Computationally Guided Searches for E ffi cient Catalysts through ChemicalMaterials Space Progress and Outlook - Griego et al. - 2021 -

《Computationally Guided Searches for E ffi cient Catalysts through ChemicalMaterials Space Progress and Outlook - Griego et al. - 2021 -》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

pubs.acs.org/JPCCFeatureArticleComputationallyGuidedSearchesforEfficientCatalyststhroughChemical/MaterialsSpace:ProgressandOutlook††CharlesD.Griego,AlexM.Maldonado,LingyanZhao,BarbaroZulueta,BrianM.Gentry,EliLipsman,TaeHoonChoi,andJohnA.Keith*CiteThis:J.Phys.Chem.C2021,125,6495−6507ReadOnlineACCESSMetrics&MoreArticleRecommendationsABSTRACT:Computationalquantumchemistrypromisestohelpguidethedesignofcatalyststhataremoresustainableandeconomical.ThisFeatureArticlegivesatutorialoverviewofhowourgroupaccountsforthethermodynamicsandkineticsofchemicalreactionsincomplexenvironments.Westartwithexplanationsofhowtoincludeenvironmentalcontributionswhenmodelinghomogeneousandheterogeneouscatalyticprocesses.Wealsoprovideexamplesofschemesthatusemachinelearningandalchemicalperturbationdensityfunctionaltheorythateschewhighcomputationalcostswhileprovidingusefulinsightsintochemicalreactionmechanisms.Withthistoolboxofcomputationalmethods,wehighlightprogressinunderstandinghowtoreliablymodelrenewableenergycatalysisreactionmechanismsthatoccurincomplexenvironments.5■INTRODUCTIONenvironment.Alternatively,heterogeneousmechanismswilloccurataninterfaceoftwo(ormore)differentphases,forCatalysisplaysacriticalroleinsocietybyproducingfuelsand1example,asolid/liquidinterface(SEI).Ifthesolidphasehereisvalue-addedchemicals,butthecontinuallyevolvingsocio-aconductor,thenthestandardgeneralizedgradientapprox-economicandenvironmentallandscapesrequireadvancesto2−4imation(GGA)-basedKohn−Shamdensityfunctionaltheorymakecatalyticprocessesmoresustainable.Scienceand(DFT)mightbeadequate.However,GGA-basedfunctionalsengineeringhavetraditionallyusedknowledge,intuition,andalsomightnotbetrustworthy,includingthecaseofCOingenuitytoguidecatalystdesign,butweareinanexcitingadsorptionontransition-metalsurfacesbeingimpreciselyepochwherecomputationalcodesandalgorithmsarebecoming6,7describedbyPerdew−Burke−Ernzerhof(PBE).Additionally,faster,moreaccurate,andbetterautomated.Thismeansthatstudyingthewater-gasshiftreaction(WGSR)oncopperwithanDownloadedviaUNIVOFCONNECTICUTonMay16,2021at10:28:55(UTC).moreandmorehypotheticalcatalystscanbecomputationallyincreasinglycomplexDFTfunctionalcanresultindifferentscreenedtofindthemostpromisingcandidatesthatwarrant8kineticpredictionsoftheWGSR.Incasessuchasthese,andifconsiderationinthearduousandcostlyprocessofexperimentalSeehttps://pubs.acs.org/sharingguidelinesforoptionsonhowtolegitimatelysharepublishedarticles.thesolidphaseinanSEIisahighlycorrelatedsemiconductor,synthesis,testing,andimplementation.thenhigherlevelcalculationssuchasDFT+U,othermethodsIncatalysisapplications,computationalchemistryiswieldedfromtheJacob’sladderofDFTfunctionals,oroneofavarietyoftopredictquantitativetrendsinhowlocalchemicalbondinganddifferentembeddedDFTtheoriesmaybeneededtophysicallysolventeffectsinfluencethethermodynamicsandkineticsofdescribethecorrectstate.Additionally,studiesoftheWGSRonreactionsteps.TherequiredrobustnessofthemodelingcopperwithincreasinglycomplexGGAexchange-correlationultimatelydependsonthecomplexityofthesystem.Thefunctionalscanresultinvaryingkineticpredictionsoftheproverbialzooofmultiscalemethods,modelchemistries(i.e.,WGSR,includingthedeterminationofthesurfacemechanism.levelsoftheoryandbasissets),andassortedkeywordscanmakeIncasessuchasthese,thesolidphaseintheSEIisahighlyitalmosttooeasyfornovicestouseamethodtoobtainacorrelatedsemiconductor;then,higherlevelcalculationssuchaspreconceivedresult,butforexperts,thiszoorepresentsagalleryofornatemethodsthataccountfordifferentphysicochemicalinteractionswithinmoleculesandmaterialsinhierarchicalReceived:December21,2020degreesofphysicalrigorandcomputationalexpense.Revised:January28,2021Dependingontherelevantphysicsofthesystem,homoge-Published:February19,2021neousreactionsmightbesufficientlymodeledwithinavacuumwithinapolarizingcontinuumsolventmodel,ortheymayrequireamorecomplexandlocallyheterogeneoussolvating©2021AmericanChemicalSocietyhttps://dx.doi.org/10.1021/acs.jpcc.0c113456495J.Phys.Chem.C2021,125,6495−6507

1TheJournalofPhysicalChemistryCpubs.acs.org/JPCCFeatureArticle9−12DFT+U,othermethodsfromtheJacob’sladderofDFTthatleverageasmallsetofQCcalculationstocreatemuchmore1338−40functionals,oroneofavarietyofdifferentembeddedDFTusefuldatawithoutsacrificingmuchaccuracy.14−22theoriesmaybeneededtophysicallydescribethecorrectIntheory,APDFTprovidesmanyadsorbateBEsbyanstate.Besidesthechoiceofmodelchemistryfortheelectronicapproximatedrelationshipofhowelectrostaticpotentialsinastructure,thesolventphasecouldbemodeledusingoneormorereferenceadsorbate−catalystsystemandtheBEchange(ΔBE)well-orderedsolventlayersorpseudoamorphousblobsofuponacompositionalchange,thatis,analchemicaltrans-solventmolecules(perhapsincludingionsandcounterions),mutation.ThisprocedureisillustratedinFigure1,whereathesolventinteractionsmightbetreatedwithimplicit23−26models,orsolvationmaybeneglectedaltogether.Thenatureofatomswithinthelocalenvironmentaroundhomogeneousorheterogeneouschemicalstepsalsodependsontheoutershellenvironmentalfactors:thechemicalpotentialsofprotons,electrons,electricfields,andanypotentiallymass-transferlimitedspecies.Thereadmittedlyisnotaconsensusaboutthe“correct”waytomodelalloftheseeffects;however,whatevertheapproachchosen,computationalmodelsmustprovideusefulinsightintoadvancingtheunderstandingandguidingthedesignofnewandimprovedtechnologies.Oneofourgroup’sprimaryinterestsisunderstandinghowtobestmodelchemicalreactionsonacomputerandnavigatethevastnessofchemicalandmaterialspaceforimproveddesign.Figure1.IllustrationofathermodynamiccyclethatdepictstheEvenwiththeadvancesincomputationalquantumchemistryenergeticpathwaysofadsorption,theatomictransmutations,andhow27−31thesetransmutationsimpacttheadsorption.Wecommonlyrefertothis(QC)andmoreefficientcalculationplatforms,thetypeofcyclewhenconsideringAPDFTpredictionsofBEsonalloychemicalandmaterialsspaceissotremendouslymassivethatsurfacesthatwerehypothesizedfromareferencematerial.PathwaysareQCcalculationsonallpossiblecandidatesshouldbeconsideredcomposedoftheBEsofanadsorbateonasurface(horizontallegs)andintractableforquitesometime.Withcatalysts,thevastnessofatomictransmutations(verticallegs).ΔE0|andΔE0|denotetheλ=0λ=1chemicalspacearisesfromdifferenthypotheticalatomBEsforthetopandbottomhorizontallegs,respectively.ΔEsandλ→1configurationsatactivesites,theirsurroundingligands,andΔEadenotetheenergychangesassociatedwiththeatomicλ→1theirdegreesofstrainaswellasenvironmentalaspectssuchastransmutationfortheleft(s=surface)andright(a=ads-site)verticalsolvation,pH,andionicstrength.Thusthereisrecentinterestinlegs,respectively.Reprintedwithpermissionfromref41.Copyrightdevelopingandapplyingcost-efficientandapproximate2020JohnWiley&Sons.methodsbasedonmachinelearning(ML)orotherphysical32theories.Thevalidityofthesemethodscanbeassessedwithstraightforwardfundamentalapplicationstudiesthatconsiderthermodynamiccycledepictstheenergeticchangesofadsorbatemetricsfortheirreliabilityincatalysisapplications,forexample,bindingtoasurfaceafteranatomictransmutation.Thepredictedbindingenergies(BEs)ofreactionintermediates,hypotheticalenergycontributionarisingfromthisalchemicalactivationenergiesofchemicalreactions,orthermodynamictransmutationisapproximatedasaTaylorseriesdescriptorssuchasacidityconstantsandstandardredox0002102potentials.Asfollows,wediscussourprogressinunderstandingΔEEE|=Δ|+∂ΔΔ+∂ΔΔ+λλ==10λλλλE...2howtomodelthesefactors.