MachineLearningguidedhigh-throughputsearchofnon-oxidegarnets

Abstract

Garnets,knownsincetheearlystagesofhumancivilization,havefoundimportantapplicationsinmoderntechnologiesincludingmagnetorestriction,spintronics,lithiumbatteries,etc.Theoverwhelmingmajorityofexperimentallyknowngarnetsareoxides,whileexplorations(experimentalortheoretical)fortherestofthechemicalspacehavebeenlimitedinscope.Akeyissueisthatthegarnetstructurehasalargeprimitiveunitcell,requiringanenormousamountofcomputationalresources.Toperformacomprehensivesearchofthecompletechemicalspacefornewgarnets,wecombinerecentprogressingraphneuralnetworkswithhigh-throughputcalculations.Weapplythemachinelearningmodeltoidentifythepotential(meta-)stablegarnetsystemsbeforesystematicdensity-functionalcalculationstovalidatethepredictions.Inthisway,wediscovermorethan600ternarygarnetswithdistancestotheconvexhullbelow100 meV/atomwithavarietyofphysicalandchemicalproperties.Thisincludessulfide,nitrideandhalidegarnets.Forthese,weanalyzetheelectronicstructureanddiscusstheconnectionbetweenthevalueoftheelectronicbandgapandchargebalance.

