• Aucun résultat trouvé

A posteriori global error estimator based on the error in the constitutive relation for reduced basis approximation of parametrized linear elastic problems

N/A
N/A
Protected

Academic year: 2021

Partager "A posteriori global error estimator based on the error in the constitutive relation for reduced basis approximation of parametrized linear elastic problems"

Copied!
15
0
0

Texte intégral

(1)

HAL Id: hal-01305736

https://hal-mines-paristech.archives-ouvertes.fr/hal-01305736

Submitted on 22 Jan 2018

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

A posteriori global error estimator based on the error in the constitutive relation for reduced basis approximation

of parametrized linear elastic problems

Laurent Gallimard, David Ryckelynck

To cite this version:

Laurent Gallimard, David Ryckelynck. A posteriori global error estimator based on the error in the constitutive relation for reduced basis approximation of parametrized linear elastic problems.

Applied Mathematical Modelling, Elsevier, 2016, 40, pp.4271-4284. �10.1016/j.apm.2015.11.016�. �hal-

01305736�

(2)

A posteriori global error estimator based on the error in the constitutive relation for re duce d basis approximation of parametrized linear elastic problems

L. Gallimard

a,

, D. Ryckelynck

b

aLaboratoire Energétique Mécanique Electromagnétisme EA4416, Université Paris Ouest Nanterre La Défense, 50 rue de Sèvres Ville d’Avray 92410, France

bMINES ParisTech, PSL, research university, Centre des Matériaux, CNRS UMR 7633, France

Keywords:

Finite element analysis Model reduction Error bounds Global error estimator Reduced basis error indicator Finite element error indicator

a b s t r a c t

Inthispaperweintroduceaposteriorierrorestimatorbasedontheconceptoferrorinthe constitutiverelationtoverifyparametricmodelscomputedwithareducedbasisapproxima- tion.Wedevelopaglobalerrorestimatorwhichleadstoanupperboundfortheexacterror andtakesintoaccountalltheerrorsources:theerrorduetothereducedbasisapproximation aswellastheerrorduetothefiniteelementapproximation.Weproposeanerrorindicatorto measurethequalityofthereducedbasisapproximationandwededuceanerrorindicatoron thefiniteelementapproximation.

1. Introduction

FiniteElementMethodisacommontoolusedtoanalyzeanddesignparametrizedmechanicalsystems.However,whena largesetofparametersneedstobeintroducedinthemodelthecomputationaleffortincreasesdrasticallyandmanyauthorshave recentlyshowninterestindevelopingmodelreductionmethodsthatexploitthefactthattheresponseofcomplexmodelscan oftenbeapproximatedbytheprojectionoftheinitialmodelonalow-dimensionalreducedbasis[1–4].Reducedbasismethods aimatspeedingupthecomputationaltimeforcomplexnumericalmodels.Theyarebasedonanoffline/onlinecomputational strategywhichconsistindetermininginafirststepasetofsnapshotsorareducedbasis(offlinecomputations)thatwillbe abletorepresentaccuratelythesolutionsfortheproblemstudied.Differenttechniquesareusedtogeneratethisbasis,themore commonlyfoundintheliteraturearetheproperorthogonaldecompositionandthegreedysamplingapproach[2,5].Inbothcase, thenumberoftermsinthereducedbasisisassumedtobeverysmallcomparedtothenumberofdegreeoffreedomofthefinite elementcomputation.Then,theapproximatesolutionsoftheparametrizedproblemarecomputedviaperformingaGalerkin projectionontothereducedbasisspace(onlinecomputations).

However,theaccuracyoftheobtainedsolutionsdependsonthequalityofthemeshusedaswellasonthequalityofthe chosenreducedbasis.Ifwedenotebyμthevectorofparameters,theglobalerroregisdefinedforanyμby

eg

( μ )

=u

( μ )

urb

( μ )

, (1)

Corresponding author. Tel.: +33 140974865.

