combining traditional modelling, model order reduction and

36
Combining traditional modelling, model order reduction and big data Wil Schilders, TU Eindhoven (Netherlands) 4TU.AMI symposium, June 8, 2018

Upload: others

Post on 26-Mar-2022

4 views

Category:

Documents


0 download

TRANSCRIPT

Combiningtraditionalmodelling,modelorderreductionandbigdata

WilSchilders,TUEindhoven(Netherlands)

4TU.AMIsymposium,June8,2018

Aimoftalk

• Sketchaglobalpicture− Relationsbetweenphysical

modeling,modelorderreductionandtheuseofdata

• Startwithaspecificexamplefromtheelectronicsindustry,owingtotheworkthatwasdoneatPhilipsandNXPSemiconductors

• Discusssomemajorchallengesanddevelopments

• Presentthewayforward

PAGE 2

GOINGBACKINTIME30YEARS

PAGE 3

Semiconductordevices

• Semiconductordevicesarethebuildingblocksofallelectronicsproducts

− Resistors,Capacitors,Inductors,Diodes

− Bipolartransistors− MOStransistors(metaloxidesemiconductor)

PAGE 4

SemiconductordevicesimulationatPhilipsResearch• Semiconductordevicesaredescribedbyusing2typesofparticles:electronsandholes(=“absenceofelectron”)

• StartingpointforthemodellingistheBoltzmannTransportEquation(BTE),forbothparticles(“n”referringtoelectrons,“p”toholes):

PAGE 5

The BTE need to be solved, in 6-dimensional phase space, for the density functions f

àvery time consuming…alternative?

Themethodofmoments

• MomentsoftheBTEhaveaphysicalmeaning:thezeroordermomentrelatestotheconcentrationofparticles

PAGE 6

• Thefirstordermomentrelatestothecurrentdensity:

Usingthemethodofmomentsuptoorder1,wearriveatthedrift-diffusionmodel

• Thedrift-diffusionmodelconsistsofthePoissonequationfortheelectricpotential:

PAGE 7

• Plusthecontinuityequationsforholesandelectrons:

• Andconstitutiverelationsforthecurrentdensities:

Usingthemethodofmomentsuptoorder1,wearriveatthedrift-diffusionmodel

• Thedrift-diffusionmodelconsistsofthePoissonequationfortheelectricpotential:

PAGE 8

• Plusthecontinuityequationsforholesandelectrons:

• Andconstitutiverelationsforthecurrentdensities:

D is the doping profile, specific for the device

Usingthemethodofmomentsuptoorder1,wearriveatthedrift-diffusionmodel

• Thedrift-diffusionmodelconsistsofthePoissonequationfortheelectricpotential:

PAGE 9

• Plusthecontinuityequationsforholesandelectrons:

• Andconstitutiverelationsforthecurrentdensities:

The recombination R and the mobilities μare parameters to be determined/given

Modelsformobilityandrecombination

• AtPhilipsResearch,therewasagroupofphysicistsandelectronicengineersworkingconstantlyonnewmodelsformobilityandrecombination

• Alsoatothercompanies,suchgroupsexisted,andtheywerecompetingonaworldwidescale

PAGE 10

Howwerethesemodelsconstructed?

• ThePhilipsgroupperformedmanyexperimentsandmanysimulations

• Thenusedphysical/electronicinsight,curve-fitting,interpolationandothermethodstocomeupwithamodelthat

− Hadincreasedfunctionality− Improvedtheaccuracyofsimulations− Broughtexperimentalresultsandsimulationsclosertogetherintheprocessofon-goingminiaturizationofsemiconductordevices

PAGE 11

Hierarchyofmodelsformobility

• Simpleinitialsimulationsusingconstantmobilities• Modelsdescribinglatticescattering(particlesinteractwiththeatomsinthesemiconductorlattice)

• Ionizedimpurityscattering(interactionswiththeionizedimpuritiesà doping)

• Carrier-carrierscattering(electronsandholeswitheachotherandwiththedifferentspecies)

• Effectsofultra-highconcentration• Velocitysaturation,fielddependence,…

PAGE 12

Similar story for the recombination R

Electroniccircuits

• Semiconductordevicesmaybethebasicbuildingblocks,alevelhigherwefindtheelectroniccircuitsthatconsistofhundreds,thousandsandoftenmillionsofsemiconductordevices

