probabilityoffailure.Furthermassoftheprobabilisticoptimumwas19.4kg/m2withdeterministicprobabilityoffailure.Reliabilitybasedoptimizationdidnotinvolvehardconstraints(safetyfactors)thusdemonstratingthatriskallocationplaysanimportantroleinndinganon-conservativebutsafedesign.Bootstrappingprovidedreasonableestimatesofuncertaintyintheprobabilityoffailureestimates. 115
PAGE 116
CHAPTER8CONCLUSIONS 8.1Conclusion ThehighcomputationalcostofthedeterministicandprobabilisticoptimizationoftheITPSpanelwasefcientlyaddressedusingvarioustechniquesatdifferentphasesoftheoptimizationprocess.Toreducethecomputationalexpensesinvolvedinthestructuralanalysis,niteelementbasedhomogenizationmethodwasemployed,homogenizingthe3DITPSmodeltoa2Dorthotropicplate.HoweveritwasfoundthathomogenizationwasapplicableonlyforpanelsthataremuchlargerthanthecharacteristicdimensionsoftherepeatingunitcellintheITPSpanel.Henceasingleunitcellwasusedfortheoptimizationprocesstoreducethecomputationalcost.Thefailureconstraintsinvolvedintheoptimizationprocessdemandedcomputationallyexpensiveniteelementanalyseswhichwasreplacedbyefcient,lowdelitysurrogatemodels.Thelimitedcomputationalresourcesweredirectedtowardstargetregionsforaccuraterepresentationofconstraintsusingadaptivesamplingstrategies. Itwasfoundthatcomputationalcostofthedeterministicoptimizationwasreducedto40%oftheavailablecomputationalbudgetbyemployingEfcientGlobalReliabilityAnalysis.Whenitwasfoundthatthecomputationalcostwasdrasticallylowered,theadaptivesamplingmethodwasalsoemployedtoimprovetheapproximationofthereliabilityindexinthereliabilitybasedoptimizationprocess.Thesafetyfactorbaseddeterministicoptimumyieldedanoptimaldesignwithstructuralmassof21.3kg/m2with2.7%probabilityoffailure.Reliabilitybasedoptimizationwithoutsafetyfactorbasedconstraintsallocatedhigherrisktothefailureresponsesthatfurtherminimizedmass.Themassoftheprobabilisticoptimumwas19.4kg/m2withdeterministicprobabilityoffailure.Thisshowedthattheriskallocationinprobabilisticoptimizationplayedanimportantroleindeterminingnon-conservativebutsafedesigns.Theaccuracyofthe 116
PAGE 117
probabilityoffailurewasestimatedusingthebootstrappingmethod.Thebootstrappingestimateswerecomparabletotheempiricalestimatesdeterminedthroughrepetitions. 8.2FutureWork Possiblefutureworkinthestructuralanalysis Thestressfailureinthepanelwaspredictedthroughmaximumstresstheoryasthemainfocuswasonthemethodologyinvolvedintheoptimizationprocess.ThiscanbereplacedwithamoresuitablecompositefailurecriterionsuchasTsai-HillorTsai-Wucriteriontopredictfailure. Asbucklingoccursinthepanelandnotinaunitcell,bucklingcouldbeanalysedforanentirepanel. 117
PAGE 118
APPENDIXATHERMALPROPERTIESOFITPSCOMPONENTS TableA-1.DensityoftheITPSComponents ITPScomponentmaterialDensitykg/m3 BFSGraphite/epoxy1576.2TFS&WrapSiC/SiC2500FoamAETB86.1 ThethermalconductivityandspecicheatcapacityoftheITPScomponentsaretemperaturedependentproperties. TableA-2.ThermalPropertiesofthebottomfacesheet-GraphiteEpoxy TemperatureThermalConductivitykW/m/kSpecicheatCpW/m2 2880.026512.93060.027554.85580.0441172.36480.0501381.67380.0541591.010080.0702177.111880.0762533.013680.0842899.415480.0913223.817280.0973558.819080.1033967.0 TableA-3.ThermalPropertiesofthetopfacesheetandWrap-SiC/SiC TemperatureThermalConductivitykW/m/kSpecicheatCpW/m2 2939.50150012737.603000 118