(1)whereΔE0|istheenergyoftheadsorbatebindingonthe■λ=1MATERIALSSCREENINGWITHALCHEMYhypotheticalsystemresultingfromthealchemicaltrans-In2011,theUnitedStateslaunchedtheMaterialsGenomemutation.ThisisequaltotheBEonthereferencesystemInitiative(MGI)tocreatenewinfrastructurefortherapid(ΔE0|)plusalchemicalderivativetermsthatarewrittenas33,34λ=0computationalpredictionandscreeningofnovelmaterials.n0∂ΔλE,wherenistheorderofthederivativeandΔλ=1.Catalysisresearchershavefollowedsuitandpursuedsimilar40−5635InpastapplicationsofAPDFT,thisexpressioniseffortswiththecatalystgenome.Themaintargetofinteresttypicallytruncatedtojustthefirst-orderderivativetohasbeentheatomicallyprecisenatureoftheactivesite,whichisapproximatethechangeinBEbetweenthetwostatesascharacterizedinpartbythelocalcoordinationofatomswithinthesurfacefacet.Usefuldescriptors,suchasthereactingΔ=BEΔ|−EEEEE0Δ|=0Δ−aΔ=s0∂ΔΔλλλλλ==→→1011λadsorbateBEtoasurfacesiteandthereactionstepenergetics,(2)provideastraightforwardwaytoassessthekineticsofsaheterogeneouscatalyzedreactions.Indeed,thereareconcertedwhereΔEλ→1andΔEλ→1aretheenergychangesassociatedwitheffortstodevelopdatabasesoftheseproperties,suchasanalchemicaltransmutationdonetoabaresurfacemodel(s)Citrination,36CatalysisHub,37andtheMaterialsProject,33toandasurfacemodelwithanadsorbate(a),respectively.Whennameafew.thealchemicaltransmutationismade,thenuclearchargeofanHowever,withmorepromisingclassesofmaterialsbeingatom(NI)isalteredbyanintegeramount(ΔZ),resultinginandiscoveredexperimentallyandcomputationally,thecommunityenergychangeequaltotheenergygradientsofthenuclearmustadoptproceduressuitableforsystematicstudies.GGAchemicalpotential(ΔμnI)withrespecttothisvariationintheNI.methods,ingeneral,aresuitableformanysystems,butsomeWhenalchemicaltransmutationsaredoneisoelectronically(therequirehigherlevelsoftheorythatbottleneckhigh-throughputnumberofelectronsinthesystemisconserved)andtheatomicscreening.Ourgrouphasbeeninvestigatingcomputationallypositionsremainthesame,ΔBEisequaltoasimplifiedfirst-efficientalchemicalperturbationDFT(APDFT)approachesorderderivative6496https://dx.doi.org/10.1021/acs.jpcc.0c11345J.Phys.Chem.C2021,125,6495−6507

2TheJournalofPhysicalChemistryCpubs.acs.org/JPCCFeatureArticleΔBE=∂ΔΔ=EN0λμ∑Δ∂thesurface(Figure2).Withtheseprofiles,wealsopredictedEaλλnIII(3)forall32pathwaysthatagreedwithDFTcalculationswithin0.3eV.Withthisapproximation,simplearithmeticmanipulationsinvolvingtheelectrostaticpotentialsofareferencecatalystmodelareusedtopredicttheBEsofhypotheticalmaterialswithminimalcomputationalcost.Foramoredetailedexplanationoffirst-orderapproximations,wedirectthereadertoourrecentarticle,whichoffershands-onresourcesthatallowusersto41performAPDFTanalyseswithJupyterNotebooks.HereweoverviewourrecentworkwithAPDFTfortheheterogeneouscatalystandofferourperspectiveonthefutureimplicationsofthismethod.InourfirstpublishedworkonAPDFT,weshowedthatfirst-orderapproximationsarequiteaccurateforhigh-throughputBEFigure2.EnergyprofilesfortheCH4dehydrogenationmechanismon54hypotheticalalloysofPt.Herewecomparethereactionenergypathwaypredictionsfordopedactivesitesintransition-metalsurfaces.UsingoxygenreductionreactionintermediatesbindingonontheoriginalPt(111)referencesurface,whichwascalculatedwithhypotheticalalloysofPt,Pd,andNi,webenchmarkedAPDFT-DFT,andtheAPDFT-predictedreactionenergiesofthesamepathwaypredictedBEsagainstDFT-predictedvaluesandfoundthattakenover32transmutedvariationsofthereferencesurface.ThereferencepathwayonpurePtisdenotedwithredasterisks.Pathwaystheseestimatesagreedwithin0.1eV.withthemostsignificanteffectfromatransmutationwithanuclearInastudywhereweappliedAPDFTtoBEpredictionsonchargechangeΔZ=+1areshownindarkblue,pathwayswiththemostcarbides,nitrides,andoxides,wefoundthatAPDFTagreeswithsignificanteffectfromatransmutationwithanuclearchargechangeΔZDFTpredictionswithin0.33eVforrocksaltTiC(111),=−1areshownindarkgreen,andotherreactionpathwayscomputedTiN(100),andTiO(100)materials,whichdonotexhibitawithAPDFTareshowninalessvisiblelight-blue/green.Becauseofthe55bandgap.Conversely,wefoundthatAPDFThassignificantenergyprofilesbeingverysimilar,thereissignificantoverlapformanyshortcomingswithBEpredictionsonmaterialsbasedonalloys.Reprintedwithpermissionfromref41.Copyright2020JohnsemiconductorslikerutileTiO2(110),rutileSnO2(110),andWiley&Sons.rocksaltZnO(100).Ourhypothesiswasthatfirst-ordercorrectionsusingAPDFTwerebenefitedbyerrorcancellationWehavealsoinvestigatedprocedurestobuttressAPDFTpresentinconductivesystemsthataresubjectedtoelectronicmodelsusingΔ-MLprocedures,58showingthatthecomputa-screening.WetestedthishypothesisbyaddingPtdopantstotionaleffortindevelopingMLmodelsisjustifiedwhentheyaresurfacelayersinTiO2,whichdecreasedthereferencematerial’svalidfordatasetmagnitudeslargerthanthedatasetsusedforbandgap,andwefoundthathigheraccuracywasachievedwithtraining.56ByrankingtherelativeimportanceoftheinputthoseAPDFTpredictions.Wecontinuetosearchforamorefeaturestoourmodels,weidentifiedthevariablesthatprecisephysicalexplanationforwhyAPDFTischallengedbycontributedmosttolow-accuracyAPDFTBEpredictionsonthesesystemssothatitcanbeusedmoregenerallyincertaintargetalloys,thesizeofthetransmutation(ΔZ),andthesemiconductingandinsulatingsystems.numberoftransmutations.OnthebasisoftheworkofvonWearealsointerestedinunderstandingthekineticsofLilienfeldandcoworkersdemonstratingtheaccuratescreeningreactionsonsurfacesthataredictatedbytheenergybarrier(Ea)ofdeprotonationenergiesusingAPDFTwiththird-orderbetweentworeactionsteps.Astandardapproachforpredicting59corrections,weexpectthattheshortcomingsweobservetheEaforasurface-boundreactionistoemploythenudgedwithfirst-ordercorrectionswouldlikelybetreatablewithhigher57elasticband(NEB)algorithmthatinterpolatesimagesordercorrectionsandwouldthusallowtheaccuratescreeningofbetweentheinitialandfinalstatesofareactionstep.ThismoretargetalloysmadewithmultipletransmutationsoflargerproceduredoesnotrequireaHessian,buttheelectronicΔZ.Usingthird-orderalchemicalderivatives,asymmetricfiniteenergiesandforcesonallimagesarecalculated,thusmakingitdifferencesprocedurerequires1+2N+N2single-pointenergymoderatelycomputationallyexpensive.Theexpenseforeachcalculations,whereNisthenumberofsitesthatcouldbeNEBcalculationmakesitchallengingtousethisapproachtotransmuted.Foracatalystsurfaceslabmodelwith16determinemanyhypotheticalEavaluesfordifferentelementarytransmutablesites,onewouldneedonly289single-pointprocessesorforsystemsinvolvingdifferentsurfaceatoms.JustasenergiestoaccuratelyscreenthousandsofadsorbateBEsonwhencalculatingaBE,thenumberofrequiredcalculationsforhypotheticalalloys.WealsoanticipatethepossibilityofhighereachNEBpathwaywilllinearlyscalewiththenumberofbarriersordercorrectionstreatingtheshortcomingsweobservedwithwewanttopredict.semiconductingsystems.Withsecond-andthird-orderenergyWetackledthisissuebyusingoneNEBcalculationtoderivatives,systemsthatarenotsubjectedtoscreeningeffectsestablishanelementaryreactionpathwayasasetofreferencemaybebetterdescribed,andthesederivativesmaybeabletodataandthenAPDFTtogeneratemanyapproximateNEBaccuratelypointusinthedirectionofmorecomplexmaterialspathwaysbasedonref41.UsingthesamethermodynamiccyclewithhighercatalyticactivitywithoutalargerelianceonotherschemetoapproximatetheBEchangefollowinganalchemicalmoreexpensivemethodslikeDFT+Uorusingmorecomplextransmutation(showninFigure1),weapproximatedtheenergyexchangecorrelationfunctionals.Finally,inotherongoingwork,changeforalchemicallytransmutedtargetsystemsbasedonweareinvestigatinghowAPDFTprocedureswouldberelatedtoeachimagefromatraditionalNEBcalculationforCH4othertraditionaltheoreticalmodelstounderstandcatalyst60dehydrogenationonasurfaceofPt(111)thatcontained10descriptorssuchasNewns−AndersonHamiltonianmethods.images.Fromthis,weusedAPDFTtogeneratetheapproximateBecauseAPDFTbringshighcomputationalefficiency,weenergyprofilesofthispathwayon32transmutedvariationsofforeseeitmakingatransformativechangeinhowtoalleviatethe6497https://dx.doi.org/10.1021/acs.jpcc.0c11345J.Phys.Chem.C2021,125,6495−6507