E-mail address: laurent.gallimard@u-paris10.fr (L. Gallimard).

(3)

whereu(μ)istheexactsolutionoftheparametrizedproblemandurb(μ)isitsreducedbasisapproximation.Thisglobalerror canbesplitintotwoparts:

eg

( μ )

=u

( μ )

uh

( μ )

+uh

( μ )

urb

( μ )

=eh

( μ )

+erb

( μ )

(2)

whereuh(μ)isthefiniteelementsolutionoftheparametrizedproblemanderb(μ)=uh(μ)urb(μ)(resp.eh(μ)=u(μ)uh(μ))istheerrorduetotheprojectionofthefiniteelementsolutioninthereducedbasisspace(resp.theerrorduetothefinite elementapproximation).Withintheframeworkandreducedbasisapproximationsbasedonproperorthogonaldecomposition orgreedysamplingmethods,manyworkshavebeendevotedtothecomputationofaposteriorierrorestimatorstomeasurethe errorduetotheprojectionofthefiniteelementsolutioninthereducedbasisspace.Anerrorestimatorisproposedforelliptic partialdifferentialequationsin[1,6],forparabolicproblemsin[7,8],forcomputationalhomogenizationin[9],forstochastic computations[10].However,theproposederrorestimatorsarefocusedontheestimationoftheerrorduetothereducedbasis approximationerbandassumethattheerrorduetothefiniteelementapproximationehisnegligible.Withintheframeworkof thepropergeneralizeddecomposition,aglobalerrorestimatorbasedontheconceptoferrorintheconstitutiverelation[11], hasbeenrecentlyproposedfortransientthermalproblemsin[12],andforlinearelasticproblemsin[13].Thiserrorestimator requirestodevelopadoublereducedbasisapproximationduringtheofflinestep,akinematicapproachandastaticapproach, forsolvingtheparametrizedproblem.

In thispaper, we focus on parametrized linear elastic models where the parametric bilinear form a is dependingon parameters-dependentfunctionsinanaffinemanner.Theobjectiveofthispaperistoextendtheconstitutiverelationerror estimatortoreducedbasisapproximationsbasedongreedysampling.Theuseoftheconstitutiverelationerror(CRE)requires thecomputationofanadmissiblepair(uˆ,σˆ)foranyparameterμduringtheonlinecomputations.Unliketheapproachproposed in[12,13],weusedirectlytheinitialreducedbasis(i.e.thereducedbasisusedtocomputethesolutionurb)tobuildtheadmissible pair(uˆ,σˆ)andwedonotneedtocomputeasecondreducedbasisbyagreedysamplingalgorithm.Additionally,twoerrorindi- catorsaredevelopedtoseparateintheglobalerrorestimatorthepartoftheerrorduetothefiniteelementapproximationfrom thepartduetothereducedbasisapproximation.Thisfamilyoferrorestimatorsallowstoconstructerrorboundsoftheenergy normandhasbeenappliedtoseparatethecontributionofthedifferentsourcesoferrorinfiniteelementcomputationsfornon- linearproblems[14–16],fordomaindecompositionproblems[17,18].Weshowthattheglobalerrorestimatorandthereduced basiserrorestimatorareupperboundsofthecorrespondingexacterrors.Tocomputeefficientlytheerrorestimateswithinthe frameworkofanoffline/onlinereducedbasismethodweneedafurtherassumption,andweassumethatthecomplementary energyis,aswellasthebilinearforma,dependingonparameters-dependentfunctionsinanaffinemanner.

Thepaperisorganizedasfollows:InSection2,weintroduceparametricproblemtobesolved,webrieflyrecallthereduced basismethodology,andwedefinetheapproximationerrorsintroduced.Theformulationoftheglobalerrorestimator,thefi- niteelementerrorindicatorandthereducedbasiserrorindicator,aswellastheoffline/onlinestrategytocomputethem,are describedinSection3.Finally,inSection4thedifferenterrorsareanalyzedthroughanumericalexample.