PAGE 119
TableA-4.ThermalConductivityoftheinsulationfoam-AETB TemperatureThermalConductivitykW/m/k 0.0045255.50.0055394.40.0067533.30.0083672.20.0102811.00.0121949.90.01441088.80.01681227.70.01941366.50.02201505.40.02441644.30.02491783.2 TableA-5.Specicheatoftheinsulationfoam-AETB TemperatureSpecicheatCpW/m2 1884.0255.52637.6394.43165.2533.33454.0672.23617.3811.03717.8949.93768.11088.83805.71227.73805.71783.2 119
PAGE 120
APPENDIXBSTRENGTHPROPERTIESOFITPSMATERIALS TableB-1.DensityoftheITPSComponents Strength(MPa)SiC/SiCGraphite/EpoxyAETB Tensilestrengthinthe1-directionS1T3232,3120.689Tensilestrengthinthe2-directionS2T32339.20.689Compressivestrengthinthe1-directionS1C63218090.689Compressivestrengthinthe2-directionS2C63297.20.689ShearstrengthS1217633.20.379 120
PAGE 121
APPENDIXCDESIGNOFEXPERIMENTSADDEDBYEGRASAMPLINGTECHNIQUE TableC-1.DesignpointsacquiredusingEGRAtoimprovetheaccuracyofTemperatureboundary.Target=473K tWmmtBmmtTmmhmmTmax;BFS 0.46.03.027.42474.31.02.03.039.20475.60.726.01.031.49474.81.02.01.039.10473.50.46.01.027.38470.70.42.03.033.33472.0 TableC-2.DesignpointsacquiredusingEGRAtoimprovetheaccuracyofTemperatureboundary.Target=215.3MPa tWmmtBmmtTmmhmm1;W 0.42.01.7125.0225.41061.02.01.7140.0213.41061.05.153.040.0214.21060.44.503.025.0214.71060.46.03.031.4218.31060.46.02.3840.0217.8106 TableC-3.DesignpointsacquiredusingEGRAtoimprovetheaccuracyofbucklingloadboundary.Target=1.5 tWmmtBmmtTmmhmm 1.06.01.040.01.650.42.01.025.01.700.626.01.025.01.850.42.01.6540.01.440.46.01.8840.01.291.02.01.040.01.61 121
PAGE 122
TableC-4.DesignpointsacquiredusingEGRAtoimprovetheaccuracyofreliabilityindexboundary.Target=1.93 tWmmtBmmtTmmhmmReliabilityindex 0.523.811.6134.502.110.423.292.9535.401.770.442.142.6435.101.480.983.542.9130.701.960.702.652.6831.101.780.632.992.2829.801.820.842.491.9333.902.020.404.001.4629.002.130.402.001.2429.001.940.864.001.0036.001.850.402.001.5335.781.700.402.763.0029.001.801.002.001.0033.151.320.594.003.0036.003.131.002.003.0036.002.280.584.003.0029.001.750.702.003.0036.002.150.404.002.0736.001.931.004.001.9831.861.650.404.001.2932.031.720.402.002.4529.001.451.002.993.0034.092.010.403.161.2229.001.441.004.003.0032.591.900.622.001.0036.001.691.004.001.0033.992.380.403.051.6836.001.83 122
PAGE 123
REFERENCES [1] M.Biancolini,Evaluationofequivalentstiffnesspropertiesofcorrugatedboard,CompositeStructures,vol.69,no.3,pp.322,July2005. [2] J.F.Davalos,P.Qiao,X.F.Xu,J.Robinson,andK.E.Barth,Modelingandcharacterizationofber-reinforcedplastichoneycombsandwichpanelsforhighwaybridgeapplications,CompositeStructures,vol.52,2001. [3] N.Buannic,P.Cartraud,andT.Quesnel,Homogenizationofcorrugatedcoresandwichpanels,CompositeStructures,vol.59,no.3,pp.299,Feb.2003. [4] J.WallachandL.Gibson,Mechanicalbehaviorofathree-dimensionaltrussmaterial,InternationalJournalofSolidsandStructures,vol.38,no.40-41,pp.7181,Oct.2001. [5] O.A.Martinez,B.V.Sankar,R.T.Haftka,S.K.Bapanapalli,andM.L.Blosser,MicromechanicalAnalysisofCompositeCorrugated-CoreSandwichPanelsforIntegralThermalProtectionSystems,AIAAJournal,vol.45,no.9,pp.2323,Sept.2007. [6] A.Sharma,B.V.Sankar,andR.T.Haftka,Multi-delityanalysisofcorrugated-coresandwichpanelsforintegratedthermalprotectionsystems,in50thAIAAASMEASCEAHSASCStructuresStructuralDynamicsandMaterialsConference,2009,vol.45,pp.1. [7] A.Sharma,Multi-FidelityDesignofanIntegralThermalProtectionSystemforFutureSpaceVehicleduringRe-entry,Ph.D.dissertation,2010. [8] R.C.KuczeraandZ.P.Mourelatos,OnEstimatingtheReliabilityofMultipleFailureRegionProblemsUsingApproximateMetamodels,JournalofMechanicalDesign,vol.131,no.12,pp.121003,2009. [9] H.Arenbeck,S.Missoum,A.Basudhar,andP.Nikravesh,Reliability-BasedOptimalDesignandTolerancingforMultibodySystemsUsingExplicitDesignSpaceDecomposition,JournalofMechanicalDesign,vol.132,no.2,pp.021010,2010. [10] C.TuandR.Barton,ProductionYieldEstimationbytheMetamodelMethodWithaBoundary-FocusedExperimentDesign,inASMEDesignEngineeringTechnicalConference,AmericanSocietyofMechanicalEngineers,1997. [11] P.Ranjan,D.Bingham,C.Va,G.Michailidis,andA.Arbor,SequentialExperimentDesignforContourEstimationfromComplexComputerCodes,Technometrics,pp.527,2008. [12] E.VazquezandJ.Bect,ASequentialBayesianAlgorithmtoEstimateaProbabilityofFailure,inProceedingsof15thIFACSymposiumonSystemIdentication SYSID ,Saint-Malo,France,2009. 123
PAGE 124
[13] V.Picheny,D.Ginsbourger,O.Roustant,R.T.Haftka,andN.-H.Kim,AdaptiveDesignsofExperimentsforAccurateApproximationofaTargetRegion,JournalofMechanicalDesign,vol.132,no.7,pp.071008,2010. [14] B.J.Bichon,EfcientSurrogateModelingforReliabilityAnalysisandDesign,Ph.D.dissertation,2010. [15] B.P.Smarslok,Measuring,Using,andReducingExperimentalandComputationalUncertaintyinReliabilityAnalysisofCompositeLaminates,Ph.D.dissertation,2009. [16] B.P.Smarslok,R.T.Haftka,L.Carraro,andD.Ginsbourger,ImprovingaccuracyoffailureprobabilityestimateswithseparableMonteCarlo,InternationalJournalofReliabilityandSafety,vol.4,no.4,pp.393,2010. [17] B.Ravishankar,B.P.Smarslok,R.T.Haftka,andB.V.Sankar,SeparableSamplingoftheLimitStateforAccurateMonteCarloSimulation,in50thAIAAAS-MEASCEAHSASCStructuresStructuralDynamicsandMaterialsConference,2009,numberMay,pp.1. [18] V.Picheny,N.H.Kim,andR.T.Haftka,Applicationofbootstrapmethodinconservativeestimationofreliabilitywithlimitedsamples,StructuralandMultidisci-plinaryOptimization,vol.41,no.2,pp.205,Aug.2009. [19] B.Ravishankar,B.P.Smarslok,R.T.Haftka,andB.V.Sankar,ErrorEstimationandErrorReductioninSeparableMonteCarloMethod,AIAAJournal,vol.48,no.11,pp.2624,Nov.2010. [20] D.E.Myers,C.J.Martin,andM.L.Blosser,ParametricWeightComparisonofCurrentandProposedThermalProtectionSystem(TPS)Concepts,33rdAIAAThermophysicsConference,June1999. [21] N.Haddard,H.McWilliams,andP.Wagoner,NASAEngineeringDesignChallenges:ThermalProtectionSystems,NASAReport,2007. [22] J.Wilson,NASAFacts,NasaTechnicalMemorandum,1997. [23] M.Blosser,C.Martin,K.Darybeigi,andC.Poteet,ReusableMetallicThermalProtectionSystemsDevelopment,NASATechnicalMemorandum,vol.1,no.11,pp.1829,2004. [24] D.J.Rasky,F.S.Milos,andT.H.Squire,TPSmaterialsandCostsforFutureReusableLaunchVehicles,NasaTechnicalMemorandum,1997. [25] J.Domoulin,SpaceShuttleOrbiterSystems,1988. [26] R.Oakes,SpaceShuttleCeramicTiles,1992. 124