PAGE 13

Electroniccircuitsimulation

• Simulatingthebehaviorofanelectroniccircuitbyusing(coupled)semiconductordevicesimulationforalldevicesinthecircuitisanimpossibletask

• Fortunately,notalldevicesaredifferent,oftenonlyafewtypesoftransistorsareused

• Forthesedevices,so-calledcompactdevicemodelsareconstructed,inawaythatisverysimilartoconstructingmodelsformobilityandrecombination:

− Performmanymeasurementsandsimulations− Usephysical/electronicinsight,discardcertainphenomena

• Bigdifference/complication:− Themodelmustbeparameterized!− ModelslikeMEXTRAMcontain50parameters

PAGE 14

PAGE 1522 March 2011

Developmentsinlast15years

• Besidesthetransistormodelsdevelopedinthemajorsemiconductorcompanies,theso-calledBSIMmodelsweredevelopedatBerkeley

• Thesemodelsareconstructedinanautomaticway,byusingalsomanyexperimentalresultsaswellasresultsfromsimulations

• Majorquestions:− cantheworkdonebythespecializedgroupsbereplacedbyautomaticproceduresgeneratingadvancedmobilityandrecombinationmodels,orevencompletedevicemodels?

− Can“insight”beincorporatedinsomeway?

PAGE 16

Robin Bornoff

Delphi4LEDTask 2.5 – Model Calibration

Market Development ManagerMechanical Analysis Division

January, 2018

Restricted © 2017 Mentor Graphics Corporation

Summary of Calibration Studies –Royal Bluen Cost Function a

quantification of the difference between measured and simulated SFs

Your Initials, Presentation Title, Month Year18

Restricted © 2017 Mentor Graphics Corporation

Summary of Calibration Studies –Royal Bluen 100 (Computational) Design of

Experiments set and solved

Your Initials, Presentation Title, Month Year19

Restricted © 2017 Mentor Graphics Corporation

Summary of Calibration Studies –Royal Bluen Global minima

identified from the cost function response surface— Simulated for

verification

Your Initials, Presentation Title, Month Year20

Lessonslearnedintheelectronicsindustry

• Modellingofsemiconductordevicesandelectroniccircuitsisdoneinamixedway:

− Usingphysical/electronicinsight,severaleffectsaredescribedbypartialandordinarydifferentialequations,aswellasbyalgebraicrelationsthatmimicthephysicaleffects

− Manymeasurementsandsimulationsareperformedinordertoproducetablesofvaluesforparametersinthemodels

− Compactdevicemodelsareneededtoperformelectroniccircuitsimulationinanefficientway– thesemodelsarealsoconstructedasacombinationofphysicaleffectsandparameterextraction

• Physicalinsightmustbeusedtoreducethecomplexityofmodels,butdataarenecessarytomakesuchmodelsaccuratedescriptionsofreality

PAGE 21Automatic via MOR?

RECENTDEVELOPMENTS

PAGE 22

PAGE 23January 2010

PAGE 24January 2010

Product Lifecycle Management

PAGE 25January 2010

PAGE 26January 2010

PAGE 27January 2010

PAGE 28January 2010

PAGE 29January 2010

PAGE 30January 2010

PAGE 31January 2010

PAGE 32January 2010

PAGE 33January 2010

CONCLUSION

PAGE 34

Theparticularapproachdependsupontheexistingstateofsimulationand/orproxymodelimplementation

PAGE 35January 2010

Importanttostart/joinnewinitiatives

• WerecentlygotourMSCAEIDproposalBIGMATHfunded:“BigDataChallengesforMathematics”(4years,7PhDstudents,8academic/industrypartners)

• NWA(NationaleWetenschapsAgenda):callwaspublishedbyNWOonMay24

− 2-stepprocedure,1st proposalinSeptember

• NWOCross-Over:calltobepublishedsoon• WeproposedatopicforFETProactive:“DigitalTwinsforIndustryandInnovation”,outcomeexpectedsoon

− Ifsuccessful,~20Meuroavailableforcallsin2020PAGE 3622 March 2011