suppl

I Introduction

Garnetscanbefoundthroughouttheworldindiversegeologicalenvironments,andhavebeenknownsinceprehistorymainlyduetotheiruseinjewelryasgemstones.Theyarealsorelativelyhardminerals,apropertythatmakesthemusefulforaseriesofindustrialapplications,suchasinwaterjetcuttingorasabrasives.Generally,thegarnetscrystallizeinacubicstructure(spacegroupIa\overline{3}d)withchemicalcompositionA_3B_2(B’C_4)_3,wheretheAatomsarelocatedinthe24cdodecahedralsites,theBatomsareinthe16aoctahedral,andB’atomsoccupythe24dtetrahedralsites.Internarygarnets,BandB’sitesareoccupiedbythesamechemicalelement.Around1950 yoder1951complete; bertaut1956structure; 10.1103/PhysRev.110.1311somerare-earthgarnets,especiallyyttrium-basedmaterials,startedtoattractattention.ThosegarnetshaveageneralformulaofRE3B2(BO4)3whereREstandsforrare-earthandBisa3dmagnetictransitionmetal(usuallyiron)oragroupIIIAelement.Amongthese,oneofthemostusedonesisyttriumaluminumgarnet(YAG),Y3Al2(AlO4)3,usedasasyntheticsimulanttodiamondduetoitshighrefractiveindex(>1.8) palik1997.DopedYAGswithotherrare-earthelementshavefoundnumerousapplicationsaslasingmediainmodernmedicallaserdevices luke2019lasersorintunableopticaldevices 10.1016/j.ceramint.2020.01.242; 10.1016/j.jallcom.2019.151903; 10.1002/bio.3108; 10.1063/1.111283.Otherimportantcompounds,withinterestingferrimagneticproperties,aretherare-earthirongarnets(RE3Fe2(FeO4)3,RIG).IntheRIGstructure,fiveFeatomsoccupytwodifferentsublattices,andtheantiferromagneticcouplingbetweensub-latticesandferromagneticcouplingwithinthesublatticeleadstoaferrimagneticconfiguration.RIGscandisplayaratherhighCurietemperature(around560 K 10.1109/TMAG.2011.2180732),andsomesystemsexhibitgiantmagnetorestriction 10.1016/0304-8853(86)90456-7.Moreover,RIGsmaterialshaveabandgapwithvaluesaround2.6to2.9 eV 10.1134/1.2161290; 10.1016/0038-1098(74)90760-1.Amongthesematerialsyttriumirongarnet(YIG)standsoutbecauseithasanexceptionallylowGilbertdamping.YIGhasfirstbeenusedasbulkmaterialinopticalinsulators,circulatorsandFaradayrotators.Sincethelasttwodecades,YIGisalsomoreandmorefrequentlyusedasthinfilmmaterialforspintronicapplications wu2013recentbecauseitallowsthetransmissionofspincurrentsalthoughbeinganinsulatorbyitself.InrecentyearswehavewitnessedtheattempttoreplaceyttriumbylanthanidestoincreasethespinorbitcouplingandintroduceevenDzyaloshinskii–Moriyainteractionsinhybridsystems.AnotherinterestinggroupofquaternarygarnetsisthelithiumgarnetsLN3M2(LiO4)3,whereLNisalanthanideandMiseitherTe,Ta,orNb.WithpartialLi-fillingoftheinvasivepositionsoftetrahedralanddistortedoctahedralsites,thestuffedlithiumgarnets(LN3M2Li2(LiO4)3)haveapromisinglithium-ionconductivityandchemicalstability,showingpotentialassolidelectrolytesinLi-batteries 10.1016/j.est.2020.102157; 10.3389/fchem.2020.00468.Themanyapplicationsofgarnets,andofYIGinparticular,hasincreasedtheneedforgarnetswithdifferentproperties.Unfortunatelythedepositionofgarnetthinfilmswithhighqualityisonlypossibleongarnetsubstratesduetothespecialcrystalstructure Schmidt2020; Althammer2018; Serga2010.Moreoverforspintronicsapplicationsitwouldbeahugestepforwardtobeabletopair,forexample,thinYIGfilmsinhybridstructureswithotherconductingfilmsbeingeithermetallicorhavingsuchalowbandgapthatreasonableelectronconductivityisachievableatroomtemperature Schmidt2020; Althammer2018; Serga2010.Againtoachieveaperfectinterfaceinthesestructuresonewouldneedtocreateanall-garnethybridwhichiscurrentlypreventedbytheobviouslackofroomtemperatureconductingormetallicgarnets.Suchamaterialclasswoulddramaticallyextendtheapplicabilityofgarnetthinfilms.However,despitethesediverseapplicationsandthetechnologicalrelevanceofgarnets,mostoftheresearchingarnetsisconfinedtooxides ye2018deep,andonlyafewhalides(alsocalledcryolithionites)areknownexperimentally 10.2138/am.2013.4201.Thisiscanbeeasilyunderstood,asoxidesareusuallysimplertoworkwithunderambientexperimentalconditions.Furthermore,acomputationalhigh-throughputsearchofnewcompositionsforthegarnetprototypeischallenging,asthegarnetcubicprimitiveunitcellcontains80atoms,whichisanorderofmagnitudelargerthanmoststructureprototypesusedinrecenthigh-throughputsearchessuchas(double-)perovskites CGAT; schmidt2017,(half)-Heuslers doi:10.1021/acs.chemmater.6b02724,dichalcogenides 10.1039/D0MA00999G,etc.Fortunately,withtheaidfromstate-of-the-artmachinelearningtechniques,theproblemofsearchingthroughtheentirecombinatorialchemicalspacecanbesignificantlyaccelerated.Withpre-trainedmachinemodels,wecanfiltermillionsofcompositionsaccordingtothepredictedstabilitywithoutperformingcostlydensityfunctionaltheory(DFT)calculationsforallthecompositions.Nevertheless,DFTvalidationofthestablecompoundsisstillnecessaryasapost-processingstep.Inthepresentpaper,wefollowedsuchaproceduretoexplorepossible(meta-)stablecompoundsbeyondoxy-garnets.Therestofthepaperisstructuredasfollows.InsectionIIweexplainthemachinelearningmodelandthecomputationalmethodsweapplied.InsectionIIIwepresentthemostinterestingcrystalphasesweuncoveredinourworkanddiscussthepotentialapplicationsofthenewproposedcompounds.Finallywepresentourconclusionsandanoutlook.

