Accelerated Discovery of Zeolite Structures with Superior Mechanical Properties via Active Learning - Kim, Min - 2021 - Unknown

《Accelerated Discovery of Zeolite Structures with Superior Mechanical Properties via Active Learning - Kim, Min - 2021 - Unknown》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/JPCLLetterAcceleratedDiscoveryofZeoliteStructureswithSuperiorMechanicalPropertiesviaActiveLearningNamjungKim*andKyoungminMin*CiteThis:J.Phys.Chem.Lett.2021,12,2334−2339ReadOnlineACCESSMetrics&MoreArticleRecommendations*sıSupportingInformationABSTRACT:ABayesianactivelearningplatformisdevelopedfortheaccelerateddiscoveryofmechanicallysuperiorzeolitestructuresfrommorethanhalfamillionhypotheticalcandidates.Aninitialdatabasecontainingthemechanicalpropertiesofsynthesizablezeolitesistrainedtodevelopthemachinelearningregressionmodel.Then,aBayesianoptimizationschemeisimplementedtoidentifyzeoliteswithpotentiallyexcellentmechanicalproperties.Thenewlyaccumulateddatabaseconsistsof871labeledstructures,andtheuncertaintyofthepredictivemodelisreducedby40%and58%forthebulkandshearmoduli,respectively.Themodelconvergenceshowsthatnofurtherimprovementoccursafterthe10thiterationofoptimizations.Theproposedplatformisabletodiscover23newzeolitestructuresthathaveunprecedentedshearmoduli;inonecase,theshearmodulus(127.81GPa)is250%higherthanthepreviousdataset.Theproposedplatformacceleratesthematerialdiscoveryprocesswhilemaximizingcomputationalefficiencyandenhancingthepredictiveaccuracy.12iliceouszeolites,crystallinemetastablephasesofSiO2withapproacheswithagreateramountofdata.Furthermore,toShighsurface-to-volumeandmassratios,exhibitpromisinggeneratezeolitestructuresfortheadsorptionofmethane,thechemicalandphysicalpropertiesthatareadjustabledependinginversedesignapproachwasimplementedwiththeaidofontherequirementsoftheirapplications,suchastheadsorptiongenerativeadversarialnetworks,whichcangreatlyreducethe13ofpolarmaterials,separationbasedontheirporousnature,andnumberoftrials-and-errorsinthematerialdesignprocess.ionexchange.1,2Inaddition,theyarewidelyemployedinNaturallanguageprocessingisanothergreattoolwhichcanbeindustryasefficientcatalystsforbiomasstransformation,appliedtoanalyzezeolitesynthesisprocessesandparametersbyelectrodeandmembranematerialsoffuelcells,andadsorptionextractingsynthesisinformationandtrendsfromnumerous3−6publishedjournalarticles.10ThisinformationwasfurtherusedmaterialsforgasessuchasCO2.Zeoliteframeworkshavevariousintrinsicstructuralcharacter-todevelopamachinelearningregressionmodelforpredictingistics,includingporeconfiguration,dimensionality,andatomicthezeoliteframeworkdensityasanindicatorofstructuralDownloadedviaUNIVOFCAPETOWNonMay14,2021at15:07:25(UTC).bondconnectivity,leadingtoimmensestructuralvariety.stability.Inaddition,combinedwithacomputationalgeometryTherefore,thenumberofpossiblehypotheticalstructuresofapproach,machinelearningwasalsoimplementedtoaccurately14,15zeoliteframeworksismorethanhalfamillion,accordingtothepredictandclassifythetopologicaltypesofzeolitecrystals.Seehttps://pubs.acs.org/sharingguidelinesforoptionsonhowtolegitimatelysharepublishedarticles.PredictedCrystallographyOpenDatabase(PCOD).7,8How-Forpredictingandfindingchemicallysynthesizablezeoliteever,thenumberofexperimentallyvalidatedstructuresisstructures