3TheJournalofPhysicalChemistryCpubs.acs.org/JPCCFeatureArticlenecessityofrunningmanycalculationstounderstandthestablestatesofcatalystsunderambientreactionconditionsacrosschemicalspace.Catalystdescriptorsarenormallymodeledassumingidealconditionsonanidealizedsurfacefacet,butin61,62reality,materialsmaysuccumbtodefects,alloysegregations,reconstruction,ornoninnocentphasereconstructionwhensubjectedtoareactionenvironment.Infuturework,wewilltacklethesechallengeswithAPDFTschemescoupledtoothermethods,asdescribedinthefollowingsection,thatourgrouphasappliedtomodelambientcatalyststates.■MODELINGCATALYSTSUNDERAMBIENTCONDITIONS63Combiningthecomputationalhydrogenelectrodemodelwith64−68atomisticthermodynamicsallowsonetoconstructelectro-Figure3.DFTcalculationsconsideredtheinterplayofcoadsorbedchemicalphasediagrams(includingPourbaixdiagrams)thatspeciestorationalizeexperimentalobservationsofnanoporousPdXidentifystablerestingstatesofcatalystsunderambientreactionskinalloysasCO-poisoning-tolerantelectrocatalystsforCOreduction2conditions.WeshowedthatpurportedcatalystsforCO2toformate.Tabulatedabovearebindingenergies(inelectronvolts)onreductionhavedifferentandcomplementaryaccessiblestatesPd-skinnedPd3Xalloys(X=Co,Ni,Cu,Ag)forCObindingtoacleanthatwouldfacilitateenergeticallyefficientshuttlingofmultiplesurface(CO*-1),CObindingtoasurfacewithH*(CO*-2),COprotonsandelectrons.67,69Theproceduresoutlinedtherewouldbindingtoasurfacewith2H*(CO*-3),Hbindingtoacleansurfacebeusefulforidentifyingandunderstandingbioinspired(H*-1),HbindingtoasurfacewithH*(H*-2),andHbindingtoasurfacewithCO*(H*-3).Reprintedwithpermissionfromref71.orthogonalhydridetransfersthathavebeenidentifiedasa70Copyright2019AmericanChemicalSociety.promisingcatalystdesignstrategy.Wealsoshowed,inacollaborationwithSnyder’sgroupatxzDrexelUniversity,thataccountingformultiplethermodynami-C32++→N2HCN232−−xyxzHC+x+y(4)22callyaccessiblestatescanhelpdeconvolutecomplexexper-imentalobservations.71NanoporousPdXskinalloys(X=Co,(NotethatC32referstothecleanGBPmodel.)WethencalculatedtheGibbsenergy(G=EDFT+Gvib)ofeachspeciesinNi,Cu,andAg)werepresentedaselectrocatalyststhattheequationwiththeelectronicenergyfromDFT(EDFT)andproducedformatefromCO2withhighselectivityandavoidedvibG,whichiscomposedoftheenergycorrectionfromthezerodeactivationfromCOpoisoning.Amongthesealloys,Pd-skin/PdCowasfoundtobemostpromising,whichourgrouppointenergy(EZPE)andthevibrationalentropy(TSvib).The3free-energychangeofthereaction,asshownineq4,iswrittenastheoreticallyexplainedbyevaluatingstableconfigurationsofCOfollows,wherethechemicalpotentials(μ)ofatomicspeciesareandHbindingonthesurfacewithDFT.Moreover,wehadtoincludedcomputationallyreconcilethedestabilizationofbothCO(poisoningtolerance)andH(facilitatingCO2hydrogenationΔGG=−(C32−−xyxzNH)G(C)32+(xy+)(C)μμμ−x(N)−z(H)toformate)onthesesurfaces.BEcalculationswithcoadsorbed(5)specieswererun(showninFigure3),andtheresultsshowedNext,wecouldreferencethisenergywiththestandardhydrogenthatadsorbedCOandHwereboththemostdestabilizedonPd-electrode(SHE),andfinally,wegetthefree-energyexpressionskin/Pd3Co,regardlessofthecoadsorbedspecies.Again,simplefortheGBPreactionasafunctionofpHandappliedpotentialcalculationmodelsforunderstandingthecatalystactivitywould(U)below,whereΔμNisthethermodynamicdrivingforceforNbesignificantlyamplifiedbyusingaccurateAPDFTmodeling.bindingInotherwork,weusedelectrochemicalphasediagramstounderstandwhichmolecularandmaterialstateswouldbeΔ=GEDFT−EDFT+Δ−GxEvibijj1+ΔμyzzCN32−−xyxzHC32j2N2Nzrelevantunderdifferentelectrochemicalreactionconditions.Ink{67onestudy,electrochemicalphasediagramsillustratedthe++−()xyEzijj1μ(H)−2.303kTpH−UyzzstabilityofdifferentintermediatesontheSnO2(110)surfaceasaC2Bj2zk{functionofthethermodynamicdrivingforceforCO2bindingto(6)asurface(i.e.,ΔμCO)andtheelectrochemicalpotential(i.e.,U).2Toplotanelectrochemicalphasediagram,thefreeenergyofWeplottedandoverlaidthetheoreticalPourbaixdiagramovereachintermediateiscalculatedwitheq6assumingaconstantpHtheexperimentalPourbaixdiagramforCO2intermediatesintheenvironment,andthisgivesarelationbetweenΔGandΔμNforaqueousphasetoshowtheproximityofboundariesbetweeneachintermediate.Next,ΔGisconvertedtoareductionthem.ThisledustoproposereactionintermediatesthatwouldpotentialEredwithE=−ΔG/(nF)tohavearelationbetweenredexplaintheexperimentallyobservedoverpotentialsforCO2thereductionpotentialandthethermodynamicdrivingforce,reductionatmaximumFaradaicefficiency.wherenisthenumberofelectronstransferredandFisFaraday’sTocreatethesediagrams,wefirstcalculatedthefree-energyconstant.Finally,wecouldplottheboundarylinesofeachchangefordifferentreactionsinthesystem.Wereferbacktoourintermediateandfindthemostthermodynamicallystablestate69workontheelectroreductionofCO2onN-dopedgrapheneasatdifferentvoltagesandΔμNvalues.PlottingaPourbaixdiagramanexamplederivation.Weconsideredthefollowingchemicalcanbedonewithasimilarprocedure,wherewesetΔμNaszeroformulaforreducinggraphenebasalplanes(GBPs)inaqueousandfindouttherelationshipbetweenUandthepHforeachenvironmentaccordingtoatomisticthermodynamicsintermediate.Withthat,wedeterminethemoststable6498https://dx.doi.org/10.1021/acs.jpcc.0c11345J.Phys.Chem.C2021,125,6495−6507

4TheJournalofPhysicalChemistryCpubs.acs.org/JPCCFeatureArticleintermediateatdifferentUandpHvaluestogetthefinalreductionstartingwithspeciesA(oranoxidationendingwithPourbaixdiagram.speciesA).WebelievetheseatomisticthermodynamicsschemesshouldForexample,supposeacatalystwasexperimentallyfoundtobeandwillbeusedmoreoftenfortheoreticallycompletereduceCO2(AinFigure4),butitsactualmechanismwascatalystscreeningstudies.Forexample,thereareexcellentunclear,andcomputationaltheorywasneededforinsight.Fromopportunitiesforcombiningschemesforidentifyingthermody-standardQCcalculations,themostfavorablestateofCO2andnamicallyrelevantstructuresusinghigh-throughputgenerationsthecatalystwithrespecttothepHandtheappliedpotential(a72,73ofmicrokineticmechanismparameters,active-siteensem-Pourbaixdiagram)canbedetermined,asshowninFigure5.The74,75blesofmetastablestates,andmodelsbasedonactive-siteboundarylinesofthecalculatedPourbaixdiagramsshowunder76−78coordinationnumbers.ThisisnotdonenowbecauseitwhichconditionsofappliedpotentialandpHspeciesonwouldrequireverylargenumbersofQCcalculations,butthisoppositesidesoftheboundaryhavethesamechemicalpotentialwouldalsobealleviatedwithnewadvancesinAPDFTandML(i.e.,whenareactionfromonetotheotherwouldbringΔG=0).approaches.Thisisanimportantpointofreferenceforunderstandingreactionmechanisms,butreactionbarrierswouldalsoneedto■TOWARDMECHANISTICUNDERSTANDINGbeintroducedtomodelthereactionkinetics.BecausetheseAnygivenchemicalspeciesmayundergocountlessreactionreactionsverylikelycaninvolvetheparticipationofsolventmechanisms.Hydrogenationsareonesuchexample,andFiguremolecules,wenowturntohowourgroupapproachessolvated4shows23hypotheticalsteps(consistingofcovalenthydrogensystems.■MODELINGSOLVENTENVIRONMENTSWeconsiderthreemainclassificationsofsolventmodels:implicit(orcontinuum),mixedimplicit/explicit(orcluster−continuum),andexplicit.Allbringdifferentstrengthsandweaknesses,butusefulinsightscanbelearnedbycomparingresultsfromdifferentmodels.Implicitmodelsisabroadcategory,butitgenerallyreferstoanysolventmodelsthatdonotexplicitlyincludesolventmolecules,asshowninFigure6A.Manyareparametrizedtopredictsolvationenergiesbasedona80,81homogeneousdielectricmediuminteractingwiththesolute.Otherimplicitmodelsusemorecomplicatedsolute−solventorensembledescriptions.Theconductor-likescreeningmodelforrealsolvents(COSMO-RS),forexample,usesQCtodescribeFigure4.GeneralhydrogenationofAcanbeconsideredasaseriesofthesolutepolarizationofaconductor’scavitysurfaceandelementaryelectronandprotontransfers,proton-coupledelectronstatisticalthermodynamicstocomputethemolecularchemical82transfers,orformalhydridetransfers,andtheenergeticsofeachpotential.Thereferenceinteractionsitemodel(RISM)pathwaymayvarydependingontheenvironment.modelsthecorrelationfunctionsofsoluteandsolventmolecularsitestocomputethesolventdistributionsandthermodynam-83atomtransfers,stepwiseorcoupledprotonandelectronics.Thisisnotanexhaustivelist,andtherearemanymoretransfers,andformalhydridetransfers)thatmightwarrantimplicitmodelsandvariationsthateachhavetheirownconsiderationwhencomputationallyanalyzinganarbitraryapproximations,formulations,andapplications.Figure5.(A)Pourbaixdiagramshowingstablestatesofthereactant,CO2.(B)Pourbaixdiagramshowingstablestatesofahypotheticalmolecularcatalyst,1,10-phenanthroline.(C)OverlaidPourbaixdiagramsfrompanelsAandBshowingsimilarboundariesforhydrogenshuttlingandCO2reduction.VerticallinesrepresentpKavalues,horizontallinesrepresentthepH-independentstandardredoxpotentials,anddiagonallinesrepresentthepH-dependentproton-coupledelectron-transfersteps.Adaptedwithpermissionfromref79.Copyright2019WileyPeriodicals,Inc.6499https://dx.doi.org/10.1021/acs.jpcc.0c11345J.Phys.Chem.C2021,125,6495−6507