II Methods

II.1 MachineLearningModel

Inthisworkweappliedcrystalgraphattentionnetworks,developedandpretrainedinRef. CGAT,topredictthermodynamicallystablematerials.Thenetworksuseanattention-basedmessagepassingapproachbasedonthecrystalgraphrepresentationofthecrystalstructure.Replacingthenormaldistanceinformationthatistypicallyusedasedge-representationincrystalgraphnetworkswithsolelythegraphdistanceoftheatomstotheirneighborsallowsforprecisepredictionsofunrelaxedstructures.Asgarnetscrystallizeinacubicstructurethelistofneighborsandconsequentlythegraphdistancesaremostlyconstantthroughoutthegeometryrelaxation.Thisremovestheneedtoperformpredictionswithmultiplecellconstantratios.WestudyallgarnetcompositionsA3B5C12consideringallelementsuptoBi,excludingonlytherare-gases,andspanningaspaceofaround550kpossiblecompounds.Obviously,mostofthesehypotheticalcompoundsarehighlyunstablefromthethermodynamicpointofview.AsdiscussedinRef. CGAT,currentmachinelearningapproachesforthepredictionofthermodynamicstabilitysufferfromaconsiderableerrorduetothestrongbiaspresentintheexistingdatasets.Tocircumventthisproblem,weusethefollowingworkflow:(i) Themachine-learningmodelisusedtopredictthedistancetotheconvexhullofstabilityforall550kcompounds.Atthestartweusethepre-trainedmachinefromRef. CGAT;(ii) WeperformDFTgeometryoptimizationstovalidateallcompoundspredictedbelow200 meV/atomfromthehull;(iii) WeaddthesecalculationstoadatasetcontainingallDFTcalculationsforgarnetsystems.(iv) Weusetransferlearningtotrainanewmodelusingthisdatasetwithatraining/validation/testingsplitof80%/10%/10%.(v) Thecycleisrestarted.Intotalwerepeatedthecyclethreetimes.Inthefirst,weperformedDFTcalculationsfor3320compounds.Themeanabsoluteerror(MAE)oftheinitialpre-trainedmachinewas0.497 eV/atom.Thisisaveryhighvalue,thatwasexpectedastherewereveryfewgarnetsinthedatasetusedinRef. CGAT,andtheyspannedaverysmallchemicalspace.Inthesecondcyclewevalidated7336compounds,andthetransfer-learningmodelperformedmuchbetter,withaMAEof0.064 eV/atom.Finally,inthethirdcyclewecomputed3844materialswithDFT.ThefinalmodelhadaMAEof0.058 eV/atom,showingthatthetranfer-learningworkflowismeaningfulandconvergesquickly.

II.2 DFTCalculations

WeperformDFTcalculationsusingthepackagevaspwithPAW pawdatasetsofversion5.2.andwiththePerdew-Burke-Ernzerhof pbe(PBE)exchange-correlationfunctional.FollowingtheMaterialsProject MP_Jain2013recommendations,weuseextraon-sitecorrectionsforoxides,fluoridescontainingCo,Cr,Fe,Mn,Mo,Ni,V,andW.Theon-sitecorrectionsarerepulsiveandcorrectthed-statesbyrespectively3.32,3.7,5.3,3.9,4.38,6.2,3.25,and6.2eV.Acutoffof520 eVisappliedtotheplanewaves,and\Gamma-centeredk-pointgridswithauniformdensityof1000 k-pointsperreciprocalatomareusedtosampletheBrillouinzone.Forgeometryoptimizationsallforcesareconvergedtolessthan0.005 eV/Å.Allcalculationsareperformedwithspin-polarization,startingfromaferromagneticground-stateasiscustomaryinhigh-throughputsearches.Unfortunately,thismeansthat,inmostcases,antiferromagneticorferrimagneticsystemswillconvergetoanincorrectferromagneticground-state.Thisisimportant,inourcontext,particularlyforferrimagneticgarnetshavinga3dtransitionmetalsuchasFe,Ni,Co,Cr,Mn,orVintheBposition.Wenotethatnotonlythespin-stateofthesegarnetsbutalsoofanti-andferrimagneticsystemsontheconvexhullaretreatedincorrectly.Consequently,theestimationofE_\text{hull}forferrimagneticgarnetsisfarlessaccuratethanfornon-magneticones.Forexample,theexperimentallyknowGd3Fe5O12ispredictedtobemorethan1 eVabovethehullaccordingtoMaterialsProjectdatabase MP_Jain2013,avaluethatiscertainlygrosslyoverestimated.ToproperlyestimateE_\text{hull}fortheferrimagneticgarnetswouldrequireobtainingthecorrectmagneticorderingforaportionoftheconvexhullaswellasforthegarnets.Thisisacomplexandcomputationallyexpensivetask,thatiswellbeyondthescopeofthiswork.Therefore,wemadethechoicetorestrictourdiscussiontosystemsnotcontainingthe3dmetalsmentionedaboveintheB-site.Itiswell-knownthattheelectronicbandgapscalculatedwiththePBEfunctionalareseverelyunderestimated benchmark1.Therefore,toobtainamorereliableestimationofthisimportantphysicalpropertyweperformedcalculationswiththemodifiedBecke-Johnson(mBJ)approximation PhysRevLett.102.226401,asthisisbynowrecognizedasoneofthemostaccuratefunctionalforthistask benchmark2.TocalculatetheaveragedcarriereffectivemassesfromtheinterpolatedeigenvalueswefollowtheapproachofRef. HautierEff.Consideringatemperatureof300 K,wededucethechemicalpotentialrequiredtoreachareferencecarrierconcentration(10{}^{18} cm{}^{-3})byusingBoltzTraP2.Weuseak-pointmeshwitharegulardensityof2000 k-pointsperreciprocalatomandinterpolatedthecalculatedeigenvaluesusingBoltzTraP2 Boltztrap; madsen2018boltztrap2.Thecalculatedaveragedcarriereffectivemassescanbeseenastheintrinsictendencyforcreatingmobilechargecarriersinmaterials HautierEff.