,anumberofresearchstudieshavebeenconductedapproximately200,asdeterminedbytheInternationalZeolitewithaidofdataminingtools,machinelearningalgorithms,andAssociation(IZA).9Thisisbecauseofthedifficultiesinautomateddatacrawlingplatformsforthepublishedliter-16−18ature.synthesizingzeoliteswithoptimizedstructural,chemical,andphysicalpropertiesfortargetapplications.10Furthermore,Thecurrentstudyisprimarilyfocusedonfindingzeoliteframeworkswithsuperiormechanicalpropertiesbyimplement-predictingthepropertiesofthesezeolitesbeforetheirsynthesisingBayesianactivelearning,whichincludesmachinelearningisclosetoimpossiblebecauseoftheirvaststructuralfeatures.andBayesianoptimization.SincezeolitesemployedingasInthisrespect,severalstudieshavebeenattempted,whichadsorptionorselectiveionexchangemembranesareexpectedtomainlyfocusedonsearchingforzeolitestructureswithidealpropertiesusingmachinelearninganddeeplearning.Forexample,toobtainacompletemapoftheelasticpropertiesofReceived:January30,2021hypotheticalzeolites,machinelearningbasedonaregressionAccepted:February23,2021modelwasimplementedbytrainingadatasetpreviouslyPublished:March2,2021constructedfromtheresultsofdensityfunctionaltheory(DFT)11calculationsandzeolite-specificfeatures.Thisworkwasfurtherextendedtodiscoverauxeticzeolitestructuresviasimilar©2021AmericanChemicalSocietyhttps://dx.doi.org/10.1021/acs.jpclett.1c003392334J.Phys.Chem.Lett.2021,12,2334−2339

1TheJournalofPhysicalChemistryLetterspubs.acs.org/JPCLLetterwithstandvariousloadingconditions,highmechanicalproper-SVMwassetto2.ThenumbersoftheminimumleafandBTE11,12,19−21tiesarecriticaldesignrequirements.Asmentioned,alearnerwere8and30,respectively.ThelearningrateofBTEwasnumberofrelevantstudiesforpredictingtheelasticpropertiesofsetto0.1.Marten5/2wasusedasthecovariancefunctionofzeolitestructuresareavailable,butadditionalcriticalstepsareGPR.Thehyper-parametervaluesofGBRwereobtainedfrom11stillrequired,including(1)validationofthepredictedpropertiesthestudybyEvansandCoudertinwhichthebestpredictionofthestructures(aroundhalfamillion)and(2)optimizationofaccuracyusingtheIZAdatasetwasreported.thedevelopedregressionmodelbytrainingitwithadditionalActivelearningisastrategyoflearningalgorithmsthatdatabasestominimizethepredictionuncertainty.iterativelyfindsthecandidatesforlabeling,thusmaximizingthe29Inthisstudy,Bayesianactivelearningwasimplementedtomodelaccuracywithminimumcomputationalcost.Amongsearchefficientlyforzeolitestructureswithsuperiormechanicalthevariousmethodsthathavebeenproposedforactivelearning,properties.First,toconstructtheinitialdatabase,themechanicalBayesianoptimizationwaschosentoreducetheuncertaintyandpropertiesoftheexperimentallyvalidatedzeoliteframeworksinfindthebestcasesofthepredictivemodelbecauseitiswelltheIZAdatabaseweredeterminedusingDFTcalculations.suitedforcaseswherethecostofobtainingnewlabelsis30Then,basedonthisdatabase,amachinelearningregressionexpensive.Inthisstudy,expectedimprovement(EI)wasused31modelwastrainedandtheelasticpropertiesoftheremainingastheacquisitionfunctionforBayesianoptimization.TheEIzeolitestructuresinthePCODdatabasewerepredicted.TheconsistsoftwocomponentsthatcontroltheexploitationandBayesianoptimizationmethodwasthenfollowedtochoosetheexploration,andthecoefficientforthistrade-offwassetto0.01.materialsforfurthervalidationtoachieveoptimalperformanceThepurposeofthecurrentworkistocombineamachineintermsoffindingstructureswithbettermechanicalpropertieslearningpredictivemodelwithDFTcalculationstofindtheandtoimprovethepredictionaccuracywithminimaltrials.mechanicallysuperiorzeolitestructuresinanexistingdatabaseTocomputethemechanicalpropertiesofthezeolite,DFTofhypotheticalzeolites(PCOD)inwhichthenumberofcalculationswiththeprojectoraugmentedwave(PAW)methodhypotheticalstructuresisapproximately590000.Sincebrute-intheViennaabinitiosimulationpackage(VASP)wereforcecalculationofallofthestructuresistooexpensiveand22,23implemented.Thegeneralizedgradientapproximationhighlyineffective,itiscriticalthatanoptimizationscheme(GGA)byPerdue,Burke,andErnzerhof(PBE)wasusedforshouldbeintroducedwhichcangreatlyreducethenumberof24theexchange-correlationfunctional.Aplane-wavecutoffoftrials-and-errors.Thespecificgoalsofthisstudyarethe650eVwasusedtoprovidebetterconvergence.Anenergyfollowing:(1)tosuggestzeolitestructureswithshear(G)andconvergencecriterionof10−6eVandak-pointdensityof1000bulk(B)modulithatarelargerthanthosereportedintheIZAperreciprocalatomwereused.Theelastictensormatrixwasdatabase(G∼55GPa,B∼110GPa);(2)toreducetheconstructedbasedonthreedisplacements(centraldifference)predictionuncertainty;and(3)toconfirmconvergenceofthewithastepsizeof0.015Å.TheelasticpropertieswereobtainedelasticpropertiesoftheunexploredzeolitestructuresinPCOD.fromthistensormatrix(shearmodulus,bulkmodulus,Poisson’sInthisrespect,toconstructtheinitialdatabase,121structures11,20ratio,andelasticanisotropy),andtherelevantequationsarewithpreviouslycalculatedelasticpropertieswerechosenshownintheSupportingInformation(SI).fromtheIZAdatabase,buttheirelasticpropertieswereToconstructthepredictivemodelforthebulkandshearrecalculatedwiththecurrentcalculationscheme.Thisisbecausemoduliofthezeolitesinthedatabase,weusedagradientapreviousstudyimplementedtheB3LYPhybridfunctionalwith25boosting-basedalgorithm,calledLightGBM.TheLightGBMlocalizedbasissets,whereasourworkemployedthePAW-basedalgorithmutilizesgradient-basedone-sidesamplingandDFTmethodwithadifferenttypeofpseudopotential.exclusivefeaturebundlingtoincreaseitsaccuracyandefficiency.Subsequently,theresultsbetweenthetwodifferentmethodsTherefore,itiswidelyusedforclassicalmachinelearningtaskswerecompared,whichrevealednodistinctdifference,asshown26,27withlarge-scaledatasets.ThisalgorithmwaschosenoverinFigureS1,thusvalidatingthereliabilityofthecurrenttheothermodels,suchasdecisiontree(DT),supportvectorapproach.machine(SVM),Gaussianprocessregressor(GPR),andTheBayesianactivelearningprocessispresentedinFigure1,gradientboostingregressor(GBR),becauseLightGBMisandthedetailsofeachstepareasfollows:effectiveandaccurate,especiallywithlarge-scaledataand(1)Basedonthecalculatedmechanicalpropertiesofthefeatures,whichcorrespondstoourproblem.zeolitestructuresintheIZAdatabase,amachinelearningmodelTomaximizethepredictionaccuracy,weoptimizedthewasconstructedtopredicttheshearandbulkmoduli.Previously11hyper-parametersofLightGBMusingRandomizedSearchCVindevelopedandvalidatedzeolitedescriptorswereadopted.28thescikit-learnlibrary.Therefore,thebesthyper-parameters(2)Theconstructedregressionmodelwasthenappliedtowerefoundwithrelativelylowcomputationalresources.predicttheelasticpropertiesoftheremainingzeolitestructuresTrainingandtestingused80%and20%ofthetotaldataset,inPCOD.Toquantifythepredictionuncertainty,50regressionrespectively.The5-foldcross-validationmethodwasalsomodelsweregeneratedbychangingthecompositionoftheperformedtosplitthetrainingsetandthevalidationset.trainingset.Stratifiedsamplingwasutilizedtosampleeachdataset(3)Theresultsfromtheuncertaintyanalysis(predictedmeanuniformlyovertheentiredomain.Thenumberofbinsfortheandstandarddeviationvalues)wereprovidedtotheBayesianstratifiedsamplingwassetto8,andtherangeofeachbinwasoptimizationmethodtoidentifythematerials.FiftystructuresdeterminedbasedontheminimumandmaximumvaluesofthewithlargervaluesofEIwerechosen,andtheirelasticpropertiesdataset.werecalculatedusingDFT.Zeolitestructures,inPCOD,withToselecttheidealregressionmodelforthecurrentstudy,themorethan66atomswerenotconsideredduringthepredictiveaccuraciesofsixdifferentmachinelearning(ML)-optimizationforcomputationalefficiency.DuringtheDFTbasedregressionmodels,suchasDT,boostingtreeensemblecalculations,structureswithnegativeshearandbulkmoduli,(BTE),SVM,GPR,GBR,andLightGBM,werecompared.Thestructureswithinvalidmoduli,andstructuresthatdidnotminimumleafsizeofDTwas4andthepolynomialdegreeofconvergeduringenergycalculationwereremovedfromthe2335https://dx.doi.org/10.1021/acs.jpclett.1c00339J.Phys.Chem.Lett.2021,12,2334−2339

2TheJournalofPhysicalChemistryLetterspubs.acs.org/JPCLLetterupdatinglist.Theupdateddatasetwastrainedagainwiththeregressionmodel.(4)Steps(2)and(3)wererepeated20timestoreducethepredictionuncertaintyandtodiscoverzeoliteswithsuperiormechanicalproperties.ThebulkmodulusdeterminedbyDFTcalculationswasplottedwiththosedeterminedbytheML-basedregressionmodels,asdisplayedinFiguresS2andS3.Theresultsindicatethat,amongthesixdifferentregressionmodels,LightGBMisthebestregressionmodelforpredictingthebulkandshearmoduliofzeolitesintermsofaccuracyandrepeatability;hence,itwasusedfortherestofthestudy.Afterconfirmingtheaccuracyofthepredictivemodelwiththelimitedamountofdata,activelearningwasperformedtoenhancethemodelaccuracythroughthegradualadditionoflabeleddatabyaniterativelearningprocess.Theconvergenceandaccuracyofthepredictivemodelweremonitoredastheiterationsproceeded.Figure2a−cshowtherepresentativepredictionsofthetestingdatasetforthebulkandshearmoduliatthe0th,10th,and20thiteration.First,forthebulkandshearmoduliatthezerothiteration,themeanabsoluteerrors(MAEs)were3.72GPa(bulk)and2.16GPa(shear)andtheR2scoreswere0.947(bulk)and0.885(shear).AstheiterationsFigure1.Bayesianactivelearningplatformtosearchformechanicallysuperiorzeolitestructures.proceeded,theMAEsslightlyincreasedto7.59GPa(bulk)and2.98GPa(shear)atthe10thiterationandplateauedaround6.75GPa(bulk)and2.69GPa(shear)atthe20thiteration.TheR2scoreatthe10thand20thiterationshowedasimilartrendasFigure2.(a−c)Representativepredictionsofthetestingdatasetofthebulkandshearmoduliatthe0th,10th,and20thiteration.(d)Standarddeviation(R2score)ofthepredictivemodelsasiterationproceeds.2336https://dx.doi.org/10.1021/acs.jpclett.1c00339J.Phys.Chem.Lett.2021,12,2334−2339