PAGE 125
[27] M.Blosser,DevelopmentofMetallicThermalProtectionSystemsfortheReusableLaunchVehicle,NasaTechnicalMemorandum110296,1996. [28] S.Bapanapalli,DesignofanIntegratedThermalProtectionSystemforFutureSpaceVehicles,Ph.D.dissertation,2007. [29] W.YuandT.Tang,VariationalAsymptoticMethodforUnitCellHomogenizationofPeriodicallyHeterogeneousMaterials,InternationalJournalofSolidsandStructures,vol.44,pp.1,2007. [30] C.-Y.LeeandW.Yu,Homogenizationanddimensionalreductionofcompositeplateswithin-planeheterogeneity,InternationalJournalofSolidsandStructures,vol.48,no.10,pp.1474,May2011. [31] O.A.Martinez,MicromechanicalAnalysisandDesignofanIntegratedThermalProtectionSystemforFutureSpaceVehicles,Ph.D.dissertation,2007. [32] A.Sharma,C.Gogu,O.A.Martinez,B.V.Sankar,andR.T.Haftka,Multi-FidelityDesignofanIntegratedThermalProtectionSystemforSpacecraftReentry,in49thAIAA/ASME/ASCE/AHS/ASCStructures,StructuralDynamicsandMaterialsConference,2008,numberApril,pp.1. [33] P.W.ChristensenandK.Anders,AnIntroductiontoStructuralOptimization,2008. [34] Z.Gurdal,R.T.Haftka,andP.Hajela,DesignandOptimizationofCompositeLaminates,JohnWiley&Sons,INC,1999. [35] X.QuandR.T.Haftka,DeterministicandReliability-BasedOptimizationofCompositeLaminatesforCryogenicEnvironments,AIAAJournal,vol.41,no.10,2003. [36] S.VenkataramanandP.Salas,OptimizationofCompositeLaminatesforRobustandPredictableProgressiveFailureResponse,AIAAJournal,vol.45,no.5,pp.1113,May2007. [37] Y.TsompanakisandM.Papadrakakis,Large-scalereliability-basedstructuraloptimization,StructuralandMultidisciplinaryOptimization,vol.26,no.6,pp.429,Apr.2004. [38] O.DitlevsenandH.O.Madsen,StructuralReliabilityMethods,NumberSeptember.2007. [39] R.E.Melchers,StructuralReliabilityAnalysisandPrediction,Wiley,1999. [40] A.HaldarandM.Sankaran,Probability,reliability,andstatisticalmethodsinengineeringdesign,Wiley,2000. [41] H.W.LehetaandA.E.Mansour,Reliability-basedMethodforOptimalStructuralDesignofStiffenedPanelsXp,MarinStructires,pp.323,1997. 125
PAGE 126
[42] M.DiSciuvaandD.Lomario,AcomparisonbetweenMonteCarloandFORMsincalculatingthereliabilityofacompositestructure,CompositeStructures,vol.59,no.1,pp.155,Jan.2003. [43] S.Kumar,D.Villanueva,B.V.Sankar,andR.T.Haftka,ProbabilisticOptimizationofIntegratedThermalProtectionSystem,in12thAIAA/ISSMOMultidisciplinaryAnalysisandOptimizationConference,2008,numberSeptember. [44] D.Villanueva,A.Sharma,R.T.Haftka,andB.V.Sankar,RiskAllocationbyOptimizationofanIntegratedThermalProtectionSystem,in8thWorldCongressonStructuralandMultidisciplinaryOptimization,2009,pp.1. [45] A.Sharma,B.V.Sankar,R.T.Haftka,andN.C.Ebaugh,Multi-FidelityAnalysisofCorrugated-CoreSandwichPanelsforIntegratedThermalProtectionSystems,in50thAIAAASMEASCEAHSASCStructuresStructuralDynamicsandMaterialsConferencethAIAA/ASME/ASCE/AHS/ASCStructures,StructuralDynamicsandMaterialsConference,2009,numberMay,pp.1. [46] D.Villanueva,R.T.Haftka,andB.V.Sankar,IncludingFutureTestsintheDesignofanIntegratedThermalProtectionSystem,in51stAIAAASMEASCEAHSASCStructuresStructuralDynamicsandMaterialsConference,2010,numberApril,pp.1. [47] T.Matsumura,R.T.Haftka,andB.V.Sankar,ReliabilityEstimationIncludingRedesignFollowingFutureTestforanIntegratedThermalProtectionSystem,in9thWorldCongressonStructuralandMultidisciplinaryOptimization,2011,pp.1. [48] A.I.KhuriandJ.A.Cornell,Khuri,A.I.;Cornell,J.A.1996:Responsesurfaces:designsandanalyses.,MarcelDekker,NewYork,2ndeditioedition,1996. [49] R.MyersandD.Montgomery,ResponseSurfaceMethodology:ProcessandProductOptimizationUsingDesignedExperiments,Wiley-Interscience,1995. [50] F.A.C.VianaandR.T.Haftka,Surrogate-basedOptimizationwithParallelSimulationsusingtheProbabilityofImprovement,in13thAIAA/ISSMOMultidis-ciplinaryAnalysisandOptimizationConference,2010,numberSeptember,pp.1. [51] F.A.C.Viana,C.Gogu,andR.T.Haftka,DETC2010-28813,inProceedingsoftheASME2010InternationalDesignEngineeringTechnicalConferences&ComputersandInformationinEngineeringConference,2010,pp.1. [52] F.A.C.Viana,SurrogateToolbox,MATLAB,2009. [53] M.PapilaandR.T.Haftka,ResponseSurfaceApproximations:Noise,ErrorRepair,andModelingErrors,AIAAJournal,vol.38,no.12,pp.2336,Dec.2000. 126
PAGE 127
[54] S.ShanandG.G.Wang,SpaceExplorationandGlobalOptimizationforComputationallyIntensiveDesignProblems:ARoughSetBasedApproach,StructuralandMultidisciplinaryOptimization,vol.28,2004. [55] D.R.Jones,M.Schonlau,andJ.William,EfcientGlobalOptimizationofExpensiveBlack-BoxFunctions,JournalofGlobalOptimization,pp.455,1998. [56] A.Guinta,A,J.M.McFarland,L.P.Swiler,andM.S.Eldred,Thepromiseandperilofuncertaintyquanticationusingresponsesurfaceapproximations,StructureandInfrastructureEngineering:Maintenance,Management,Life-CycleDesignandPerformance,vol.2,no.3-4,2006. [57] L.Wang,D.Beeson,S.Akkaram,andG.Wiggs,GaussianProcessMeta-ModelsforEfcientProbabilisticDesigninComplexEngineeringDesignSpaces,inProc.ofASMEInternationalDesignEngineeringTechnicalConference&ComputersandInformationinEngineeringConference,2005. [58] P.Ranjan,D.Bingham,andG.Michailidis,ErrataSequentialExperimentDesignforContourEstimationFromComplexComputerCodes,Technometrics,vol.50,no.4,pp.4. [59] B.J.Bichon,M.S.Eldred,L.P.Swiler,S.Mahadevan,andJ.M.McFarland,EfcientGlobalReliabilityAnalysisforNonlinearImplicitPerformanceFunctions,AIAAJournal,vol.46,no.10,pp.2459,Oct.2008. [60] L.P.Swiler,T.L.Paez,andR.L.Mayes,EpistemicUncertaintyQuanticationTutorial,inProceedingsoftheIMAC-XXVII,SocietyofExperimentalMechanics,2009. [61] E.Acar,R.T.Haftka,andN.H.Kim,EffectsofStructuralTestsonAircraftSafety,AIAAJournal,vol.48,no.10,pp.2235,Oct.2010. [62] C.C.a.AntonioandL.N.Hoffbauer,Fromlocaltoglobalimportancemeasuresofuncertaintypropagationincompositestructures,CompositeStructures,vol.85,no.3,pp.213,Oct.2008. [63] D.H.OhandL.Librescu,Freevibrationandreliabilityofcompositecantileversfeaturing-uncertainproperties,ReliabilityEngineeringandSystemSafety,vol.56,pp.265,1997. [64] A.K.Noor,J.H.Starnes,andJ.M.Peters,Uncertaintyanalysisofcompositestructures,ComputerMethodsinAppliedMechanicsandEngineering,vol.185,no.2-4,pp.413,May2000. [65] A.K.Noor,J.H.Starnes,andJ.M.Peters,UncertaintyAnalysisofStiffenedCompositePanels,CompositeStructures,vol.51,2001. 127
PAGE 128