5TheJournalofPhysicalChemistryCpubs.acs.org/JPCCFeatureArticle92Thesmoothoverlapofatomicposition(SOAP)representa-93tionwithsketch-mapdimensionalityreductionwasusedtoquantitativelycomparestructuresfromanautomated,multistepfilteringprocedureofcandidatesolute−solventclustersofdifferentsizes.Wethencomparedthecloseness,ordistancebetweenpoints,todeterminethelocalsolventenvironmentsimilarityofthesolute−solventstructures.Oncelargersolute−solventclustersoverlapwithsmalleronesonthesketch-map,onecanassumethattheadditionalsolventmoleculesarefarenoughawaytohaveminimalimpactonthelocalsolventenvironment.Figure7showsthecaseofNa+hydration.Aglobal94optimizationcode,ABCluster,wasusedtogeneratehundredsFigure6.Varioussolventmodelingtechniquesareillustratedonafictitioussolute(purple)solvatedinwater.(A)Localandbulksolventinteractionscanbetreatedwithaimplicitsolventmodel;afictitioussolutecavityisshown.(B)Localsolventeffectsarecapturedwithexplicitsolventmolecules.Thelessimportantbulkcontributionsareefficientlydescribedwithanimplicitmodel.(C)Theentiresolventismodeledexplicitly.Thewell-knownproblemofimplicitmodelsistheirrespectiveapproximationsandparametrizationsthatdictatetheiraccuracy,reliability,andtransferability.ByincludingsomeexplicitsolventmoleculeswiththesoluteinQCcalculations,thesolventisessentiallyseparatedintolocal(inner)andbulk(outer)84contributions.Thistechniqueiscommonlycalledmixedimplicit/explicitorcluster−continuummodeling(Figure6B).Quasi-chemicaltheory(QCT)isaphysicallyrigorouswaytoseparatesolvationfreeenergiesintostatisticalcontributionsand85−89isthoroughlyexplainedelsewhere.Essentially,theexcesschemicalpotential(i.e.,molarsolvationfreeenergy)ofsomearbitrarysolute,X,inapuresolvent,L,isexpressedinQCTas(ex)(0)n(ex)(ex)μX=−kTBln(Knρμ)+kTBlnpnX()+[LXn−nμL](7)Thetermsineq7canbeconceptuallydescribedasdesolvatingnFigure7.(A)SOAP/sketch-maprepresentationsofsolute−solventindividualsolventmolecules(−nμ(ex)),associatingthesolute+LclusterscontainingNawithvariousnumbersofwatermolecules.Anandsolventmolecules(nL+X⇌(L)nX)intoacavitywithaexampleofaclusterwithfourwatermoleculesisshown.Thecolorbar(0)representsthenumberofwatermoleculesinthecluster.(B)solventdensity,ρ,andequilibriumconstant,Kn,intheideal-Boltzmann-weightedaverageofsolvationfreeenergiesofNa+witha(0)nvariablenumberofwaterligandsfromB3LYP-D3BJ/def2-SVPgasphase(−kBTlnKnρ),solvatingtheassociatedsolute−(ex)geometriesandωB97X-D3/def2-TZVPenergies.Anexampleofasolventcluster(μ),thenreleasingthegeometricconstraintLXnclusterwitheightwatermoleculesisshown.Adaptedwithpermissioninsidethecavity(kBTlnpX(n))wherepX(n)istheprobabilityoffromref5.Copyright2020AIPPublishing.observingnligandsinsolutionwithinapredefinedregion.Variouscontributionsineq7,specificallylnpX(n),requireaofsolventandsolute−solventclusters.Thefivelowestenergyprioriinformationonsolventcoordinationnumbersordynamicstructureswerethenoptimizedusingarelativelyinexpensivesimulationstoexplicitlyquantifythem.Alternatively,onecouldQCmethod,BP86-D3BJ/def2-SVP,andthesolute−solvent90usepreformedsolventclusters,asproposedbyBryantsevetal.clustersweremappedusingSOAP/sketch-map,asshowninThisclusterthermodynamiccyclecanbethoughtofasapplyingFigure7A.Onedotrepresentsasinglesolute−solventstructure.QCTtoboththesolventandsolute−solventclusterstoprovide(Anexamplecaseisgivenforfourwatermolecules.)Aerrorcancellation,allowingonetoforegoaprioriinformationBoltzmann-weightedaverageoftheclusterswasusedtorequirementsanddynamicsimulations.ThechallengewithcalculatethesolvationfreeenergyofNa+inwater.Onthetheseapproaches,however,isdeterminingthequantityandbasisofourprocedure,themostreliablestructurescamefrom12configurationsofsolventmoleculestobeused.Thiswouldwaterclusters,anddatacloselyagreedwiththeexperimentalinvolvemolecularsimulationsorcomparisonstoexperiment.solvationfreeenergy.OurapproachhasalsobeensuccessfullyTotacklethischallenge,weemployedanunsupervisedMLdemonstratedonionsolvationfreeenergiesspanning2−to2+91proceduretoidentifysolute−solventclustersthatresultincharges.Reiherandcoworkershaverecentlydevelopedasingle-ionsolvationenergiesthatappeartoconvergetowardsimilarcomputationalimplementationthatfocusesonrigor-91valuesinreasonableagreementwiththeexperimentaldata.ouslyachievingstatisticallyrelevantensemblesoflocalsolvent6500https://dx.doi.org/10.1021/acs.jpcc.0c11345J.Phys.Chem.C2021,125,6495−6507

6TheJournalofPhysicalChemistryCpubs.acs.org/JPCCFeatureArticle95moleculesinanautomatedmanner.Theirapproachisanthese,QCexplorationscanrigorouslydiscernpathwaystofindimprovedimplementationcomparedwithwhatwepublishedwhichwouldbemostlikely.Next,wewilloverviewourapproachpreviously,butitdoesnotmakeanexplicitconnectiontoQCT,andgeneralguidelinesformodelingsolvatedreactionwhichisausefulwaytoreducecomputationalexpensebymechanismswithminimaluserbias.leveragingimplicitsolventmodels.EventhoughtheirapproachBOMDsimulationshaveproventobeapowerfultoolfor105,106hasnotyetbeentestedforsingle-ionsolvationenergyanalyzingsolvatedreactionmechanisms,butsuchpredictions,thesimilaritiesofourandtheirapproachesindicatemethodsareusuallycostly,whichmakesthemusedlessoften.thattheirsshouldbeusefulandaccurateaswell.Usingmixedimplicit/explicitmodelingwithchain-of-statesAkeypointtoreiterateisthatcluster-continuummodeling,methodssuchasthenudgedelasticband(NEB)methodorwhendonecorrectly,canbeausefulandcost-effectivegrowingstringmethod(GSM)canbedoneonsmallersystems,calculationschemeforpropertypredictionsandmechanisticandthesewillbemoreamenabletohigherlevelsoftheory.On84,96,97studiesinsolvents.However,thedynamicsofthesolventthecontrary,thisbenefitoflowercomputationalcostcomes98areoftencrucialforaccuratepredictions,andwhenimplicitorwithitsowncomplexities.Forexample,wecomputationallymixedimplict/explicitmodelsarenotsufficient,theyshouldbeinvestigatedtheaqueoussodiumborohydride(NaBH4)modeledwithexplicitsolventmodeling(Figure6C),suchasreductionofcarbondioxide(CO)toformate(HCOO−)via299−101−Born−Oppenheimermoleculardynamics(BOMD)orahydride(H)transferusingseveraldifferentmodeling102−104107,108quantummechanics/molecularmechanics(QM/MM).approaches.Hydridetransfersarechargemigrations,Bothmethodsofferanunparalleledassessmentofthevastandonemightexpectthesetobesensitivetothelocalsolventconfigurationalspaceobservableinreactions.environment.Assuch,itmightbeexpectedtobecomputation-Accountingforradialandspatialdistributionfunctionsallydemandingtomodelthiskindofsystemusingexplicit(SDFs)inexplicitsolvationcanprovidemolecularinsightssolventmolecules.Thereactionpathwayenergeticsusingthisintothesolventeffectsofreactionsinelectrolytesolutions.modelingschemeisshowninFigure9(labeled“ExplicitFigure8showstheSDFsofourgroup’sBOMDsimulationsinFigure8.ComputationalinsightintothelocalsolventenvironmentofBH−liquidsimulationsof(a)pureHOand(b)7mol/LNaOH42(isovalueof48nm−3).Oxygenandhydrogendistributionsinthe3Dspatialdistributionfunctions(SDFs)areshowninorangeandwhite,Figure9.Free-energydifferencesandcomparisonofcontinuumandrespectively.Thecentralbluesphererepresentstheboronatom.cluster-continuumsolventmodelingofag-SSNEBpathwayofsodiumborohydridereductionofcarbondioxide.Differentsubsystemswerecraftedfromanexplicitg-SSNEBpathwaycontaining70watertwodifferentsolventenvironmentsandtheeffectsofhighbasemoleculeswithanimplicitsolventmodelreplacingtheremovedconcentrations.InpureH2O(Figure8a),theoxygenexplicitwatermolecules.107Thecontinuumsolventmodelwasunabledistributioninthefirstsolvationshellformsasphericalcagetocapturethequalitativetrendoftheg-SSNEBpathwayunlessthearoundBH−.Moreover,hydrogenhassomedistributionsinside4counterionandthefirstsolventshellwereincluded.ABOMDPMFthisoxygenshellthatresultintheformationofdihydrogencalculationat300Kquantifiedthefree-energydifferencesnotincludedbondsbetweenBH−andHO.However,in7mol/LNaOH,intheg-SSNEBenergies.10842theoxygendistributionhasnosignificantsphericalshape.Furthermore,thehydrogendistributioninsideoftheoxygenPMF”).108Becausethisapproachinvolvestheleastempiricism,shellhasasmallerdistributionrangecomparedwiththatinthethisisthereactionpathwayallotherlessexpensivecomputa-BH−liquidsimulations.Theseobservationsindicatethat4tionalschemeswouldideallyreproduce.increasedconcentrationsofNaOHresultinweakerinteractionsUsingthesamecollectivevariableastheexplicitpotentialofbetweenHOandBH−,andthismayhaveramificationsin24meanforce(PMF),weidentifiedastaticpathwayusingthereactionmechanismandkineticsstudiesinrealisticelectrolytegeneralizedsolid-stateNEB(g-SSNEB)109methodwhileenvironments.keepingtheexplicitsolventmodeledwithQC.107Thispathway,“ExplicitNEB”inFigure9,isessentiallythesameasthePMF■SOLVATEDREACTIONMECHANISMSwithoutanyfree-energy(entropic)contributions.Thequanti-ConsideragainthehydrogenationreactionnetworkinFigure4.tativesimilarityinenergyprofilesshowsthatentropiesalongthisWhereasalocalenvironmentcaninfluencewhichstatesarepathwaydonotsignificantlyimpactthebarrierforthisreaction.thermodynamicallyfavorable,eachfundamentalreactionstepThisisgoodtoknowbecauseavoidingdynamicssimulationsmayhaveasignificantenergybarriertoconsider.Experimentalwouldsignificantlylowerthecomputationalcostsformecha-observations(orstrongchemicalintuition)wouldhelpclarifynismsstudies.Forexample,insteadoffulldynamicsinanexplicitwhichpathwayswouldbelikely;however,intheabsenceofsolvent,onemighttreatthebulkcontributions(i.e.,everything6501https://dx.doi.org/10.1021/acs.jpcc.0c11345J.Phys.Chem.C2021,125,6495−6507