2. Problemtobesolved 2.1. Linearelasticmodel

Letusconsideranelasticstructuredefinedinadomainboundedby.Theexternalactionsonthestructurearerepre- sentedbyasurfaceforcedensityTdefinedoverasubsetNoftheboundaryandabodyforcedensitybdefinedin.Weassume thataprescribeddisplacementu=udisimposedonD=

N.Thematerialisassumedtobelinearelastic,beingCthe Hooketensor.WeconsiderthattheproblemisdependentofasetofparametersμD⊂Rp.Theproblemcanbeformulated as:findadisplacementfieldu(μ)Uandastressfieldσ(μ)definedinwhichverify:

thekinematicconstraints:

u

( μ )

=udon

D (3)

theequilibriumequations:

div

σ ( μ )

+b

( μ )

=0in

and

σ ( μ )

n=T

( μ )

in

N (4)

theconstitutiveequation:

σ ( μ )

=C

( μ ) ε (

u

( μ ))

in

(5)

ndenotestheoutgoingnormalto.U isthespaceinwhichthedisplacementfieldisbeingsought,U0thespaceofthe fieldsinU whicharezeroonD,andε(u)denotesthelinearizeddeformationassociatedwiththedisplacement:[ε(u)]i j= 1/2(ui,j+uj,i).

Thestrongformoftheproblem(3-5)isequivalenttotheclassicalweakformformulation:findu∈{vU; v|

D=ud}such

that:

a

(

u

( μ )

,u;

μ )

=f

(

u;

μ )

uU0 (6)

(4)

where

a

(

u

( μ )

,u;

μ )

=

C

( μ ) ε (

u

( μ ))

:

ε (

u

)

d

and f

(

u;

μ )

=

b

( μ )

.ud

+

N

T

( μ )

.ud

Following[1],weassumethattheparametricbilinearformaandtheloadingfareaffinelydependentoffunctionsoftheparam- eterμ,bywhichweshallmean:

a

(

u,u;

μ )

= Q

q=1

q

( μ )

aq

(

u,u

)

andf

(

u;

μ )

=

Q

q=1

Hq

( μ )

fq

(

u

)

(7)

whereq(μ)(q∈{1,… ,Q})andHq(μ)(q∈{1,… ,Q})areknownfunctionsofμ,aq(q∈{1,… ,Q})arebilinearformsindependent ofμandfq(q∈{1,… ,Q})arelinearfunctionsindependentofμ.

Remark. This‘affine’decompositionhasbeensuccessfullyappliedbyRozzaetal.fordifferentkindofparametrization:geometry, loadintensityanddirection,materialproperties,andmultisubdomainsformodularstructures(see[2]formoredetails).

2.2.Finiteelementapproximation

Tocomputethesolutionu(μ)ofEq.(6),weintroduceafiniteelementapproximationuhofuinafiniteelementspaceUh. Thefinite-dimensionspaceUhisassociatedwithafiniteelementmeshofcharacteristicsizeh.LetPhbeapartitionofinto elementsEk.Thispartitionformedbytheunionofallelements,isassumedtocoincideexactlywiththedomainandanytwo elementsareeitherdisjointorshareacommonedge.WeassumethatudcanberepresentedbyadisplacementfieldinUh.The discretizedproblemis:finduh∈{vUh;v|

D=ud}suchthat:

a

(

uh

( μ )

,uh;

μ )

=f

(

uh;

μ )

uhUh0 (8) whereUh0={vUh;v|D=0}.