[66] E.Acar,R.T.Haftka,andT.F.Johnson,EffectsofUncertaintyReductiononWeightofCompositeLaminatesatCryogenicTemperatures,inASME2005Inter-nationalDesignEngineeringTechnicalConferences&ComputersandInformationinEngineeringConference,2005,pp.1. [67] C.E.PapadopoulosandH.Yeung,UncertaintyestimationandMonteCarlosimulationmethod,FlowMeasurementandInstrumentation,vol.12,pp.291,2001. [68] M.H.KalosandP.A.Whitlock,MonteCarloMethods,Wiley-IntersciencePublications,JohnWileyandSons,NewYork,1986. [69] D.Padmanabhan,H.Agarwal,J.E.Renaud,andS.M.Batill,AstudyusingMonteCarloSimulationforfailureprobabilitycalculationinReliability-BasedOptimization,OptimizationandEngineering,vol.7,no.3,pp.297,Sept.2006. [70] R.Y.Rubinstein,VarianceReductionTechniques:SimulationandtheMonteCarloMethod,JohnWiley&Sons,NewYork,N.Y.,1981. [71] N.H.Kim,P.Ramu,N.V.Queipo,andM.Carlo,TailModelinginReliability-BasedDesignOptimizationforHighlySafeStructuralSystems,47thAIAA/ASME/ASCE/AHS/ASCStructures,StructuralDynamicsandMaterialsConference,pp.1,2006. [72] B.P.Smarslok,A.Dylan,andR.T.Haftka,SeparableMonteCarloSimulationAppliedtoLaminatedCompositePlatesReliability,in49thAIAA/ASME/ASCE/AHS/ASCStructures,StructuralDynamicsandMaterialsConference,2008,pp.1. [73] T.K.JacobsenandP.Brndsted,MechanicalPropertiesofTwoPlain-WovenChemicalVaporInltratedSiliconCarbide-MatrixComposites,JournalofAmericanCeramicSociety. [74] NASA,CeramicMaterialsDatabase,. [75] NASA,TPSXNASAMaterialsDatabase,. [76] P.Fenici,a.J.FriasRebelo,R.Jones,a.Kohyama,andL.Snead,CurrentstatusofSiC/SiCcompositesR&D,JournalofNuclearMaterials,vol.258-263,pp.215,Oct.1998. [77] ABAQUS,CAEUser'sManual,ABAQUSdocumentation6.9,. [78] ABAQUS,AnalysisUser'sManual,ABAQUSDocumentation6.9,. [79] R.V.MarreyandB.V.Sankar,MicromechanicalModelsforTextileStructuralComposites,Tech.Rep.,1995. 128
PAGE 129
[80] R.V.MarreyandB.V.Sankar,AMicromechanicalModelforTextileCompositePlates,JournalofCompositeMaterials,vol.31,no.12,pp.1187,June1997. [81] R.Gibson,PrinciplesofCompositeMaterials,McGraw-HillInc.,1994. [82] J.Whitney,StructuralAnalysisofLaminatedAnisotropicPlates,Technomic,1987. [83] C.Stamblewski,B.V.Sankar,andD.Zenkert,AnalysisofThree-DimensionalQuadraticFailureCriteriaforThickCompositesusingtheDirectMicromechanicsMethod,JournalofCompositeMaterials,vol.42,no.7,pp.635,Apr.2008. [84] K.PriceandR.Storn,DifferentialEvolutionforContinuousFunctionOptimization,1996. [85] F.Viana,R.Haftka,andL.Watson,SequentialSamplingforContourEstimationwithConcurrentFunctionEvaluations,StructuralandMultidisciplinaryOptimiza-tion,vol.45,no.4,2012. [86] B.J.Bichon,M.Sankaran,andM.Eldred,Reliability-BasedDesignOptimizationusingEfcientGlobalReliabilityAnalysis,in50thAIAA/ASME/ASCE/AHS/ASCStructures,StructuralDynamicsandMaterialsConference,2009. 129
PAGE 130
BIOGRAPHICALSKETCHBharaniRavishankarwasbornin1985inthestateofTamilNadu,India.ShegraduatedwithaBachelorofEngineeringinAeronauticalEngineeringfromMadrasInstituteofTechnologyin2006.ShestartedhergraduatestudiesatUniversityofFloridaasamaster'sdegreestudentinthedepartmentofMechanicalandAerospaceEngineeringinJanuary2007.ShejoinedtheCenterforAdvancedCompositesLabinMay2007wheresheworkedasResearchAssistantwithDr.BhavaniSankar.InAugust2008,shecontinuedtopursueherPhDdegreeinMechanicalEngineering.ShewasawardedtheAmeliaEarhartFellowshipfromZontaInternational. 130
|
|