7TheJournalofPhysicalChemistryCpubs.acs.org/JPCCFeatureArticlebutthefirstsolventshell)usinganimplicitmodel.Theresultingsuppressingtheoxygenreductionactivity,andthiswaslaterreactionpathwaysareshowninFigure9.Wefoundthatcluster-confirmedbyexperimentinourpublication.Thus,inthiscase,continuummodelingwasactuallysuitableforqualitativelyReaxFFwasstillveryusefulforidentifyingsalientstructuresforcapturingthemetastableintermediate,thesecondtransitionreactions.state,andtheproduct.ThecounterionwasalsocrucialforHowever,whenweturnedtoanalyzesimilarreactiondescribingthereactionmechanismswiththestaticcalculations.mechanismsonmorecomplicatedTiAl2O5materials,wesoughtAsexpected,staticcalculationsbenefitfromcluster-continuumtodeterminewhetherthesematerialscouldformstablemodelingwhenafullsolventshellismodeled.amorphousstructures.AttemptstotrainaReaxFFmodelModelingmultistepreactionsbringsadditionalcomplexitiesusingautomatedprocedureswereunsuccessful(whichcouldbeaswell.PlataandSingletondemonstratedthatimplicitsolventduetoanynumberoffactors),soalternativeautomatablymodelsoftenfallshortofreliablycapturingthelocalsolventtrainedforce-fieldmethodsweresought.Ideally,BOMDwitheffectsofthefive-stepMorita−Baylis−Hillman(MBH)110on-the-flyelectronicstructurecalculationswouldbeuseful,butreaction.Liuetal.illustratedthatexplicitmethodswithcomputationalcostsareseverelylimiting.MLisnowapopularperturbationsolvationtreatmentsontopofhigh-levelwavealternativetoreactiveforcefieldsusedinmolecularfunctionmethods(i.e.,DLPNO−CCSD(T))canaccurately119−121111simulation.Neuralnetworks(NNs)haverecentlyreproducereactionenergeticsfromkineticsexperiments.becomepopularforlearningenergiesandforcesacrosschemicalTheirworkshowedthatpredictionsofcomplicatedreactionsarespace,oneexamplebeingtheBehler−ParrinelloNN(BPNN)possiblewithsignificantcomputationalresources,something122approach.TounderstandhowBPNNmethodscomparewiththatisnotalwaysavailableorrecommendedbeforeproperlyvettingcollectivevariablesusinglessintensivemethods.ReaxFFmethodstrainedusingthesameDFTdatasets(largely123Cluster-continuummodelingoftheborohydridemechanismbasedoffofpreviousdatasets),wecollaboratedwithabovereliedontakingstructuresoutofaperiodic,explicitlyKitchin’sgrouptostudyAubulk,surfaces,andclusterstotaling124solvatedchain-of-statecalculationderivedfromQMBOMD9972Kohn−Shamdensityfunctionalcalculations.Thesimulations.ThemultiplestepsoftheMBHmechanismmakeitoptimallytrainedBPNNpotentialwastrainedon9734challengingtoemploythismethodofdeterminingthesolventcalculations,whereasReaxFFrequiredonly848datapoints.shells.Instead,BP86-D3BJ/def2-SVP-optimizedstructuresBPNNoutperformedReaxFFinallcases,butthecomputationalfromABClusterwereusedtodescribethelocalsolventcostfortrainingtheBPNNwassubstantiallyhigher.Becausethe112environmentofeachMBHintermediate.Solute−solventexistingReaxFFparameterswerenotsufficientlyaccurateandclusterscontainingbetween0and10explicitmethanolBOMDsimulationsweretooexpensive,weturnedtogeneratingmoleculeswereconsidered.WeunexpectedlyobservedalackaccurateBPNNpotentialsusingKhorshidiandPeterson’sAMPofcorrelationbetweenmoremethanolmoleculesandthe125code.Wethencouldpredictthatamorphousstructuresofaccuracy,showingthataddingmoreexplicitsolventmoleculesTiAl2O5wereunlikelytoformeventhoughTiO2systemsdid,doesnotguaranteebetterpredictionsduetointrinsicerrorsinbutanalysesofdopantsonTiAl2O5systemswereinconclusivecalculationprocedures.withrespecttoexperimentalresultsfromourcollaboratorsatthe126NavalResearchLab.Thisshowedthatagreementbetween■REACTIVEATOMISTICMODELINGANDMACHINEexperimentandcomputationonthesemorecomplicatedLEARNINGsystemsisstillanopenquestion.Inparticular,thenextstepWehavestressedthatdifferentmodelingschemescanbeusedforthisworkmightbetoaccountforsolvationinteractionsatthedependingonthecomplexityofthesystemsandthecomputa-TiAl2O5/electrolyteinterface.tionalresourcesavailable.Still,computationalinsightisInprinciple,adifferentvarietyofMLpotentialscouldbeusedgenerallymostreliablyderivedwhenmodelingasmuchoftheheretoaccountforsolvatedinterfaces,butweareintriguedatactualreactionenvironmentaspossible.Reactiveforcefields,thepossibilitiesofkernel-basedsymmetricgradientdomain113,114suchasReaxFF,areaframeworktoprovideveryusefulmachinelearning(sGDML)methods127thatcanbetrainedinsights,evenifthesemethodsmightnotalwaysbeasaccurateagainstpotentialenergysurfacesofmedium-sizedmoleculesortransferableastheQCmethodstheyaretrainedon.withfarfewerdatapointsthanstandardBPNNpotentials.Asanexample,wepreviouslystudiedcatalyticreaction(sGDMLmethodsusuallyrequireonlyafewhundredtrainingmechanismsusingthecomputationalhydrogenelectrodemodelpoints,whereasBPNNscanrequiremanythousands.)TheneedtounderstandhowdopingTiO2mightdeoptimizethereductionforlessdataagainallowsonetofocuscomputationalcostsonkineticsandthereforeresultinanimprovedanticorrosion115obtainingmuchhigherqualitydata,forinstance,abinitiocoating.TheactualsysteminvolvedTimaterialsthatformedmoleculardynamicsenergiesandforcesusingcorrelatedwavenativeoxidesthatwerebelievedtohavelargelyamorphousfunctionmethods.AkeycurrentlimitationwithsGDMLforcestructures,butQCanalysescouldnotbecarriedoutuntilfields,however,istheirrelianceonafewmolecularatomic-scaleamorphousstructureswereestablished.UsingReaxFFparametersfromanotherstudy,116weusedcomputa-representationsaswellastheirinabilitytobetransferabletotionallyefficientmoleculardynamicstoannealcrystallineTiO2systemswithdifferentnumbersofatomsthanisusedintheirstructuresintolessorderedstructuresfollowinganalogoustrainingset,andthiscurrentlylimitstheirusetosimulationsonproceduresusedbyJohnsonandcoworkersforstudiesofjustsinglemoleculesorclusters.Currenteffortsbyourgroupinamorphoussilica.117TheamorphousTiOstructureswecollaborationwiththeTkatchenkogrouphavebeentoward2obtaineddidnotwellrepresentexperimentalstructuresforexploringwaystoovercometheselimitations.Onceovercome,smallnanoparticles,118butfurthergeometryoptimizationsusingweseeopportunitiesforautomatedandefficientdevelopments5DFTresultedinstructuresthatwereingoodagreementwithofMLmodelssuitableforstudiesofmixed-solventsystemsandexperimentaldata.Furthermore,ouranalysisofatomicdopantstheuseofthoseinstudiesofelectrochemicalinterfaces,forpredictedthataluminumandvanadiumwouldbeusefulforexample,inthecontextofrefs128−131.6502https://dx.doi.org/10.1021/acs.jpcc.0c11345J.Phys.Chem.C2021,125,6495−6507

8TheJournalofPhysicalChemistryCpubs.acs.org/JPCCFeatureArticle■CONCLUSIONSin2017.HeiscurrentlyanNSFGraduateResearchFellowandfourth-WehavegivenanoverviewandoutlookforhowourgroupusesyearPh.D.studentattheUniversityofPittsburgh.Hisresearchfocusescomputationalchemistrytoexaminethethermodynamicandonthedevelopmentandapplicationofalchemicalperturbationdensitykineticpropertiesofhypotheticalcatalystsincomplexenviron-functionaltheorytoheterogeneouscatalystsystems.ments.APDFT,especiallyusinghigherordercorrections,hasAlexM.Maldonadoreceivedhisbachelor’sinchemicalengineeringattremendouspromiseforacceleratingcomputationalscreeningWesternMichiganUniversityin2018andwasanMI-LSAMPscholar.effortswhilerelyingononlyrelativelysmallsetsofQCHeiscurrentlyaPittSTRIVEscholar,anR.K.MellonGraduateFellow,calculations.Forexample,APDFTcanbeusedonasinglesetandathird-yearPh.D.studentattheUniversityofPittsburgh.Hisofimagesalongareactionpathwaytogenerateinsightfuldataforresearchfocusesonthedevelopmentandapplicationofcomputationalmanyotherrelatedprocesses.Asitbecomesmoretested,itwillchemistrymethodstoelucidatesolvatedreactionmechanisms.becomemorelikelytotransformhowconventionalcomputa-LingyanZhaoreceivedherbachelor’sdegreesinmaterialsscienceattionalmodelingworkflowsareusedwhenexploringthechemicalBeijingUniversityofChemicalTechnologyin2018andmaster’sdegreespacefornewcatalyststhataresubjectedtoconditionsintheiratCarnegieMellonUniversityin2019.Sheiscurrentlyasecond-yearlocalenvironmentunderambientconditionsthatincludePh.D.studentattheUniversityofPittsburgh.HerresearchfocusesonappliedpotentialandpHeffects.Finally,whereaslessexpensivemodelingelectrocatalytichomogeneousandheterogeneousreactionmodelingschemescanbeeffective,weenvisionthatmechanisms.comprehensiveinvestigationsofexplicitlysolvatedreactionmechanismswillbecomepossiblewithcontinueddevelopmentsBarbaroZuluetaobtainedhisbachelor’sdegreesinphysicsandofMLforcefields.Allcombined,onecanenvisionelaboratechemistryfromTempleUniversityin2019.Heisafirst-yearPh.D.workflowsthatwouldbesuitableformicrokineticpredictionsstudentattheUniversityofPittsburgh.Hisresearchfocusesonthebasedonmassivesearchspacesforhypotheticalcandidatesdevelopmentandapplicationsofalchemicalperturbationdensityacrossthechemicalandmaterialsspace.functionaltheory.BrianGentrycompletedhisbachelor’sdegreeinmechanicalengineer-■AUTHORINFORMATIONingattheUniversityofPittsburghin2020,wherehisresearchfocusedCorrespondingAuthoronmodelingmolecularchelantsandgelatormolecules.HeisnowaJohnA.Keith−DepartmentofChemicalandPetroleumgraduatestudentandanNSFGraduateResearchFellowatCarnegieEngineering,UniversityofPittsburgh,Pittsburgh,PennsylvaniaMellonUniversitypursuingajointPh.D.inmechanicalengineeringand15261,UnitedStates;orcid.org/0000-0002-6583-6322;engineeringandpublicpolicy.Email:jakeith@pitt.eduEliLipsmanwasanundergraduateresearcherwhoworkedonimplementingdescriptorsformachinelearningforcefieldswhileAuthorsfinishinghisbachelor’sdegreeinchemicalengineering.CharlesD.Griego−DepartmentofChemicalandPetroleumEngineering,UniversityofPittsburgh,Pittsburgh,PennsylvaniaTaeHoonChoiearnedhisPh.D.inchemistryfromtheUniversityof15261,UnitedStatesPittsburgh,andhedidpostdoctoralresearchattheUniversityofAlexM.Maldonado−DepartmentofChemicalandPetroleumChicago.After6yearsofbeingaprofessoratChungnamNationalEngineering,UniversityofPittsburgh,Pittsburgh,PennsylvaniaUniversityinKorea,hereturnedtotheUniversityofPittsburghin201715261,UnitedStatestobearesearchassistantprofessor.HisresearchfocusesonLingyanZhao−DepartmentofChemicalandPetroleumimplementingmultiscalesimulationsformetal−ligandbindingandEngineering,UniversityofPittsburgh,Pittsburgh,solvatedreactions.Pennsylvania15261,UnitedStatesJohnA.Keithearnedabachelor’sdegreefromWesleyanUniversityandBarbaroZulueta−DepartmentofChemicalandPetroleumaPh.D.fromCaltech.AfteranAlexandervonHumboldtpostdoctoralEngineering,UniversityofPittsburgh,Pittsburgh,fellowshipattheUniversityofUlm(Germany),hewasanAssociatePennsylvania15261,UnitedStatesResearchScholaratPrincetonUniversity.HebeganhisindependentBrianM.Gentry−DepartmentofChemicalandPetroleumcareerin2013attheUniversityofPittsburghintheDepartmentofEngineering,UniversityofPittsburgh,Pittsburgh,ChemicalandPetroleumEngineeringasaninauguralR.K.MellonPennsylvania15261,UnitedStatesFacultyFellowinEnergy.HereceivedanNSF-CAREERawardin2017EliLipsman−DepartmentofChemicalandPetroleumandwaspromotedtoassociateprofessorwithtenurein2019.HisEngineering,UniversityofPittsburgh,Pittsburgh,researchdevelopsandappliescomputationalchemistrymethodsforPennsylvania15261,UnitedStatesinsightsintotheimproveddesignofmoleculesandmaterials.TaeHoonChoi−DepartmentofChemicalandPetroleumEngineering,UniversityofPittsburgh,Pittsburgh,■Pennsylvania15261,UnitedStatesACKNOWLEDGMENTSWeacknowledgesupportfromtheR.K.MellonFoundationandCompletecontactinformationisavailableat:theU.S.NationalScienceFoundation(grantnos.CBET-https://pubs.acs.org/10.1021/acs.jpcc.0c113451653392,CBET-1705592,andCHE-1856460).C.D.G.ac-knowledgessupportfromtheNationalScienceFoundationAuthorContributions†GraduateResearchFellowshipundergrantno.1747452.C.D.G.andA.M.M.contributedequallytothisworkComputationalresourcesandtechnicalsupportwereprovidedNotesbytheUniversityofPittsburghCenterforResearchComputing.Theauthorsdeclarenocompetingfinancialinterest.Biographies■REFERENCESCharlesD.Griegoreceivedhisbachelor’sdegreeinchemical(1)Schlögl,R.HeterogeneousCatalysis.Angew.Chem.,Int.Ed.2015,engineeringattheNewMexicoInstituteofMiningandTechnology54,3465−3520.6503https://dx.doi.org/10.1021/acs.jpcc.0c11345J.Phys.Chem.C2021,125,6495−6507

9TheJournalofPhysicalChemistryCpubs.acs.org/JPCCFeatureArticle(2)Montoya,J.H.;Seitz,L.C.;Chakthranont,P.;Vojvodic,A.;(21)Knizia,G.;Chan,G.K.-L.Densitymatrixembedding:AsimpleJaramillo,T.F.;Nørskov,J.K.Materialsforsolarfuelsandchemicals.alternativetodynamicalmean-fieldtheory.Phys.Rev.Lett.2012,109,Nat.Mater.2017,16,70−81.186404.(3)Seh,Z.W.;Kibsgaard,J.;Dickens,C.F.;Chorkendorff,I.;(22)Knizia,G.;Chan,G.K.-L.Densitymatrixembedding:Astrong-Nørskov,J.K.;Jaramillo,T.F.Combiningtheoryandexperimentincouplingquantumembeddingtheory.J.Chem.TheoryComput.2013,9,electrocatalysis:Insightsintomaterialsdesign.Science2017,355,1428−1432.No.eaad4998.(23)Marenich,A.V.;Cramer,C.J.;Truhlar,D.G.Universalsolvation(4)Bullock,R.M.;Chen,J.G.;Gagliardi,L.;Chirik,P.J.;Farha,O.K.;modelbasedonsoluteelectrondensityandonacontinuummodelofHendon,C.H.;Jones,C.W.;Keith,J.A.;Klosin,J.;Minteer,S.D.;etal.thesolventdefinedbythebulkdielectricconstantandatomicsurfaceUsingnature’sblueprinttoexpandcatalysiswithEarth-abundanttensions.J.Phys.Chem.B2009,113,6378−6396.metals.Science2020,369,No.eabc3183.(24)Barone,V.;Cossi,M.Quantumcalculationofmolecularenergies(5)Maldonado,A.M.;Basdogan,Y.;Berryman,J.T.;Rempe,S.B.;andenergygradientsinsolutionbyaconductorsolventmodel.J.Phys.Keith,J.A.First-principlesmodelingofchemistryinmixedsolvents:Chem.A1998,102,1995−2001.Wheretogofromhere?J.Chem.Phys.2020,152,130902.(25)Klamt,A.;Schüürmann,G.COSMO:anewapproachto(6)Feibelman,P.J.;Hammer,B.;Nørskov,J.K.;Wagner,F.;dielectricscreeninginsolventswithexplicitexpressionsfortheScheffler,M.;Stumpf,R.;Watwe,R.;Dumesic,J.TheCO/Pt(111)screeningenergyanditsgradient.J.Chem.Soc.,PerkinTrans.21993,Puzzle.J.Phys.Chem.B2001,105,4018−4025.2,799−805.(7)Stroppa,A.;Termentzidis,K.;Paier,J.;Kresse,G.;Hafner,J.CO(26)Sundararaman,R.;Goddard,W.A.,IIIThecharge-asymmetricadsorptiononmetalsurfaces:Ahybridfunctionalstudywithplane-nonlocallydeterminedlocal-electric(CANDLE)solvationmodel.J.wavebasisset.Phys.Rev.B:Condens.MatterMater.Phys.2007,76,Chem.Phys.2015,142,064107.195440.(27)Prentice,J.C.A.;Aarons,J.;Womack,J.C.;Allen,A.E.A.;(8)Tameh,M.S.;Huang,C.LargeImpactofApproximateExchange-Andrinopoulos,L.;Anton,L.;Bell,R.A.;Bhandari,A.;Bramley,G.A.;CorrelationFunctionalsonModelingtheWaterGasShiftReactiononCharlton,R.J.;etal.TheONETEPlinear-scalingdensityfunctionalCopper.J.Phys.Chem.C2020,124,22506−22520.theoryprogram.J.Chem.Phys.2020,152,174111.(9)Liechtenstein,A.I.;Anisimov,V.I.;Zaanen,J.Density-functional(28)Hutter,J.;Iannuzzi,M.;Schiffmann,F.;VandeVondele,J.cp2k:theoryandstronginteractions:OrbitalorderinginMott-Hubbardatomisticsimulationsofcondensedmattersystems.WileyInterdiscip.insulators.Phys.Rev.B:Condens.MatterMater.Phys.1995,52,R5467−Rev.:Comput.Mol.Sci.2014,4,15−25.R5470.(29)Blum,V.;Gehrke,R.;Hanke,F.;Havu,P.;Havu,V.;Ren,X.;(10)Anisimov,V.I.;Aryasetiawan,F.;Lichtenstein,A.I.First-Reuter,K.;Scheffler,M.Abinitiomolecularsimulationswithnumericprinciplescalculationsoftheelectronicstructureandspectraofstronglyatom-centeredorbitals.Comput.Phys.Commun.2009,180,2175−correlatedsystems:theLDA+Umethod.J.Phys.:Condens.Matter2196.1997,9,767−808.(30)Neese,F.TheORCAprogramsystem.WileyInterdiscip.Rev.:(11)Dudarev,S.L.;Liechtenstein,A.I.;Castell,M.R.;Briggs,G.A.Comput.Mol.Sci.2012,2,73−78.D.;Sutton,A.P.SurfacestatesonNiO(100)andtheoriginofthe(31)Kresse,G.;Furthmüller,J.Efficientiterativeschemesforabinitiocontrastreversalinatomicallyresolvedscanningtunnelingmicroscopetotal-energycalculationsusingaplane-wavebasisset.Phys.Rev.B:images.Phys.Rev.B:Condens.MatterMater.Phys.1997,56,4900−Condens.MatterMater.Phys.1996,54,11169−11186.4908.(32)vonLilienfeld,O.A.;Müller,K.-R.;Tkatchenko,A.Exploring(12)Dudarev,S.L.;Botton,G.A.;Savrasov,S.Y.;Humphreys,C.J.;chemicalcompoundspacewithquantum-basedmachinelearning.Nat.Sutton,A.P.Electron-energy-lossspectraandthestructuralstabilityofRev.Chem.2020,4,347−358.nickeloxide:AnLSDA+Ustudy.Phys.Rev.B:Condens.MatterMater.(33)Jain,A.;Ong,S.P.;Hautier,G.;Chen,W.;Richards,W.D.;Phys.1998,57,1505−1509.Dacek,S.;Cholia,S.;Gunter,D.;Skinner,D.;Ceder,G.;etal.(13)Perdew,J.P.;Schmidt,K.Jacob’sladderofdensityfunctionalCommentary:TheMaterialsProject:Amaterialsgenomeapproachtoapproximationsfortheexchange-correlationenergy.AIPConf.Proc.acceleratingmaterialsinnovation.APLMater.2013,1,011002.2000,577,1−20.