人工智能论文20220806 2022PhD Northwestern Accelerating Materials Study and Discovery through Machine Learning.pdfVIP

NORTHWESTERNUNIVERSITY

AcceleratingMaterialsStudyandDiscoverythroughMachineLearning

ADISSERTATION

SUBMITTEDTOTHEGRADUATESCHOOL

INPARTIALFULFILLMENTOFTHEREQUIREMENTS

forthedegree

DOCTOROFPHILOSOPHY

FieldofMaterialsScienceandEngineering

By

CheolWooPark

EVANSTON,ILLINOIS

March2022

2

Abstract

Despitetheunprecedentedpaceatwhichtechnologyisadvancingnowadays,itswider

useandapplicationareoftenlimitedbytheavailabilityoflow-cost,high-performancematerials.

Thedevelopmentofbettermaterialsisoftencostlyandtime-consumingduetothesheersizeof

thesearchspaceofanewmaterial.Recently,data-drivenmaterialsdesignthroughmachine

learningtechniqueshavegainedmuchinterestforitspotentialtoacceleratethematerials

discoveryandcharacterizationbyordersofmagnitudecomparedtoconventionalcomputational

methodssuchasdensityfunctionaltheory(DFT).Thisthesisisfocusedonthedevelopmentof

machinelearningalgorithmsfortwoobjectives.Thefirstobjectiveistoacceleratethediscovery

ofnewstablematerialsthroughhigh-throughputstudiesbyaccuratelyscreeningforpotential

stablecompounds.Towardsthisend,wedevelopedtheiCGCNNandvi-iCGCNNwhichutilize

crystalgraphrepresentationsofinorganicmaterialsthatcanbeusedasinputforneuralnetworks.

Weshowausecaseofthevi-iCGCNNinpredictingalternativesuperhardmaterials.Thesecond

objectiveistoacceleratematerialcharacterizationthroughmoleculardynamicsimulations.To

achievethis,we