OnthePropertiesofSolutionsoftheAdjointEulerEquationsMichaelB.GilesNilesA.PierceOxfordUniversityComputingLaboratoryAbstractThebehaviorofanalyticandnumericaladjointsolutionsisexaminedforthequasi-1DEulerequations.Forshocked ow,thederivationoftheadjointproblemrevealsthattheadjointvariablesarecontinuouswithzerogradientattheshockandthataninternaladjointboundaryconditionisrequiredattheshock.AGreen’sfunctionapproachisusedtoderivetheanalyticadjointsolutionscorrespondingtoisentropicandshockedtransonic ow,revealingaloga-rithmicsingularityatthesonicthroatandcon rmingtheexpectedpropertiesattheshock.Numericalsolutionsobtainedusingbothdiscreteandcontinuousadjointformulationsrevealthatthereisnoneedtoexplicitlyenforcetheadjointshockboundarycondition.Adjointmethodsaredemonstratedtoplayanimportantroleintheerrorestimationofintegratedquantitiessuchasliftanddrag.1IntroductionAdjointproblemsarisenaturallyintheformulationofmethodsforoptimalaerodynamicdesignandoptimalerrorcontrol.Fordesignapplications,theadjointsolutionprovidesthelinearsensitivitiesofanobjectivefunctionsuchasliftordragtoanumberofdesignvariableswhichparameterisetheshape.Thesesensitivitiescanthenbeusedtodriveanoptimisationprocedure.Considerablee orthasbeendedicatedtothedevelopmentofoptimaldesignmethodsbasedonthisapproach[1{8].Morerecently,adjointmethodshavebeenrecognisedasameansofachievingerrorcontrolin uiddynamicssimulations[9{12].Inthiscontext,theadjointsolutionrelatesthesensitivityoftheobjectivefunctiontothelocaltruncationerrorsinthe owdiscretisation.Thisinformationcanthenbeusedtoprovideanaposteriorierrorestimateortoguideanadaptivemeshingalgorithm.Whilesigni cante orthasbeendedicatedtodevelopingpracticalmethodsbasedonad-jointformulations,therehasbeenlittlediscussionofthepropertiesoftheadjointsolutionsthemselves[13].Thepresentworkinvestigatesvariousissuesconcerningthederivationandapproximationofsolutionstothequasi-1DadjointEulerequations.ThestandardLagrangemultiplierderivationofJameson[1]isextendedtoincludethee ectofshocksintheformu-lationoftheanalyticadjointequations.ExplicitinclusionofthesteadyRankine{HugoniotconditionsviaanadditionalLagrangemultiplierdemonstratesthatattheshock,theadjointvariablesarecontinuousandthataninternaladjointboundaryconditionisrequired.Thisisconsistentwithacharacteristicviewpointwhichindicatesthatoneinternaladjointb.c.isneededduetothedisparityinthenumberofadjointcharacteristicsenteringandleavingtheshock.However,theconclusionsdi erfromthoseofpreviousinvestigators[14{16].Thediscreteadjointequationscanbeformulatedintwoways,eitherbydiscretisingtheanalyticadjointequations(theso-called‘continuous’approach)[1],orbytransposingthedis-creteequationsobtainedbylinearisingthediscretised owequations(the‘discrete’approach)12Giles&Pierce[7].Gileshaspreviouslyshownthatforquasi-1D owswithshocks,aconservativediscreti-sationwhichissecond-orderaccurateinsmoothregionsofthe owproducesasecond-orderaccurateapproximationtothe‘lift’integral[17].Hence,thelinearisationofsuchamethod(onwhichthediscreteadjointisbased)mustproducealinearisedliftperturbationwhichisatleast rst-orderaccurate.Ontheotherhand,itislessclearwhetherthediscretisationoftheanalyticadjointequationsleadstothecorrectadjointsolutionifthereisnoexplicitenforcementofthespecialshockcondition.Toinvestigatethispoint,thepaperderivestheanalyticsolutiontotheadjointequationsforshocked ow.ThisisaccomplishedbyconstructingtheGreen’sfunctionsforthelinearisedEulerequations,includingthelinearisedRankine{Hugoniotconditions,usinganextensionoftheapproachdevelopedbyGilesandPierceforshock-freequasi-1D ows[13].Theanalyticresultscompareverywellwithnumericalresultsobtainedusingboththecontinuousanddiscreteapproaches.Tounderstandwhythecontinuousapproachbehavescorrectlywithoutexplicitenforcementoftheadjointshockboundarycondition,ashootingmethodwasusedtomarchthesolutionbackfromtheexitacrosstheshock.Disregardingtheadjointshockb.c.,butmaintainingcontinuityattheshock,leadstoafamilyofsolutionsoftheadjointequations.Ofthese,itappearsthatthecontinuousapproachselectsthesmoothestmemberofthefamily,whichcorrespondstotheanalyticsolution.The nalsectionofthepaperdiscussestheuseofadjointsolutionsforerroranalysis.Theerrorintheliftintegralisshowntobeaninnerproductofthetheadjoint owvariablesandthetruncationerrorofthediscretisationoftheEulerequations.Estimatingthetruncationerrorgivesamethodofaccuratelyestimatingtheerrorintheliftintegral.Witha rst-orderdiscretisationoftheEulerequations,itisshownthattheerrorestimatecanbeusedtocorrectthecomputedvalueoftheliftintegralandobtainsecond-orderaccuracy.Alternatively,theerrorestimatecouldbeusedinthefutureasthebasisforoptimalgridadaptation[12].2AdjointproblemformulationThequasi-1DEulerequationsforsteady owinaductofcross-sectionh(x),ontheinterval 1 x 1,maybewrittenasR(U;h) ddx(hF) dhdxP=0;whereU=0@ q E1A;F=0@ q q2+p qH1A;P=0@0p01A:Ifthesolutioncontainsashockatxs,theRankine-Hugoniotjumpcondition[F]x+sx s=0connectsthesmoothsolutionsoneitherside.Fordesignapplications,linearisationofRwithrespecttoperturbationsinthe