搜索
1AControlSystemForPhotolithographicSequencesbySovarongLeang,Shang-YiMa,JohnThomson,BartBombayandProf.CostasSpanosDepartmentofElectricalEngineeringandComputerSciences,UniversityofCaliforniaatBerkeleyAbstractThegoalofourcontrolsystemistoimprovethereliability,accuracy,andeconomyofopera-tionofasequenceofinterrelatedprocesses.Weachievethistaskbyusingwellknown,rigorousstatisticaltechniquestocontinuouslymonitorprocessparameters,detectout-of-controlequip-ment,andthenoptimallyadjustrelevantmachineinputstobringtheprocessbackontarget.WehaveimplementedthesupervisorycontrolsystemonthephotolithographysequenceintheBerke-leyMicrofabricationLaboratory,whereithasbeenconclusivelyproventhatthesupervisorycon-trolsystemincreasessignificantlythecapabilityoftheentireprocess.Thesupervisorycontrolalgorithms,consistingoffeedbackandfeed-forwardcontrol,multivariate,model-basedstatisticalprocesscontrol(SPC),andautomatedspecificationmanagementalgorithms,areindependentofmachineand/orprocess,andcanbeappliedtoanysemiconductormanufacturingsequence.1IntroductionTostaycompetitive,semiconductorindustriesmustdevelopefficient,highyieldingmanufac-turingfacilities.Onewaytoincreaseyieldistoreduceprocessvariability.Thisbecomesaratherdifficulttask,sincesemiconductorprocessesoftendriftduetoequipmentaging,depletionofchemicals,fluctuationinambientconditions,oraftermachineshavebeenreplacedormaintained.Oneapproachtorelievetheproblemistodevelopasupervisorysystemthattunesprocessesdur-ingproduction.Suchasystemtakesadvantageofmodernanalyticalandprocessingequipment2thathavetheabilitytointeractwithcomputerdrivencontrollers,collectinformationandmanipu-laterecipestocompensateforprocessdrifts.Thispaperdescribesthedevelopmentandthedeploymentofsuchacontroller.Althoughheu-risticalgorithmsforcontrolhavebeenreportedinthesemiconductorindustry[1],wehavechosentobaseourapproachonformalstatisticalmethods.Statisticallybasedalgorithmsofferseveraladvantagesoverheuristicapproaches,sincetheycanbeadaptedtoalargenumberofprocessesand,onceinplace,arerobustenoughtobeusefulinanactualmanufacturingenvironment.Considerableworkhasbeendonetoformalizethisprocedure,includingtheMITrun-by-runController[2],Ultramax™[3],andTexasInstrument’sPCCController[6][4][5].ThefirstversionoftheMITRun-by-RunControllerallowsthecontrolofoneprocessparam-eterbyadaptivelychangingtheconstanttermofalinearprocessmodel[2],andlaterversionsintegratemoregeneralmodeladaptationandmultivariateapplications[7].Ultramaxisacommer-cialsoftwareforsequentialprocessoptimizationandprocesscontrolthatcanhandlemultipleinputsandoutputs.Althoughthedetailsofitsoperationareproprietary,Ultramaxusesavariantoftheevolutionaryoperationalgorithm(EVOP)[8]tofindtheoptimumoperatingpoint.Ultra-maxoffersthesignificantadvantagethatnopriormodeloftheprocessisrequired;however,itdoesrequirecontinuouschangesontheprocessinordertoderivesuchamodel.TexasInstru-ment’sMMSTcontrolarchitectureprovidesagoodframeworkwithinwhichcontrolanddiagnos-ticsystemshavebeenimplemented[6][29].ThePCCControllerisanexampleofanefficientrun-to-runcontrollerthatcanhandleprocessdrifts.ThestructureofourcontrollerisinmanywayssimilartothatofthePCCController.Themaincontributionofourcontrolalgorithmsisthattheyaredesignedforsequencesofmultipleinterrelatedprocesses,whichareoftencontrolledassingleprocessesinindustry.Althoughthestatisticaltechniquesusedinourcontrollerarenotnovel[8],themannertheyareusedforprocesscontrolisnew,andresultsinimprovedcapabilityoftheentireprocesssequence.Thecontrol3algorithmspresentedinthispaperalsohavetheadditionaladvantagethattheyoffermultivariatecontrolandcomplexadaptationofnon-linearmodels,withoutintroducingextraneousvarianceintheprocess[9][10][11][12][13].Ourdemonstrationvehicleisthephotolithographyprocesssequence.Theroleofthecontrol-leristoprovideanintelligentsystemforgeneratinginitialprocessrecipes,correctingprocessdrifts,anddetectingequipmentorprocessmalfunctionsonarun-by-runbasis.Inlaterwork,uponthedetectionofaprocessdriftormalfunction,adiagnosticsystemlinkedtothecontrollerwillofferaneducatedguessofthecauseoftheproblem.Wehaveinvestigatedtwoprocesscontrolapproachesformulti-processsequences.Thefirstonekeepstightcontrolovereachmachineintheprocesssequence.Whentheoutputsofamachinedrift,thecontrollerimmediatelygeneratesanewrecipetobringthembackontarget.Thesecondprocesscontrolapproach,ontheotherhand,keepsthefinaltargetoftheprocesssequencefixed,whileintermediatetargetsaresubjecttodynamicadjustments.Thepaperisstructuredasfollows.First,wepresentinsection2themonitoredphotolithogra-phyparametersandtheirrespectivemodels.Theninsection3,wediscussthedetectionofprocessdisturbances.Sections4and5presentthetwocontrolapproaches,namelythelocalcontrolandthesupervisorycontrolmethodologies.Finallyinsection6,wecompareusingexperimentsthetwocontrolapproachestoeachotherandtoanuncontrolledprocesssequence,runonfixedreci-pes.2PhotolithographyEquipmentModelsTocharacterizethestateofthewaferbetweeneachphotolithographystep,wehavechosentomonitorthefollowingparameters:resistthickness(Tres)andphot