(34)dePablo,J.J.;Jackson,N.E.;Webb,M.A.;Chen,L.-Q.;Moore,(14)Libisch,F.;Huang,C.;Carter,E.A.EmbeddedcorrelatedJ.E.;Morgan,D.;Jacobs,R.;Pollock,T.;Schlom,D.G.;Toberer,E.S.;wavefunctionschemes:Theoryandapplications.Acc.Chem.Res.2014,etal.Newfrontiersforthematerialsgenomeinitiative.npjComput.47,2768−2775.Mater.2019,5,41.(15)Huang,C.;Pavone,M.;Carter,E.A.Quantummechanical(35)Nørskov,J.K.;Bligaard,T.TheCatalystGenome.Angew.Chem.,embeddingtheorybasedonauniqueembeddingpotential.J.Chem.Int.Ed.2013,52,776−777.Phys.2011,134,154110.(36)O’Mara,J.;Meredig,B.;Michel,K.Materialsdatainfrastructure:(16)Manby,F.R.;Stella,M.;Goodpaster,J.D.;Miller,T.F.Asimple,Acasestudyofthecitrinationplatformtoexaminedataimport,storage,exactdensity-functional-theoryembeddingscheme.J.Chem.Theoryandaccess.JOM2016,68,2031−2034.Comput.2012,8,2564−2568.(37)Winther,K.T.;Hoffmann,M.J.;Boes,J.R.;Mamun,O.;Bajdich,(17)Goodpaster,J.D.;Ananth,N.;Manby,F.R.;Miller,T.F.ExactM.;Bligaard,T.Catalysis-Hub.org,anopenelectronicstructurenonadditivekineticpotentialsforembeddeddensityfunctionaltheory.databaseforsurfacereactions.Sci.Data2019,6,75.J.Chem.Phys.2010,133,084103.(38)vonLilienfeld,O.A.;Lins,R.D.;Rothlisberger,U.Variational(18)Goodpaster,J.D.;Barnes,T.A.;Miller,T.F.Embeddeddensityparticlenumberapproachforrationalcompounddesign.Phys.Rev.Lett.functionaltheoryforcovalentlybondedandstronglyinteracting2005,95,153002.subsystems.J.Chem.Phys.2011,134,164108.(39)vonLilienfeld,O.A.;Tuckerman,M.E.Moleculargrand-(19)Goodpaster,J.D.;Barnes,T.A.;Manby,F.R.;Miller,T.F.canonicalensembledensityfunctionaltheoryandexplorationofDensityfunctionaltheoryembeddingforcorrelatedwavefunctions:chemicalspace.J.Chem.Phys.2006,125,154104.Improvedmethodsforopen-shellsystemsandtransitionmetal(40)toBaben,M.;Achenbach,J.O.;vonLilienfeld,O.A.Guidingabcomplexes.J.Chem.Phys.2012,137,224113.initiocalculationsbyalchemicalderivatives.J.Chem.Phys.2016,144,(20)Goodpaster,J.D.;Barnes,T.A.;Manby,F.R.;Miller,T.F.104103.Accurateandsystematicallyimprovabledensityfunctionaltheory(41)Griego,C.D.;Kitchin,J.R.;Keith,J.A.Accelerationofcatalystembeddingforcorrelatedwavefunctions.J.Chem.Phys.2014,140,discoverywitheasy,fast,andreproduciblecomputationalalchemy.Int.18A507.J.QuantumChem.2021,121,No.e26380.6504https://dx.doi.org/10.1021/acs.jpcc.0c11345J.Phys.Chem.C2021,125,6495−6507

10TheJournalofPhysicalChemistryCpubs.acs.org/JPCCFeatureArticle(42)Alfe,D.;Gillan,M.J.;Price,G.D.Constraintsonthè(64)Reuter,K.;Scheffler,M.First-principlesatomisticthermody-compositionoftheEarth’scorefromabinitiocalculations.Naturenamicsforoxidationcatalysis:surfacephasediagramsandcatalytically2000,405,172−175.interestingregions.Phys.Rev.Lett.2003,90,046103.(43)Beste,A.;Harrison,R.J.;Yanai,T.Directcomputationofgeneral(65)Exner,K.S.;Anton,J.;Jacob,T.;Over,H.Chlorineevolutionchemicalenergydifferences:Applicationtoionizationpotentials,reactiononRuO2(110):Abinitioatomisticthermodynamicsstudy-excitation,andbondenergies.J.Chem.Phys.2006,125,074101.pourbaixdiagrams.Electrochim.Acta2014,120,460−466.(44)Solovyeva,A.;vonLilienfeld,O.A.Alchemicalscreeningofionic(66)Marjolin,A.;Keith,J.A.Thermodynamicdescriptorsforcrystals.Phys.Chem.Chem.Phys.2016,18,31078−31091.moleculesthatcatalyzeefficientCO2electroreductions.ACSCatal.(45)Chang,K.Y.S.;Fias,S.;Ramakrishnan,R.;vonLilienfeld,O.A.2015,5,1123−1130.Fastandaccuratepredictionsofcovalentbondsinchemicalspace.J.(67)Saravanan,K.;Basdogan,Y.;Dean,J.;Keith,J.A.ComputationalChem.Phys.2016,144,174110.investigationofCO2electroreductionontinoxideandpredictionsof(46)Chang,K.Y.S.;vonLilienfeld,O.A.AlxGa1−xAscrystalswithTi,V,NbandZrdopantsforimprovedcatalysis.J.Mater.Chem.Adirect2eVbandgapsfromcomputationalalchemy.Phys.Rev.Mater.2017,5,11756−11763.2018,2,073802.(68)Jacob,T.Theoreticalinvestigationsonthepotential-induced(47)Sheppard,D.;Henkelman,G.;vonLilienfeld,O.A.AlchemicalformationofPt-oxidesurfaces.J.Electroanal.Chem.2007,607,158−derivativesofreactionenergetics.J.Chem.Phys.2010,133,084104.166.(48)Weigend,F.;Schrodt,C.;Ahlrichs,R.Atomdistributionsin(69)Saravanan,K.;Gottlieb,E.;Keith,J.A.Nitrogen-dopedbinaryatomclusters:Aperturbationalapproachanditsvalidationinananocarbonmaterialsunderelectroreductionoperatingconditionscasestudy.J.Chem.Phys.2004,121,10380−10384.andimplicationsforelectrocatalysisofCO2.Carbon2017,111,859−(49)Weigend,F.ExtendingDFT-basedgeneticalgorithmsbyatom-866.to-placere-assignmentviaperturbationtheory:Asystematicand(70)Yang,J.Y.;Kerr,T.A.;Wang,X.S.;Barlow,J.M.ReducingCO2unbiasedapproachtostructuresofmixed-metallicclusters.J.Chem.toHCO2−atMildPotentials:LessonsfromFormateDehydrogenase.Phys.2014,141,134103.J.Am.Chem.Soc.2020,142,19438−19445.(50)Al-Hamdani,Y.S.;Michaelides,A.;vonLilienfeld,O.A.(71)Chatterjee,S.;Griego,C.;Hart,J.L.;Li,Y.;Taheri,M.L.;Keith,J.;Snyder,J.D.FreestandingnanoporouspalladiumalloysasCOExploringdissociativewateradsorptiononisoelectronicallyBNdopedpoisoningtolerantelectrocatalystsfortheelectrochemicalreductionofgrapheneusingalchemicalderivatives.J.Chem.Phys.2017,147,164113.CO2toformate.ACSCatal.2019,9,5290−5301.(72)Goldsmith,C.F.;West,R.H.AutomaticGenerationof(51)Balawender,R.;Welearegay,M.A.;Lesiuk,M.;DeProft,F.;MicrokineticMechanismsforHeterogeneousCatalysis.J.Phys.Chem.Geerlings,P.ExploringChemicalSpacewiththeAlchemicalC2017,121,9970−9981.Derivatives.J.Chem.TheoryComput.2013,9,5327−5340.(73)Blondal,K.;Jelic,J.;Mazeau,E.;Studt,F.;West,R.H.;(52)Balawender,R.;Lesiuk,M.;DeProft,F.;Geerlings,P.ExploringGoldsmith,C.F.Computer-generatedkineticsforcoupledheteroge-ChemicalSpacewithAlchemicalDerivatives:BN-Simultaneousneous/homogeneoussystems:AcasestudyincatalyticcombustionofSubstitutionPatternsinC60.J.Chem.TheoryComput.2018,14,methaneonplatinum.Ind.Eng.Chem.Res.2019,58,17682−17691.1154−1168.(74)Zhang,Z.;Zandkarimi,B.;Alexandrova,A.N.Ensemblesof(53)Fias,S.;Heidar-Zadeh,F.;Geerlings,P.;Ayers,P.W.Chemicalmetastablestatesgovernheterogeneouscatalysisondynamicinterfaces.transferabilityoffunctionalgroupsfollowsfromthenearsightednessofAcc.Chem.Res.2020,53,447−458.electronicmatter.Proc.Natl.Acad.Sci.U.S.A.2017,114,11633−(75)Cheula,R.;Maestri,M.;Mpourmpakis,G.Modeling11638.morphologyandcatalyticactivityofnanoparticleensemblesunder(54)Saravanan,K.;Kitchin,J.R.;vonLilienfeld,O.A.;Keith,J.A.reactionconditions.ACSCatal.2020,10,6149−6158.Alchemicalpredictionsforcomputationalcatalysis:Potentialand(76)Calle-Vallejo,F.;Loffreda,D.;Koper,M.T.M.;Sautet,P.limitations.J.Phys.Chem.Lett.2017,8,5002−5007.Introducingstructuralsensitivityintoadsorption−energyscaling(55)Griego,C.D.;Saravanan,K.;Keith,J.A.Benchmarkingrelationsbymeansofcoordinationnumbers.Nat.Chem.2015,7,computationalalchemyforcarbide,nitride,andoxidecatalysts.Adv.403−410.TheorySimul.2019,2,1800142.(77)Yan,Z.;Taylor,M.G.;Mascareno,A.;Mpourmpakis,G.Size-,(56)Griego,C.D.;Zhao,L.;Saravanan,K.;Keith,J.A.Machineshape-,andcomposition-dependentmodelformetalnanoparticlelearningcorrectedalchemicalperturbationdensityfunctionaltheoryforstabilityprediction.NanoLett.2018,18,2696−2704.catalysisapplications.AIChEJ.2020,66,No.e17041.(78)Roling,L.T.;Choksi,T.S.;Abild-Pedersen,F.Acoordination-(57)Henkelman,G.;Uberuaga,B.P.;Jónsson,H.Aclimbingimagebasedmodelfortransitionmetalalloynanoparticles.Nanoscale2019,nudgedelasticbandmethodforfindingsaddlepointsandminimum11,4438−4452.energypaths.J.Chem.Phys.2000,113,9901−9904.(79)Basdogan,Y.;Maldonado,A.M.;Keith,J.A.Advancesand(58)Ramakrishnan,R.;Dral,P.O.;Rupp,M.;VonLilienfeld,O.A.challengesinmodelingsolvatedreactionmechanismsforrenewableBigdatameetsquantumchemistryapproximations:TheΔ-machinefuelsandchemicals.WileyInterdiscip.Rev.:Comput.Mol.Sci.2020,10,learningapproach.J.Chem.TheoryComput.2015,11,2087−2096.No.e1446.(59)vonRudorff,G.F.;vonLilienfeld,O.A.Rapidandaccurate(80)Tomasi,J.;Mennucci,B.;Cammi,R.Quantummechanicalmoleculardeprotonationenergiesfromquantumalchemy.Phys.Chem.continuumsolvationmodels.Chem.Rev.2005,105,2999−3094.Chem.Phys.2020,22,10519−10525.(81)Tomasi,J.;Persico,M.Molecularinteractionsinsolution:an(60)Wang,S.;Pillai,H.S.;Xin,H.Bayesianlearningofchemisorptionoverviewofmethodsbasedoncontinuousdistributionsofthesolvent.forbridgingcomplexitiesofelectronicdescriptors.Nat.Commun.2020,Chem.Rev.1994,94,2027−2094.11,6132.(82)Klamt,A.COSMO-RS:FromQuantumChemistrytoFluidPhase(61)Boes,J.R.;Groenenboom,M.C.;Keith,J.A.;Kitchin,J.R.ThermodynamicsandDrugDesign;Elsevier,2005.NeuralnetworkandReaxFFcomparisonforAuproperties.Int.J.(83)Hirata,F.MolecularTheoryofSolvation;Springer:Dordrecht,QuantumChem.2016,116,979−987.TheNetherlands,2003;Vol.24.(62)Boes,J.R.;Kitchin,J.R.Neuralnetworkpredictionsofoxygen(84)Pliego,J.R.,Jr;Riveros,J.M.Hybriddiscrete-continuuminteractionsonadynamicPdsurface.Mol.Simul.2017,43,346−354.solvationmethods.WileyInterdiscip.Rev.:Comput.Mol.Sci.2020,10,(63)Nørskov,J.K.;Rossmeisl,J.;Logadottir,A.;Lindqvist,L.;No.e1440.Kitchin,J.R.;Bligaard,T.;Jónsson,H.Originoftheoverpotentialfor(85)Chaudhari,M.I.;Vanegas,J.M.;Pratt,L.;Muralidharan,A.;oxygenreductionatafuel-cellcathode.J.Phys.Chem.B2004,108,Rempe,S.B.Hydrationmimicrybymembraneionchannels.Annu.Rev.17886−17892.Phys.Chem.2020,71,461−484.6505https://dx.doi.org/10.1021/acs.jpcc.0c11345J.Phys.Chem.C2021,125,6495−6507