III ResultsandDiscussion

III.1 StabilityandChemistry

(a)

(b)

(c)

Figure 1: Thedistributionofthedistancetotheconvexhull(E_\text{hull})for(a) allcalculatedcompositions,(b) halides,and(c) chalcogenides/nitrides.Thecutofftofilter(meta-)stablesystems(100 meV/atom)isshownasdashedverticalline.

Table 1: Number(meta-)stablesystems(N_\text{stable})below100 meV/atomfromtheconvexhullofthermodynamicstabilityandhigh-throughputsuccessrate(R)foreachcategory.

Category	N_\text{stable}	R(%)
Sulfides	70	14
Selenides	68	15
Tellurides	28	7
Nitrides	64	35
Fluorides	62	17
Chlorides	68	16
Bromides	72	17
Iodides	68	15
Hydrides	69	8
Total	569	14

Intotalthemachine-learningmodelpredictedaround12300compositionsbelow200 meV/atomfromtheconvexhullofthermodynamicstabilitythatwerenotpresentintheMaterialsProjectdatabase MP_Jain2013norintheInorganicCrystalStructureDatabase ICSD(ICSD).AllthesecalculationscanbedownloadedfromtheMaterialsCloudrepository linktodata.AftertheDFTvalidationcalculationswere-evaluatedthedistancetotheconvexhull(E_\text{hull})ofthesecandidatesusingthemuchmorecompleteconvexhullofRef. CGAT.ThehistogramofthevaluesE_\text{hull}isshowninFig. 0(a).Wealsoseparatethesystemsintosulfides,selenides,tellurides,nitrides,chlorides,bromides,iodides,andhydrides.Thesecomprisethemajorityofallsystemsfound.MostofthecandidateshaveanE_\text{hull}largerthan100 meV/atom,buttherearestillmorethanonethousand(about9%)compositionsbelowthisthreshold.Thehigh-throughputsuccessrate,thatwedefinebythenumberofcompoundsthatarewithin100 meV/atomfromtheconvexhulldividedbythetotalnumberofDFTcalculations,stoodat14%,withamaximumof35%fornitridesandaminimumof8%forhydrides.Thesenumbersprovetheefficiencyofourmachine-learningassistedhigh-throughoutsearch.ThehistogramofE_\text{hull}forthesecategoriesaswellasallcalculatedcompositionsareshowninFig. 1.ThedistributionofE_\text{hull}forallsystemsfollowsthetypicalskewedGaussianwiththepeaklocatedataround200 meV/atomandafattailthatextendsbeyond1 eV/atom,inagreementwiththeMAEerrorsforourmachine-learningmodels.IndividualdistributionsforchalcogenidesarealsoskewedGaussianspeakingataround200 meV/atom.Theamountof(meta-)stablecompoundsdecreasesfromsulfidestotellurides,whichisalsoexpectedasthisisthecommontrendofstabilityforthechalcogenides.Unlikethesituationforchalcogenides,thedistributioncurvefornitrideshasmultiplepeaks,andshowsthatthereareplentyofpotentiallystablenitridegarnets.ForhalidesandhydridesthehistogramsareagainskewedGaussiansimilartothoseofchalcogenides,butthereisnocleartrendinwhatconcernsstabilityacrossthegroup.Forhydrides,thetotalnumberofsystemsismuchlarger,butalowerpercentageofthemare(meta-)stablecomparedtothehalides.Thetotalnumberof(quasi-)stablesystemsforeachcategoryislistedasTable 1.AfulllistofthesystemscanbefoundinTable 1oftheSupplementaryInformation(SI).WealsoselectedadozenofthemtoanalyzemorecloselyinTable 3.Laterdiscussionswillmainlyfocusonthesesystems.

Table 2: Theexperimentallyknownoxy-garnets(notincludinga3dmetalintheBsite),theirICSDID,MaterialsProjectID,andpredictedmeta-stablecounterpartswithdifferentCanions.ThedistancetothehullcalculatedwithDFTisinparentheses(inmeV/atom).

Formula	ICSDID	MPID	Counterparts
Y3Al5O12	20090,41144,41145,67102,67103,	mp-3050	Y3Al5S12(-4);Y3Al5Se12(36);Y3Al5Te12(84)
	93634,93635,170157,170158,236589,
	280104,17687,17688,17689,17690,
	74607,31496
La3Al5O12		mp-780432	La3Al5S12(-4);La3Al5Se12(19);La3Al5Te12(49)
Eu3Al5O12	245326	mp-21757	Eu3Al5S12(39);Eu3Al5Se12(93)
Tb3Al5O12	33602	mp-14387	Tb3Al5S12(-4);Tb3Al5Se12(29);Tb3Al5Te12(79)
Er3Al5O12	170147,280606,170146,62615	mp-3384	Er3Al5S12(6);Er3Al5Se12(50);Er3Al5Te12(98)
Gd3Al5O12	192184	mp-14133	Gd3Al5S12(-16);Gd3Al5Se12(21);Gd3Al5Te12(66)
Ho3Al5O12	409390,33603	mp-14388	Ho3Al5S12(-1);Ho3Al5Se12(44);Ho3Al5Te12(92)
Lu3Al5O12	259144,17789,182354	mp-14132	Lu3Al5S12(28);Lu3Al5Se12(70)
Y3Ga5O12	80148,14343,185862,23852	mp-5444	Y3Ga5S12(47);Y3Ga5Se12(81)
La3Ga5O12		mp-780561	La3Ga5S12(36);La3Ga5Se12(51);La3Ga5Te12(100)
Tb3Ga5O12	20831,84875,184934	mp-5965	Tb3Ga5S12(37);Tb3Ga5Se12(75)
Sm3Ga5O12	9236,84873,291192	mp-5800	Sm3Ga5S12(29);Sm3Ga5Se12(64);Sm3Ga5Te12(93)
Nd3Ga5O12	84872	mp-15239	Nd3Ga5S12(26);Nd3Ga5Se12(54);Nd3Ga5Te12(96)
Gd3Ga5O12	9237,37145,192181,84874,184931	mp-2921	Gd3Ga5S12(32);Gd3Ga5Se12(67);Gd3Ga5Te12(93)
Lu3Ga5O12	23850	mp-14134	Lu3Ga5S12(74)
Dy3Ga5O12	409391	mp-15576	Dy3Ga5S12(49);Dy3Ga5Se12(81)
Er3Ga5O12	9238	mp-12236	Er3Ga5S12(55);Er3Ga5Se12(95)
Ho3Ga5O12	409390	mp-15575	Ho3Ga5S12(52);Gd3Ga5Se12(89)