Thecorrespondingstressfieldiscalculatedusingtheconstitutiveequation:

σ

h

( μ )

=C

( μ ) ε (

uh

( μ ))

(9)

2.3.Reducedbasisapproximation

LetudirUhbeadisplacementfieldsuchthatudir|

D=ud.LetusintroduceasetofsamplesintheparameterspaceSNs=

{μ1,...,μNs},whereμiD,andforeachμncomputeafiniteelementsolutionu0h(μn)inUh0describedbythecorresponding vectorofnodalvaluesqn.

a

(

u0h

( μ

n

)

,uh;

μ

n

)

= f

(

uh;

μ

n

)

a

(

udir,uh;

μ

n

)

uhUh0 (10) Thereducedbasisspaceisthendefinedby

Urb0 =span

{

u0h

( μ

1

)

,. . .,u0h

( μ

Ns

) }

Uh0 (11)

ThechoiceofthesamplesintheparameterspaceSNsandoftheassociatedreducedbasisUrbdependsonthesamplingstrategy (see[1]formoredetails).Inthispaper,asdetailedinSection2.5,weusetheGreedyapproachproposedin[1,2].Thereduced basisapproximationconsistsinsolvingEq.(6)inUrb0+{udir}.Akeypointtojustifytheuseofthereducedbasisapproximationis thatNsisassumedtobemuchsmallerthanNfe(i.e.Ns<<Nfe).

a

(

u0rb

( μ )

,urb;

μ )

= f

(

urb;

μ )

a

(

udir,urb;

μ )

urbUrb0 (12) Remark. Thecomputationofudirisperformedoffline.Thesimplestchoiceis,foragivenμ¯ ∈D,tofindudir∈{vUh;v|

D=ud}

suchthat

a

(

udir,uh;

μ

¯

)

=0

uhUh0

Thecorrespondingstressfieldiscalculatedusingtheconstitutiveequation:

σ

dir=C

( μ

¯

) ε (

udir

)

(13)

σdirisequilibratedtozerointheFEsense:

σ

dir:

ε (

uh

)

d

=0,

uhUh0

Letusdenotebyirbthedisplacementfieldsu0h(μi)fori∈{1,...,Ns}.Thereducedbasissolutionwrites urb

( μ )

=udir+

Ns

i=1

α

i

irb (14)

(5)

Thecorrespondingstressisdefinedusingtheconstitutiveequation

σ

rb

( μ )

=C

( μ ) ε (

udir

)

+

Ns

i=1

α

iC

( μ ) ε (

irb

)

(15)

byintroducingEq.(14)inEq.(12)webuildthealgebraicsystem

[K

( μ )

][

α

]=[F

( μ )

] (16)

Theelementsof[K(μ)]andof[F(μ)]aredefinedby

Ki j

( μ )

=a

(

irb,

rbj;

μ )

andFi

( μ )

=f

(

irb;

μ )

(17)

Thankstothe‘affine’decompositionofaandf,KijandFicanbewrittenasalinearcombinationofthefunctionsq(μ)(q∈{1,… ,Q})andHq(μ)(q∈{1,… ,Q}).

Ki j

( μ )

= Q

q=1

q

( μ )

aq

(

irb,

rbj

)

andFi

( μ )

=

Q

q=1

Hq

( μ )

fq

(

irb

)

(18)

Acrucialpointinreducedbasisapproximationsistheseparationofthecomputationalprocedureintwoparts:anofflinepart devotedtothecomputationofparametersindependenttermsandperformedonlyonce,andanonlinepartdevotedtothe computationofparametersdependenttermsandperformedmanytimes.Thecomputationalcomplexityoftheofflinestageis greatanddependsonNfe(thesizeofthefiniteelementapproximation),whilethecomputationalcomplexityoftheonlinestage issmallanddependsonNs,Q,Q.Wesummarizethetwopartsofthecomputation(amoredetaileddescriptioncanbefoundin [1])

offline:Nsfiniteelementsolutionsirb=u0h(μi)aswellasthescalarsquantitiesaq(irb,rbj )andfq(irb)arecomputedand stored.

online:thematrix[K(μ)]andtherighthandside[F(μ)]areassembledfromEq.(17),thesystem[K(μ)][α]=[F(μ)]issolved andthedisplacementfieldurb(μ)iscomputedfromEq.(14).