11TheJournalofPhysicalChemistryCpubs.acs.org/JPCCFeatureArticle(86)Asthagiri,D.;Dixit,P.;Merchant,S.;Paulaitis,M.;Pratt,L.;(109)Sheppard,D.;Xiao,P.;Chemelewski,W.;Johnson,D.D.;Rempe,S.B.;Varma,S.IonselectivityfromlocalconfigurationsofHenkelman,G.Ageneralizedsolid-statenudgedelasticbandmethod.J.ligandsinsolutionsandionchannels.Chem.Phys.Lett.2010,485,1−7.Chem.Phys.2012,136,074103.(87)Pratt,L.R.;Asthagiri,D.InFreeEnergyCalculations:Theoryand(110)Plata,R.E.;Singleton,D.A.AcasestudyofthemechanismofApplicationsinChemistryandBiology;Chipot,C.,Pohorille,A.,Eds.;alcohol-mediatedMoritaBaylis−Hillmanreactions.TheimportanceofSpringer,2007;pp323−351.experimentalobservations.J.Am.Chem.Soc.2015,137,3811−3826.(88)Beck,T.L.;Paulaitis,M.E.;Pratt,L.R.ThePotentialDistribution(111)Liu,Z.;Patel,C.;Harvey,J.N.;Sunoj,R.B.MechanismandTheoremandModelsofMolecularSolutions;CambridgeUniversityreactivityintheMorita−Baylis−Hillmanreaction:thechallengeofPress,2006.accuratecomputations.Phys.Chem.Chem.Phys.2017,19,30647−(89)Pratt,L.R.;Rempe,S.B.Quasi-chemicaltheoryandimplicit30657.solventmodelsforsimulations.AIPConf.Proc.1999,492,172−201.(112)Basdogan,Y.;Keith,J.A.Aparamedictreatmentformodeling(90)Bryantsev,V.S.;Diallo,M.S.;Goddard,W.A.,IIIexplicitlysolvatedchemicalreactionmechanisms.Chem.Sci.2018,9,Computationalstudyofcopper(II)complexationandhydrolysisin5341−5346.aqueoussolutionsusingmixedcluster/continuummodels.J.Phys.(113)VanDuin,A.C.;Dasgupta,S.;Lorant,F.;Goddard,W.A.,IIIChem.A2009,113,9559−9567.ReaxFF:areactiveforcefieldforhydrocarbons.J.Phys.Chem.A2001,(91)Basdogan,Y.;Groenenboom,M.C.;Henderson,E.;De,S.;105,9396−9409.Rempe,S.B.;Keith,J.A.Machinelearning-guidedapproachfor(114)Senftle,T.P;Hong,S.;Islam,M.M.;Kylasa,S.B;Zheng,Y.;studyingsolvationenvironments.J.Chem.TheoryComput.2020,16,Shin,Y.K.;Junkermeier,C.;Engel-Herbert,R.;Janik,M.J;Aktulga,H.633−642.M.;etal.TheReaxFFreactiveforce-field:development,applications(92)Bartók,A.P.;Kondor,R.;Csanyi,G.Onrepresentingchemicaĺandfuturedirections.npjComput.Mater.2016,2,15011.environments.Phys.Rev.B:Condens.MatterMater.Phys.2013,87,(115)Groenenboom,M.;Anderson,R.;Horton,D.;Basdogan,Y.;Roeper,D.;Policastro,S.;Keith,J.DopedAmorphousTiOxidesto184115.DeoptimizeOxygenReductionReactionCatalysis.J.Phys.Chem.C(93)Ceriotti,M.;Tribello,G.A.;Parrinello,M.Simplifyingthe2017,121,16825−16830.representationofcomplexfree-energylandscapesusingsketch-map.(116)Kim,S.-Y.;Kumar,N.;Persson,P.;Sofo,J.;vanDuin,A.C.T.;Proc.Natl.Acad.Sci.U.S.A.2011,108,13023−13028.Kubicki,J.D.DevelopmentofaReaxFFReactiveForceFieldfor(94)Zhang,J.;Dolg,M.ABCluster:theartificialbeecolonyalgorithmTitaniumDioxide/WaterSystems.Langmuir2013,29,7838−7846.forclusterglobaloptimization.Phys.Chem.Chem.Phys.2015,17,(117)Ewing,C.S.;Bhavsar,S.;Veser,G.;McCarthy,J.J.;Johnson,J.24173−24181.K.AccurateAmorphousSilicaSurfaceModelsfromFirst-Principles(95)Simm,G.N.;Türtscher,P.L.;Reiher,M.SystematicThermodynamicsofSurfaceDehydroxylation.Langmuir2014,30,microsolvationapproachwithacluster-continuumschemeand5133−5141.conformationalsampling.J.Comput.Chem.2020,41,1144−1155.(118)Petkov,V.;Holzhüter,G.;Tröge,U.;Gerber,T.;Himmel,B.(96)Rogers,D.M.;Jiao,D.;Pratt,L.R.;Rempe,S.B.InAnnualAtomic-scalestructureofamorphousTiO2byelectron,X-rayReportsinComputationalChemistry;Wheeler,R.A.,Ed.;Elsevier,2012;diffractionandreverseMonteCarlosimulations.J.Non-Cryst.SolidsVol.8,pp71−127.1998,231,17−30.(97)Sunoj,R.B.;Anand,M.Microsolvatedtransitionstatemodelsfor(119)Schütt,K.T.;Chmiela,S.;vonLilienfeld,O.A.;Tkatchenko,A.;improvedinsightintochemicalpropertiesandreactionmechanisms.Tsuda,K.;Müller,K.-R.MachineLearningMeetsQuantumPhysics;Phys.Chem.Chem.Phys.2012,14,12715−12736.Springer,2020.(98)Roytman,V.A.;Singleton,D.A.Solvationdynamicsandthe(120)Behler,J.Perspective:Machinelearningpotentialsforatomisticnatureofreactionbarriersandion-pairintermediatesincarbocationsimulations.J.Chem.Phys.2016,145,170901.reactions.J.Am.Chem.Soc.2020,142,12865−12877.(121)vonLilienfeld,O.A.;Müller,K.-R.;Tkatchenko,A.Exploring(99)Hassanali,A.A.;Cuny,J.;Verdolino,V.;Parrinello,M.Aqueouschemicalcompoundspacewithquantum-basedmachinelearning.Nat.solutions:stateoftheartinabinitiomoleculardynamics.Philos.Trans.Rev.Chem.2020,4,347−358.R.Soc.,A2014,372,20120482.(122)Behler,J.;Parrinello,M.Generalizedneural-network(100)Iftimie,R.;Minary,P.;Tuckerman,M.E.Abinitiomolecularrepresentationofhigh-dimensionalpotential-energysurfaces.Phys.dynamics:Concepts,recentdevelopments,andfuturetrends.Proc.Rev.Lett.2007,98,146401.Natl.Acad.Sci.U.S.A.2005,102,6654−6659.(123)Keith,J.A.;Fantauzzi,D.;Jacob,T.;VanDuin,A.C.Reactive(101)Tse,J.S.Abinitiomoleculardynamicswithdensityfunctionalforcefieldforsimulatinggoldsurfacesandnanoparticles.Phys.Rev.B:theory.Annu.Rev.Phys.Chem.2002,53,249−290.Condens.MatterMater.Phys.2010,81,235404.(102)Lu,X.;Fang,D.;Ito,S.;Okamoto,Y.;Ovchinnikov,V.;Cui,Q.(124)Boes,J.R.;Groenenboom,M.C.;Keith,J.A.;Kitchin,J.R.QM/MMfreeenergysimulations:Recentprogressandchallenges.Mol.NeuralnetworkandReaxFFcomparisonforAuproperties.Int.J.Simul.2016,42,1056−1078.QuantumChem.2016,116,979−987.(103)Senn,H.M.;Thiel,W.QM/MMmethodsforbiomolecular(125)Khorshidi,A.;Peterson,A.A.Amp:Amodularapproachtosystems.Angew.Chem.,Int.Ed.2009,48,1198−1229.machinelearninginatomisticsimulations.Comput.Phys.Commun.(104)Lin,H.;Truhlar,D.G.QM/MM:whathavewelearned,where2016,207,310−324.arewe,andwheredowegofromhere?Theor.Chem.Acc.2007,117,(126)Groenenboom,M.C.;Anderson,R.M.;Wollmershauser,J.A.;185.Horton,D.J.;Policastro,S.A.;Keith,J.A.CombinedNeuralNetwork(105)Bolhuis,P.G.;Chandler,D.;Dellago,C.;Geissler,P.L.PotentialandDensityFunctionalTheoryStudyofTiAl2O5SurfaceTransitionPathSampling:ThrowingRopesOverRoughMountainMorphologyandOxygenReductionReactionOverpotentials.J.Phys.Passes,intheDark.Annu.Rev.Phys.Chem.2002,53,291−318.Chem.C2020,124,15171−15179.(106)Agarwal,V.;Dauenhauer,P.J.;Huber,G.W.;Auerbach,S.M.(127)Chmiela,S.;Sauceda,H.E.;Müller,K.-R.;Tkatchenko,A.Abinitiodynamicsofcellulosepyrolysis:NascentdecompositionTowardsexactmoleculardynamicssimulationswithmachine-learnedpathwaysat327and600C.J.Am.Chem.Soc.2012,134,14958−14972.forcefields.Nat.Commun.2018,9,3887.(107)Groenenboom,M.C.;Keith,J.A.Explicitlyunravelingtheroles(128)Sakong,S.;Mahlberg,D.;Roman,T.;Li,M.;Pandey,M.;Groß,ofcounterions,solventmolecules,andelectroncorrelationinsolutionA.InfluenceofLocalInhomogeneitiesandtheElectrochemicalphasereactionpathways.J.Phys.Chem.B2016,120,10797−10807.EnvironmentontheOxygenReductionReactiononPt-Based(108)Groenenboom,M.C.;Keith,J.A.QuantumchemicalanalysesElectrodes:ADFTStudy.J.Phys.Chem.C2020,124,27604−27613.ofBH−andBHOH−hydridetransferstoCOinaqueoussolution432(129)Rendón-Calle,A.;Builes,S.;Calle-Vallejo,F.Substantialwithpotentialsofmeanforce.ChemPhysChem2017,18,3148−3152.improvementofelectrocatalyticpredictionsbysystematicassessment6506https://dx.doi.org/10.1021/acs.jpcc.0c11345J.Phys.Chem.C2021,125,6495−6507

12TheJournalofPhysicalChemistryCpubs.acs.org/JPCCFeatureArticleofsolventeffectsonadsorptionenergies.Appl.Catal.,B2020,276,119147.(130)Kastlunger,G.;Lindgren,P.;Peterson,A.A.Controlled-PotentialSimulationofElementaryElectrochemicalReactions:ProtonDischargeonMetalSurfaces.J.Phys.Chem.C2018,122,12771−12781.(131)Heenen,H.H.;Gauthier,J.A.;Kristoffersen,H.H.;Ludwig,T.;Chan,K.Solvationatmetal/waterinterfaces:Anabinitiomoleculardynamicsbenchmarkofcommoncomputationalapproaches.J.Chem.Phys.2020,152,144703.6507https://dx.doi.org/10.1021/acs.jpcc.0c11345J.Phys.Chem.C2021,125,6495−6507