Fromourcalculationswerecoverthemajorityoftheoxide(andhalide)garnetsthatarealreadyknownexperimentally,butwealsoobtainawealthofdifferentcompoundsnotpresentinavailabledatabases.Manyofthesesystemsareoxides,thathavebeenthesubjectofarecenthigh-throughputsearch ye2018deepwithresultssimilartoours.Interestingly,wealsofindawealthofotherchalcogenides,nitrides,halides,andevenhydridesasshowninFigs. 0(b),0(c)andinTable1.Fromthestablecompounds,severalarecloselyrelatedtotheoxidegarnetsbythechemicalsubstitutionofoxygenbyanotherchalcogen.ThesegarnetsalongwiththeircounterpartsarepresentedinTable 2.Theexistenceofsuchcompoundsisexpectedduetothechemicalsimilarityamongchalcogens.Inthosechalcogenides,thedodecahedralsites(siteAinA3B5S12)aremostlyoccupiedbyrareearthelements,andaccordingtotheelementoccupyingtheoctahedralandtetrahedralsites(siteB),thechalcogenidescanbefurtherdividedintoseveralcategories.Thenumbersof(meta-)stablesystemsforeachcategorydecreaseinthefollowingorder:occupyingBwithgroupIIIAelements(Al,Ga,In,Tl),groupIVAelements(Ge,Sn,Pb),groupVAelements(As,Sb,Bi),andtransitionmetals(Ag,Cu,Sc,Ti).ThepreferenceofgroupIIIAelementsforsiteBcanbeunderstoodbysimplechargecompensationarguments.Themostcommonoxidationstateofthechalcogensis-2,whiletherareearthelementsinsitesAare+3:thecompositionreachesthe“balanced”or“compensated”stateifBisintheoxidationstate+3.Moreover,forbalancedA{}^{\text{III}}_3B{}^{\text{III}}_5C{}^{\text{-II}}_{12}chalcogenidecompositions,onewouldexpectthecompoundstobesemiconductors.Thisisindeedwhatwefind(seeanexampleinsection III.2).Besidesthechalcogenides,wediscover64stablenitridesasseeninTable 1.Asdiscussedabove,toreachabalancedoxidationstate,elementswithhigheroxidationstateshouldbefavoredtocombinewith-3oxydationstateofnitrogen.IndeedfornitridegarnetsA3B5N12,thepositionAismostlyoccupiedby+2or+3chemicalelements,andforpositionBthemajorityofmeta-stablesystemshaveelementswithoxidationstate+6(MoandW).Althougha+6elementisrequiredtoachieveabalancedstate,wealsofindthatrelativelystablecompoundsarepossiblefor+5(NbandTa),and+7(Re).Wecouldarguethatthebalancednitridesshouldbesemiconducting.However,duetothelowelectronegativityofnitrogenthegapmayclose(seeonesuchexampleinsection III.2).WecanalsoidentifyseveralhalidesandhydridesfromFigure 0(c).However,halogens(andhydrogeninhydrides)haveanoxidationstateof-1,whichmakesitmoredifficulttoreachabalancedoxidationstatewiththeA3B5C12stoichiometry.OneviableschemeiswithanAelementthatis+1,whiletheBelementsin16aand24dWycoffpositionsarerespectively+1and+3,i.e.havingtheformofA{}^{\text{I}}_3B{}^{\text{I}}_2(B{}^{\text{III}}C{}^{\text{-I}}_4)_{3}.Chemicalelementsexhibitingboth+1and+3oxidationstatesarequiterare.Nevertheless,westilldiscoveredsomemeta-stablesemiconductinghalidegarnets,forexampleK3In5F12.Thereareothercompositionsthatdonotbelongtoanyofthediscussedclasses,suchasforexampleSr12Zn3H5.Mostofthemare“inverted”-garnets,i.e.withcationsinsteadofanionsoccupyingtheC-sites,andhaveacomparativelyhigherE_\text{hull}thanregulargarnets.Furthermore,onlyafewanti-garnetswiththeAandBsitesbothoccupiedbythechemicalelementsofthenitrogengroupcouldpotentiallybecomeoxidationstatebalanced.Wehavetoagainemphasizethat,inordertoform(meta-)stableorinsulating/semiconductingcompounds,chargecompensationisneitheranecessarynorasufficientcondition,andwefindmanyexceptionsinTableS1intheSupplementaryInformation.However,itgivesusasimple,intuitiveargumenttounderstandwhyasystemisstabilizedorhasanelectronicbandgap.Furthermore,wehavetokeepinmindthatuncompensatedsystemsmightbefurtherstabilizedthroughdefects,suchasvacancies.Tosimplifyourdiscussion,weleavesuchpossibilitiestofutureworks,andfocusinthefollowingon(meta-)stableregulargarnetssystemswhichcouldhavebalancedcharges,specificallychalcogenides(exceptoxides),halides,hydrides,andnitrides.

Table 3: Thecalculatedlatticeconstant(a,inÅ),bandgapcalculatedwithPBE(Gap{}^{\text{PBE}})andMBJ(Gap{}^{\text{MBJ}})functional(inunitofeV)),distancetotheconvexhull(E_\text{hull}inmeV/atom),effectiveelectron(m_\text{e}^{*})andhole(m_\text{h}^{*})masses(inunitofm_\text{e}^{0}),forsomeselected(meta-)stablesulfideandnitridegarnets,dataforY3Al5O12isalsolistedforcomparison.