AlltheonlineoperationsareindependentofdimensionNfeanddependonlyonNs,QandQ. 2.4. Approximationerrors

Twoapproximationerrorsareintroducedinthecomputationofthereducedbasisapproximation

anerrorduetotheprojectioninthereducedbasisspace,

anerrorduetothefiniteelementapproximation.

Theglobalerroregisdefinedby

eg

( μ )

=u

( μ )

urb

( μ )

(19)

Thisglobalerrorcanbesplitintwoparts

eg

( μ )

=eh

( μ )

+erb

( μ )

(20)

whereehanderb arerespectivelytheerrorintroducedbythefiniteelementapproximationandtheerrorintroducedbythe reducedbasisapproximation.

eh

( μ )

=u

( μ )

uh

( μ )

anderb

( μ )

=uh

( μ )

urb

( μ )

(21)

Theenergynormisaclassicalwaytomeasuretheseerrors

e

( μ )

2μ=a

(

e

( μ )

,e

( μ )

;

μ )

=

C

( μ ) ε (

e

( μ ))

:

ε (

e

( μ ))

d

(22)

where• =g,horrb.Thefollowingequalityholdsfortheenergynorm:

eg

( μ )

2μ=

eh

( μ )

2μ+

erb

( μ )

2μ (23)

2.5. Choiceofthesnapshots

Thechoiceoftheelementsuh(μi)ofthereducedbasis(oftencalledsnapshots)isperformedfollowingtheGreedyapproach

proposedin[1,2].WedenotebyDtrainDthesetofthesampleswhichwillbeusedtogenerateourreducedbasisapproximation andNtrainthenumberofelementsofDtrain.TypicallyNtrainischosensuchthatNs<<Ntrain.Thegenerationofthesnapshotsis performedbyaniterativegreedyprocedure.Webeginwithafirstpointμ1andafirstreducedbasisspaceUrb1 =span{u0h(μ1)}, then,forN=2,...,Ns,wecomputeurb(μ)inUrbN−1+{udir}forallμDtrainandwefind

μ

N=argmaxμ∈D

train

erb

( μ )

μandu0h

( μ

N

)

=uh

( μ

N

)

udir (24)

setSN=SN−1μN,andupdateUrbN =UrbN1+span{u0h(μN)}.

(6)

Remark. Forpracticalpurposeitisnecessarytoreplaceerb(μ)μbyaninexpensiveaposteriorierrorboundasproposedin[1]. Inthispaper,weuseanerrorboundbasedontheconceptoferrorintheconstitutiverelationthatwillbedevelopedinSection (3.5).

3. Errorintheconstitutiverelation

Inthissection,weproposeanerrorestimationmethodbasedontheconceptoferrorintheconstitutiverelation.Weshow thatthiserrorestimatorisanupperboundoftheerrorintheenergynorm.Animportantpointtodevelopanefficienterror estimatorforreducedbasisapproximationsisthatthecostofconstructionoftheerrorestimatorsisindependentonNfe,and weneedtointroduceasafurtherassumptionthatthecomplementaryenergyisalsoaffinelydependentoffunctionsofthe parameterμ.

C

( μ )

1

σ

1:

σ

2d

=

Q

q=1

qσ

( μ )

Sq

σ

1:

σ

2d

(25)

whereqσ(μ)areknownscalarfunctionsandSqarefourthordertensorfieldsdefinedon. 3.1. Definitionoftheerrorestimator

Theapproachbasedontheconstitutiverelationerror[11]reliesonapartitionoftheequationsoftheproblemtobesolved intotwogroups.Inlinearelasticity,thefirstgroupconsistsofthekinematicconstraintsandtheequilibriumequations;the constitutiveequationconstitutesthesecondgroup.ForagivenμD,letusconsideranapproximatesolutionoftheproblem, denotedby(uc,σe),whichverifiesthefirstgroupofequations(3)and(4).Thissolutionwillbesaidadmissible.If(uc,σe)verifies

theconstitutiveequation(Eq.(5))inthen(uc,σe)=(u(μ),σ(μ)).If(uc,σe)doesnotverifytheconstitutiveequation,the qualityofthisadmissiblesolutionismeasuredbytheerrorintheconstitutiverelation

η

gwhichisdefinedwithrespecttothe

constitutiveequation

η

g=ecr

(

uc,

σ

e

)

=

| σ

eC

( μ ) ε (

uc

)) |

μ (26) where

| σ

eC

( μ ) ε (

uc

)) |

μ=

( σ

eC

( μ ) ε (

uc

))

:C1

( μ )( σ

eC

( μ ) ε (

uc

))

d

1/2

(27)

ThePrager–Syngetheoremgivesarelationshipbetweentheexactandadmissiblesolution[19]

e2cr

(

uc,

σ

e

)

=

| σ

σ

e

|

2μ+

uuc

2μ (28)

3.2.Admissiblesolution

Akeypointtodeveloptheerrorestimatoristheconstruction,foragivenμD,ofanadmissiblesolution(uc,σe)fromthe

reducedbasissolutionurb(μ)andthedata.Sincethereducedbasissolutionverifiesthekinematicconstraints,onetakes:

uc=urb

( μ )

in

(29)

Amethodtorecoverequilibratedstressfieldsσefromthefiniteelementsolutionandthedatahavebeenunderdevelopment forseveralyears[19–21].Inclassicallinearelasticityfiniteelementcomputation,theadmissiblestressfieldcanbeconstructed directlyfromthefiniteelementstressfield[19]becausethefiniteelementstressfieldverifiestheequilibriumequationsofthe finiteelementmodel.Inthecaseofareducedbasistheconstructionofanadmissiblestressfieldσerequirestwodistinctsteps.

Inafirstofflinestep,thereducedbasisUrbisusedtobuildanadmissiblereducedbasisforthestresses.Thisreducedbasisis builtwitharecoverytechniqueforconstructingequilibratedstressesinstarpatchesproposedin[20].

Inasecondonlinestep,theadmissiblestressiscomputedbyaminimizationofanenergynorm.

3.3.Constructionofanadmissiblestressfield

Thefirststep(offline)consistsinbuildinganadmissiblereducedbasisforthestresses.

Letμ0denoteaparticularsetofparameterssuchthatμ0/{μ1,...μNs},andletusintroduceasetofdisplacementfields uqh(μ0)Uh0,forq∈{1,...,Q},suchthat

a

(

uqh

( μ

0

)

,uh;

μ

0

)

= fq

(

uh

)

uhUh0 (30) wherefqistheqthelementoftheaffinedecompositionoftheloading(Eq.(7)).Thestressfieldsσqf=C(μ0)ε(uqh(μ0))verifyan equilibriumequationintheFEsense(Eq.(31))

σ

qf :

ε (

uh

)

d

= fq

(

uh

)

uhUh0 (31)

(7)

Foreachq∈{1,...,Q},arecoverytechnique[20]isusedtobuildastressfieldσˆqfthatverifiestheequilibriumequationdefined bythefollowingequation:

σ

ˆqf:

ε (

u

)

d

=fq

(

u

)

uU0 (32) Similarly,astressfieldσˆdirequilibratedtozeroisbuiltfromthestressfieldσdir(Eq.(13))suchthat

σ

ˆdir:

ε (

u

)

d

=0

uU0 (33) Thesetofstressesσˆqf isusedtobuildanadmissiblestressfieldσˆf(μ)foranyloadingf(u;μ).Letusdefinethestressfield σˆf(μ),whichdependsexplicitlyonμby

σ

ˆf

( μ )

=

Q

q=1

Hq

( μ ) σ

ˆqf+

σ

ˆdir (34)

whereHq(μ)areknownfunctionsofμdefinedinEq.(7).IfweintroduceEq.(32)inEq.(7),thankstothe‘affine’decomposition, weobtain

uU0 f

(

u;

μ )

=

Q

q=1

Hq

( μ )

σ

ˆqf:

ε (

u

)

d

=

Q

q=1

Hq

( μ ) σ

ˆqf :

ε (

u

)

d

(35)

Hence,thestressfieldσˆf(μ)isadmissiblefortheloadingf(u;μ).Similarlyσf(μ)definedbyEq.(36)isanequilibratedstress fieldforthefiniteelementmodel.