3TheJournalofPhysicalChemistryLetterspubs.acs.org/JPCLLetterthatoftheMAEvalues.Thisindicatedthatimprovementoftheofourpredictivemodel.Inaddition,thisresultindicatesthatthepredictivemodelwassaturatedathighaccuracy.Notably,additionofmoredatatocertainiterationsincreasesthealthoughmoredatapointswereadded,therewasnosignofpredictiveaccuracyandreducestheuncertaintyinthefurtherimprovementinpredictiveaccuracy.Althoughthisisprediction.Furthermore,smalllabeleddatasetsandtheabsenceratherunexpected,thiscanbejustifiedasfollows:(1)Theactualofdataadditionlimittheaccuracyimprovementofthenumberofstructures(∼900)withelasticpropertiesisextremelypredictivemodel.Atthefinaliteration,ourdatabasehas871smallcomparedwiththeunexploredregion(∼590000);hence,labeledzeolitestructures,whichisseventimeslargerthantheacertainlimitationintermsofaccuracymayexistfortheinitialIZAdataset.Becauseofthelargeamountofdataandthepredictivemodelbecauseoftheasymmetricalsizeofthelabeledefficientsamplingstrategy,webelievethatourpredictivemodelandunlabeleddatasets.(2)Itismoreimportanttovalidatetheproducesarefinedestimationforpredictingthebulkandshearcurrentframeworkforfindingsuperiormaterialsdespitethismoduliofzeolitescomparedwithpreviousstudies.uncertaintyandreducethepredictionuncertaintysignificantlyInordertoelucidatethedescriptorsusedinthisstudyandwherefurtheroptimizationisnotnecessary.theircorrelationchangesbeforeandafteraddingmoredatasets,Inthisregard,theconvergenceofthepredictivemodelwastheinterdependenciesoftheimplementeddescriptorswerefurthervalidatedbyobservingthestandarddeviation(R2value)obtained,asshowninFigureS5,fromtheIZAdatasetandfromforpredictionsofthebulkandshearmoduli,asshowninFigurethedatasetafterthe20thiteration.Theresultsindicatethatthe2d.TheR2scorerepresentsthequantitativedeviationofthedependenciesofthemostofpairwisedescriptorsareaffectedaspredictivevaluesamongthe50regressionmodelsgiventhethelabeleddataareadded.Hence,includingallthedescriptorstrainingandvalidationdatasets.Astheiterationsproceeded,thefortheregressionmodelisrecommendedforthecurrentmodelstandarddeviationofthepredictionsusingthevalidationsetinordertoobtainalltheunexpectedeffectsfromtheentirereducedandreachedcertainvalues(fromapproximately0.15todescriptors.Furthermore,theinterpretationofMLmodelsis0.05forbothcases).Thisvaluewasalmostcomparabletothatobtainedbyanalyzingthepartialdependenceplots,asshowninobtainedfromthetrainingset,confirmingthatthepredictiveFigureS6andTableS2.Itindicatesthattheinformationfromuncertaintywaslargelyreduced.Inotherwords,the50threecategories(local,structural,andporosity)ofzeoliteregressionmodelsproducedconsistentpredictions,anddescriptorsareequallyimportantandnecessarytopredicttheimprovementofthepredictivemodelwassaturatedatacertainbulkandshearmoduliofazeolitestructure.limit.Uponachievingmodelconvergence,thezeolitestructuresAfterconfirmingtheconvergenceofthepredictivemodelfoundduringactivelearningwereexaminedforsuperiorbasedonthelabeleddata,thepredictivemodelwiththemechanicalproperties.Figure4aand4brespectivelyshowthe∼590000unlabeleddatasetwasalsostudiedtomeasurethecalculatedbulkandshearmoduliofthequeriedpointsasconvergenceofthepredictivemodel.Thisisthemostcriticaliterationproceeds.EachpointrepresentsthebulkandshearstepinthevalidationprocessbecauseconvergenceshouldbemoduliofthezeolitestructurecalculatedbyDFT,whichwassatisfiedtofurtherutilizethecurrentpredictivemodel(tooqueriedbytheBayesianoptimizationalgorithm.ThebulkandmanyunlabeleddataremainedtobevalidatedwithDFTshearmoduliatIteration0wereobtainedfromthezeolitecalculations).Figure3displaysthestandarddeviation(R2score)structuresintheIZAdatabase.FigureS4showsthecontinuousdensitiesofthebulkandshearmoduliestimatedbykerneldensityestimation(KDE).Forthebulkmodulus,allthequeriedpointsexistedwithintherangeoftheIZAdataset(7.22−129.14GPa),indicatingthattheIZAdatabasecoveredmostofthefeasiblerangeofthebulkmodulus.However,fortheshearmodulus,ourBayesianactivelearningplatformdiscovered23newzeolitestructureswithshearmodulihigherthanthoseintheIZAdatabase(5.61−48.16GPa),aslistedinTableS1.Themaximumdiscoveredshearmoduluswas127.81GPa,whichismorethan2.5timeshigherthanthemaximumvalueintheIZAdataset.Theshearmodulusisameasureoftheresistanceagainstshearloading;azeolitewitha2.5timeshighershearmodulusisabletoresista250%highershearloading,whichcansatisfytherequirementofoutstandingmechanicalperformancewhile32maintainingchemicaladvantages.Figure3.Standarddeviation(R2score)ofthepredictivemodelwithInthiswork,aBayesianactivelearningplatformwastheunlabeleddatasetforthebulkandshearmoduli.developedfortheaccelerateddiscoveryofmechanicallysuperiorzeolitestructures.Aninitialdatabasecontainingthemechanicalofthepredictivemodelastheiterationsproceededwiththepropertiesofexperimentallysynthesizablezeolitestructureswasadditionoftheunlabeleddataset.Thestandarddeviationoftheconstructedtotrainthemachinelearningregressionmodel.predictivemodelalsoconvergedtoacertainvalue(approx-Then,Bayesianoptimizationwasappliedtothehypotheticalimately0.15forthebulkmodulusand0.04fortheshearzeolitestructures(morethanhalfamillioninthePCOD)tomodulus)afterthe15thiteration,similartothetrainingandidentifymaterialswithpotentiallyexcellentmechanicalproper-validationdatasets,indicatingconvergenceofthepredictiveties,whichsignificantlyreducesthecomputationalcost.Amodelwithrespecttotheunlabeleddataset.rigorouscomparisonoftheresultsindicatesthattheselectedOnthebasisofthisresult,weconfirmedthattheconvergenceML-basedregressionmodel,LightGBM,isthebestmodelforofourpredictivemodelandthe20iterationsofactivelearningtheactivelearningplatformintermsofaccuracyandweresufficienttoinducedistinctiveandfinalizedimprovementrepeatability.Thenewdatabasecontainsatotalof871labeled2337https://dx.doi.org/10.1021/acs.jpclett.1c00339J.Phys.Chem.Lett.2021,12,2334−2339

4TheJournalofPhysicalChemistryLetters■pubs.acs.org/JPCLLetterAUTHORINFORMATIONCorrespondingAuthorsNamjungKim−DepartmentofMechanicalEngineering,GachonUniversity,Seongnam,Gyeonggi-do13120,RepublicofKorea;orcid.org/0000-0002-2600-5921;Email:namjungk@gachon.ac.krKyoungminMin−SchoolofMechanicalEngineering,SoongsilUniversity,Seoul06978,RepublicofKorea;orcid.org/0000-0002-1041-6005;Email:kmin.min@ssu.ac.krCompletecontactinformationisavailableat:https://pubs.acs.org/10.1021/acs.jpclett.1c00339NotesDataAvailability:Machinelearningscripts,thetrainingset,andthetestsetusedinthisstudycanbefoundathttps://github.com/kminmin/zeolitebayesian.Theauthorsdeclarenocompetingfinancialinterest.■ACKNOWLEDGMENTSThisworkwassupportedbyaNationalResearchFoundationofKorea(NRF)grantfundedbytheKoreangovernment(MSIT)(2020R1F1A1073300,2020R1F1A1066519).ThisworkwassupportedbytheNationalSupercomputingCenterwithsupercomputingresourcesandincludingtechnicalsupport(KSC-2020-CRE-0012).