人工智能论文20220716 2022PhD CMU Deep Learning Methods for Catalyst Surface and Interface Structure Analysis.pdfVIP

DeepLearningMethodsforCatalyst

SurfaceandInterfaceStructureAnalysis

Submittedinpartialfulfillmentoftherequirementsfor

thedegreeof

DoctorofPhilosophy

in

DepartmentofChemicalEngineering

JunwoongYoon

B.S.ChemicalEngineering,UniversityofCalifornia,Berkeley

M.S.MachineLearning,CarnegieMellonUniversity

CarnegieMellonUniversity

Pittsburgh,PA

May2022

©JunwoongYoon,2022Allrightsreserved.

Abstract

Theincreaseinglobalenergydemandandraisedenvironmentalcon-

cernshavemotivatedthedesignofnovelmaterialsforenergy-relatedap-

plications.However,thedesignofever-complicatingmaterialsforemerg-

ingenergytechnologiesiscurrentlybottleneckedbylimitedresourcesto

understandcomplexsurfaceandinterfacestructuresandpropertyrela-

tionships.Inthefirstpartofthisthesis,wedevelopatandemframework

thatcombinesamolecularthermodynamictheoryandmoleculardynam-

icssimulationsinanattempttoinvestigatesolidinterfacialphenomena

andtodiscusshowdeeplearningmethodscanimprovetheframeworkas

anextstep.Inthesecondpart,wedevelopasetofdeeplearningmeth-

odsthatsolvevariousmaterialsandcatalystdesignproblemsincluding

property,structure,andstabilityanalysis.Wepresentagraphneural

networksarchitecturetolearntheoptimalrepresentationsofheteroge-

neouscatalysissystemsfortheaccuratepredictionofadsorption/binding

energies.Thenweextendedtheapproachtoapproximateground-state

structuresofthecatalysissystemsbyincorporatingdifferentiableopti-

mizationmethodsintothegraphneuralnetwor