σ

f

( μ )

=

Q

q=1

Hq

( μ ) σ

qf+

σ

dir (36)

Thestressfieldσˆf(μ)couldbeuseddirectlytocomputeanupperboundoftheerror,howeveritiseasytoimprovethisupper boundadaptingthetechniqueproposedin[12]forapropergeneralizeddecompositionmethod,toareducedbasisapproach.Let usconsiderthesetofstressfieldscomputedfromthesnapshotsolutionsUrb

σ

irb=C

( μ

i

) ε (

uh

( μ

i

))

fori

{

1,...,Ns

}

(37)

ThestressfieldσirbisequilibratedintheFEsensefortheloadingf(uh;μi).Therecoverytechnique[20]isagainusedtocompute admissiblestressfieldsσˆirb(i∈{1,...,Ns})suchthat

σ

ˆirb:

ε (

u

)

d

=f

(

u;

μ

i

)

uU0 (38)

ForafixedμiEq.(34)leadsto

σ

ˆf

( μ

i

)

:

ε (

u

)

d

= f

(

u;

μ

i

)

uU0 (39)

BysubtractingEq.(39)fromEq.(38)weobtain

σ

ˆirb:

ε (

u

)

d

=0

uU0 (40)

whereσˆirb=σˆirbσˆf(μi)(i∈{1,...,Ns})isasetofstressfieldsequilibratedtozero.Similarly,wecanbuildasetofstressfields equilibratedtozerointheFEsensebysettingσirb=σirbσf(μi).

Thesecondstepoftheconstructionisperformedduringtheonlineprocedure.Foreachparameterμanadmissiblestress

fieldσeissoughtinthereducedbasisofequilibratedstressfields

σ

e=

σ

ˆf

( μ )

+

Ns

i=1

β

i

σ

ˆirb (41)

FromEqs.(40)and(34)itfollowsthatσeisequilibratedwiththeloadingf(u;μ)forall

β

i∈R.

Thecoefficients

β

iarecomputedinordertominimizethe“distance” betweenthestressfieldcomputedinthereducedbasis

σrb(μ)=C(μ)ε(urb)andastressfieldequilibratedintheFEsenseσ˜(

β

1,...,

β

Ns)

J

( β

1,...,

β

Ns

)

=

|

C

( μ ) ε (

urb

)

σ

˜

( β

1,...,

β

Ns

) |

μ (42) where

σ

˜

( β

1,...,

β

Ns

)

=

σ

f

( μ )

+

Ns

i=1

β

i

σ

irb (43)

(8)

TheminimizationofEq.(42)leadstothealgebraicsystem [A

( μ )

]

β

=[G

( μ )

] (44)

where Ai j=

C1

( μ ) σ

rbj :

σ

irbd

(45)

and Gj=

σ

rbj :

( ε (

urb

)

C1

( μ ) σ

f

( μ ))

d

(46)

Remark. TheequilibratedstressfieldσecouldalsobeusedinEq.(42)insteadoftheweaklyequilibratedstressfieldσ˜,toobtain abettereffectivityindexfortheglobalerrorestimator.However,theuseoftheweaklyequilibratedstressfieldleadstominimize theerrorestimatoronthereducedbasisapproximation(asshowninEq.(61))whichisusedtobuildthesnapshots.