■REFERENCES(1)Davis,M.E.OrderedPorousMaterialsforEmergingApplications.Nature2002,417(6891),813−821.(2)Martínez,C.;Corma,A.InorganicMolecularSieves:Preparation,Figure4.DFT-calculated(a)bulkand(b)shearmoduliofthequeriedModificationandIndustrialApplicationinCatalyticProcesses.Coord.zeolitestructuresasBayesianactivelearningproceeds.ThedottedlineChem.Rev.2011,255(13),1558−1580.representstheminimumandthemaximummodulusthattheIZA(3)Moliner,M.StateoftheArtofLewisAcid-ContainingZeolites:databasecovers.LessonsfromFineChemistrytoNewBiomassTransformationProcesses.Dalt.Trans.2014,43(11),4197−4208.(4)Chen,Z.;Holmberg,B.;Li,W.;Wang,X.;Deng,W.;Munoz,R.;zeolitestructures,whichisseventimeslargerthantheIZAdataYan,Y.Nafion/ZeoliteNanocompositeMembranebyinSituset.CrystallizationforaDirectMethanolFuelCell.Chem.Mater.2006,18(24),5669−5675.Becauseofthestrategicadditionoflabeleddataandthewell-(5)Wu,H.;Zheng,B.;Zheng,X.;Wang,J.;Yuan,W.;Jiang,Z.trainedregressionmodel,theproposedplatformwasabletoSurface-ModifiedYZeolite-FilledChitosanMembraneforDirectreducetheuncertaintyofthepredictionfortheunlabeleddataMethanolFuelCell.J.PowerSources2007,173(2),842−852.setby40%and58%forthebulkandshearmoduli,respectively.(6)Siriwardane,R.V.;Shen,M.-S.;Fisher,E.P.;Losch,J.AdsorptionImportantly,thiswasachievedbylabelingonly0.15%oftheofCO2onZeolitesatModerateTemperatures.EnergyFuels2005,19totaldata.Theproposedplatformoffersauniversaltoolto(3),1153−1159.acceleratethediscoveryofnewmaterial.Moreover,itcanbe(7)Earl,D.J.;Deem,M.W.TowardaDatabaseofHypotheticalappliedtofindothermaterialswithunprecedentedmechanicalZeoliteStructures.Ind.Eng.Chem.Res.2006,45(16),5449−5454.properties,suchasunusuallyhighPoisson’sandthermal(8)Pophale,R.;Cheeseman,P.A.;Deem,M.W.ADatabaseofNewexpansionratios.The23newlydiscoveredzeolitestructuresZeolite-likeMaterials.Phys.Chem.Chem.Phys.2011,13(27),12407−12412.haveunprecedentedshearmoduli,demonstratingthehuge(9)IZAStructureCommission.DatabaseofZeoliteStructures.potentialofusingzeolitesasfunctionalmaterialsinmechanicsashttp://www.iza-structure.org/databases/.wellasinchemistry-relatedapplications.(10)Jensen,Z.;Kim,E.;Kwon,S.;Gani,T.Z.H.;Román-Leshkov,Y.;Moliner,M.;Corma,A.;Olivetti,E.AMachineLearningApproachto■ASSOCIATEDCONTENTZeoliteSynthesisEnabledbyAutomaticLiteratureDataExtraction.*sıSupportingInformationACSCent.Sci.2019,5(5),892−899.TheSupportingInformationisavailablefreeofchargeat(11)Evans,J.D.;Coudert,F.-X.PredictingtheMechanicalPropertieshttps://pubs.acs.org/doi/10.1021/acs.jpclett.1c00339.ofZeoliteFrameworksbyMachineLearning.Chem.Mater.2017,29(18),7833−7839.Detaileddescriptionofelasticpropertiesderivation,(12)Gaillac,R.;Chibani,S.;Coudert,F.-X.SpeedingUpDiscoveryofvalidationofthecurrentmethod,predictivemodelAuxeticZeoliteFrameworksbyMachineLearning.Chem.Mater.2020,accuracy,repeatabilityofthepredictionmodel,distribu-32(6),2653−2663.tionofpredictedbulkandshearmoduli,andlistof(13)Kim,B.;Lee,S.;Kim,J.InverseDesignofPorousMaterialsUsingsuperiorzeolitestructures(PDF)ArtificialNeuralNetworks.Sci.Adv.2020,6(1),eaax9324.2338https://dx.doi.org/10.1021/acs.jpclett.1c00339J.Phys.Chem.Lett.2021,12,2334−2339