3.4.Computationoftheerrorestimatorandupperboundproperty

AdirectapplicationofthePrager–Syngetheorem(Eq.(28))withuc=urbandσedefinedbyEq.(41)leadstothefollowing

upperboundpropertyfortheenergynorm

eg

( μ )

μ=

uurb

μ

η

g=ecr

(

urb,

σ

e

)

(47)

FromEq.(26)theexpressionoftheglobalerrorestimatorisgivenby e2cr

(

urb,

σ

e

)

=

C1

( μ ) σ

e:

σ

ed

+

C

( μ ) ε (

urb

)

:

ε (

urb

)

d

2

σ

e:

ε (

urb

)

d

(48)

Thecomputationoftheglobalerrorestimatorisdonebyanoffline–onlineprocedure

Duringtheofflineprocedurethestressfieldsσˆqf(q∈{1,...,Q})andσˆirb(i∈{1,...,Ns})arecomputedandstored,aswell asthefollowingscalarquantitiesfor(i,j)∈{1,...,Ns},p∈{1,...,Q}andr∈{1,...,Q}

ari j=

Sr

σ

irb:

σ

rbj d

a0r p j=

Sr

σ

0p:

σ

rbj d

ci j=

σ

irb:

ε (

rbj

)

d

(49)

and ˆ ari j=

Sr

σ

ˆirb:

σ

ˆrbj d

ˆ

a0r pi=

Sr

σ

ˆpf:

σ

ˆirbd

ˆ

a00r pq=

Sr

σ

ˆpf:

σ

ˆqfd

bˆsi j=

Cs

ε (

irb

)

:

ε (

rbj

)

d

cˆi j =

σ

ˆirb:

ε (

rbj

)

d

ˆ

c0pi=

σ

ˆpf :

ε (

irb

)

d

(50)

Duringtheonlineprocedureforeachparameterμ,thankstothe‘affine’decomposition(Eq.(25)),wecomputethecoefficients ofmatrix[A(μ)]and[G(μ)]byintroducingthecoefficientsdefinedbyEq.(49)inEqs.(45)and(46):

Ai j=

Q

r=1

qσ

( μ )

ari jandGj=

Ns

i=1

α

i

( μ )

ci j

Q

p=1 Q

r=1

Hp

( μ )

rσ

( μ )

a0r p j (51)

then,thesystem[A(μ)] β

=[G(μ)]issolved.

Theerrorestimatorecriscomputedduringtheonlineprocedurefromthecoefficients

β

i,andthecoefficientsinEq.(50)are computedduringtheofflineprocedure.

e2cr

(

urb,

σ

e

)

=

Q

r=1 Q

p=1 Q

q=1

rσHpHqaˆ00r pq+

Q

r=1 Q

p=1 Ns

i=1

rσHp

β

iaˆ0r pi

Références

Documents relatifs

In this paper, we focus on the control of the error intro- duced by a reduced basis surrogate model on the computation of the failure probability obtained by a Monte Carlo

In this paper, we have proposed a study of the enhanced error &#34; enh on the constitutive relation in plasticity and shown that the calculation of this error estimator in order

PGD driven by the Constitutive Relation Error Minimal CRE/PGD Pierre-Eric Allier, Ludovic Chamoin, Pierre Ladevèze.. To cite this version: Pierre-Eric Allier, Ludovic Chamoin,

A posteriori error estimator based on gradient recovery by averaging for discontinuous Galerkin methods.. Emmanuel Creusé,

KEY WORDS : finite element method; equilibrated stress recovery; constitutive relation error; error bounds; goal-oriented error

The essential compo- nents are (i ) (provably) rapidly convergent global reduced-basis approximations – Galerkin projection onto a space W N spanned by solutions of the

2 INRIA Rocquencourt, MICMAC Team-Project, Domaine de Voluceau, B.P.. In this stage, a certification of the approximation is possible by means of an a posteriori error bound.

tiple reciprocity method [21, 22J which allows to transform the volume integral corresponding to a body force into an infinite series of boundary integrals using