5TheJournalofPhysicalChemistryLetterspubs.acs.org/JPCLLetter(14)Carr,D.A.;Lach-hab,M.;Yang,S.;Vaisman,I.I.;Blaisten-Barojas,E.MachineLearningApproachforStructure-BasedZeoliteClassification.MicroporousMesoporousMater.2009,117(1),339−349.(15)Yang,S.;Lach-hab,M.;Vaisman,I.I.;Blaisten-Barojas,E.IdentifyingZeoliteFrameworkswithaMachineLearningApproach.J.Phys.Chem.C2009,113(52),21721−21725.(16)Daeyaert,F.;Ye,F.;Deem,M.W.Machine-LearningApproachtotheDesignofOSDAsforZeoliteBeta.Proc.Natl.Acad.Sci.U.S.A.2019,116(9),3413−3418.(17)Moliner,M.;Román-Leshkov,Y.;Corma,A.MachineLearningAppliedtoZeoliteSynthesis:TheMissingLinkforRealizingHigh-ThroughputDiscovery.Acc.Chem.Res.2019,52(10),2971−2980.(18)Muraoka,K.;Sada,Y.;Miyazaki,D.;Chaikittisilp,W.;Okubo,T.LinkingSynthesisandStructureDescriptorsfromaLargeCollectionofSyntheticRecordsofZeoliteMaterials.Nat.Commun.2019,10(1),4459.(19)Ambekar,R.S.;Oliveira,E.F.;Kushwaha,B.;Pal,V.;Machado,L.D.;Sajadi,S.M.;Baughman,R.H.;Ajayan,P.M.;Roy,A.K.;Galvao,D.S.;etal.OntheMechanicalPropertiesofAtomicand3DPrintedZeolite-TemplatedCarbonNanotube.Networks.Addit.Manuf.2021,37,101628.(20)Coudert,F.-X.SystematicInvestigationoftheMechanicalPropertiesofPureSilicaZeolites:Stiffness,Anisotropy,andNegativeLinearCompressibility.Phys.Chem.Chem.Phys.2013,15(38),16012−16018.(21)Tan,J.C.;Bennett,T.D.;Cheetham,A.K.ChemicalStructure,NetworkTopology,andPorosityEffectsontheMechanicalPropertiesofZeoliticImidazolateFrameworks.Proc.Natl.Acad.Sci.U.S.A.2010,107(22),9938−9943.(22)Kresse,G.;Furthmüller,J.EfficiencyofAb-InitioTotalEnergyCalculationsforMetalsandSemiconductorsUsingaPlane-WaveBasisSet.Comput.Mater.Sci.1996,6(1),15−50.(23)Kresse,G.;Furthmüller,J.EfficientIterativeSchemesforAb-InitioTotalEnergyCalculationsUsingaPlane-WaveBasisSet.Phys.Rev.B:Condens.MatterMater.Phys.1996,54(16),11169−11186.(24)Perdew,J.P.;Burke,K.;Ernzerhof,M.GeneralizedGradientApproximationMadeSimple.Phys.Rev.Lett.1996,77(18),3865−3868.(25)Ke,G.;Meng,Q.;Finley,T.;Wang,T.;Chen,W.;Ma,W.;Ye,Q.;Liu,T.-Y.LightGBM:AHighlyEfficientGradientBoostingDecisionTree.InAdvancesinNeuralInformationProcessingSystems30;Guyon,I.,Luxburg,U.V,Bengio,S.,Wallach,H.,Fergus,R.,Vishwanathan,S.,Garnett,R.,Eds.;CurranAssociates,Inc.:2017;pp3146−3154.(26)Wang,D.;Zhang,Y.;Zhao,Y.LightGBM:AnEffectiveMiRNAClassificationMethodinBreastCancerPatients.InProceedingsofthe2017InternationalConferenceonComputationalBiologyandBio-informatics;2017;pp7−11.(27)Sun,X.;Liu,M.;Sima,Z.ANovelCryptocurrencyPriceTrendForecastingModelBasedonLightGBM.Financ.Res.Lett.2020,32,101084.(28)Pedregosa,F.;Varoquaux,G.;Gramfort,A.;Michel,V.;Thirion,B.;Grisel,O.;Blndel,M.;Prettenhofer,P.;Weiss,R.;Dubourg,V.;etal.Scikit-Learn:MachineLearninginPython.J.Mach.Learn.Res.Pref.2011,39,2825−2830.(29)Müller,W.G.CollectingSpatialData:OptimumDesignofExperimentsforRandomFields;SpringerScience&BusinessMedia:2007.(30)Shahriari,B.;Swersky,K.;Wang,Z.;Adams,R.P.;DeFreitas,N.TakingtheHumanoutoftheLoop:AReviewofBayesianOptimization.Proc.IEEE2016,104(1),148−175.(31)Snoek,J.;Larochelle,H.;Adams,R.P.PracticalBayesianOptimizationofMachineLearningAlgorithmsJasper.Adv.NeuralInf.Process.Syst.2012,2951−2959.(32)Wang,S.;Bai,P.;Sun,M.;Liu,W.;Li,D.;Wu,W.;Yan,W.;Shang,J.;Yu,J.FabricatingMechanicallyRobustBinder-FreeStructuredZeolitesby3DPrintingCoupledwithZeoliteSoldering:ASuperiorConfigurationforCO2Capture.Adv.Sci.2019,6(17),1901317.2339https://dx.doi.org/10.1021/acs.jpclett.1c00339J.Phys.Chem.Lett.2021,12,2334−2339