Formula	a	Gap{}^{\text{PBE}}	Gap{}^{\text{MBJ}}	E_\text{hull}	m_\text{e}^{*}	m_\text{h}^{*}
Y3Al5O12	12.125	4.53	6.12	0	1.3	6.8
Y3Al5S12	14.932	2.08	3.00	0	0.7	1.5
Y3Ga5S12	15.073	1.32	2.45	46	0.5	3.2
Y3In5S12	15.709	1.35	2.38	46	0.5	2.6
Y3Al5Se12	15.752	1.46	2.15	36	0.5	1.5
Y3Al5Te12	17.120	0.60	1.07	84	0.3	2.7
Y3Ge5S12	15.324	0.00	0.00	81	–	–
Ca3W5N12	12.928	0.93	1.57	0	0.9	3.2
Ca3Re5N12	12.896	0.00	0.00	9	–	–
La3Nb5N12	13.297	0.00	0.00	99	–	–
K3In5F12	14.898	2.41	3.39	59	0.8	6.1
K3In5I12	19.977	0.00	0.00	61	–	–
Mg3Rh5H12	11.008	0.00	0.00	51	–	–
Y3Rh5H12	11.581	0.00	0.00	92	–	–

III.2 ElectronicStructure

WeillustratethepossibleelectronicstructuresofourgarnetsthroughafewselectedexamplesdepictedinFigs. 2–4.GenerallytheelectronstatesaroundtheFermilevelingarnetscanbeclassifiedinthreecategories:(i) froms–p,p–p,orevend–pbondingorbitalsbetweenBandCatoms,(ii) fromcorrespondinganti-bondingbetweenBandCatoms,and,(iii) fromd-electronsfromAatomsiftheseared-blockorf-blockmetals.Fromthesimplesttight-bindingmodel,andasexpectedfromk\cdot ptheory,weknowthatthelargerthedifferencebetweentheelectronegativityofBandCandtheshorterB–Cbond-length,thelargertheseparationsbetweenstates(i)and(ii).Forthed(orf)blockelementsoccupyingtheA-site,thepositionofstates(iii)canbebetween(i)and(ii),overlappingwith,orevenabovethelatter.Inchargecompensatedsituations,thebands(i)arecompletelyfilled,while(ii)and(iii)(ifapplicable)bandsareempty,resultingininsulating/semiconductingsystems.Otherwise,depletionofbands(i)orfillingofbands(ii)or(iii)canhappen,leadingtometallicsystems.Althoughmanydeviationsfromthissimpletight-bindingpictureappear,wewillseehowthesegeneralpatternsareusefultounderstandtheband-structures.

(a) Y3Al5O12

(b) Y3Al5S12

(c) Y3Ge5S12

(d) Y3Ga5S12

Figure 2: CalculatedmBJelectronicbandstructuresforselectedchalgogenidegarnets.TheFermilevelissetatzero.

(a) Ca3W5N12

(b) Ca3Re5N12

Figure 3: CalculatedmBJelectronicbandstructuresforselectednitridegarnets.TheFermilevelissetatzero.

ForchalcogenidegarnetswechoseasrepresentativeexamplesY3Al5S12,Y3Ge5S12,andY3Ga5S12,whilemoreexamplescanbefoundintheSI.Theband-structuresofthesecompoundsareshowninFig 2,togetherwithY3Al5O12forcomparison.Theoxidation-statebalancedY3Al5CH12(CH=S,Se,Te)garnetsaresemiconductorsasexpected.ForY3Al5S12thebandstructureisshowninFig. 1(b).SimilartoitsoxidecounterpartinFig. 1(a),theY–dstatesdominatetheconductionbands(CB)slightlyhybridizingwithAl(p)andS(p)-states.Thesebandsarefromthetype(iii)statesasdiscussedabove.Thevalencebands(VB)aroundFermilevelaremainlycomposedofthelocalizedanionicp–statesoftype(i),alsorepresentingthetypicalsituationdescribedabove.FromY3Al5O12toY3Al5Te12(seeFig. S1inSI),followingthedecreasingtrendofelectronegativityforchalcogens,thebandgapshrinksandthebandedgesbecomemoredispersed.TheelementsoccupyingtheBpositionincharge-compensatedchalcogenidesalsohaveacrucialeffectontheelectronicstructure.Forexample,Y3B5S12(B=Al,Ga,In,Tl,seeFig. S1intheSI)allhaveadirectgapat\GammawiththesoleexceptionofY3Al5S12whichhasanindirectH–\Gammagap.Moreover,goingfromAltoTlthegapdecreases,whilethecontibutionofthes-statesfromtheBatomstothebottomoftheconductionbandsincreases,leadingtomoreextendedbandsandtolowereffectiveelectronmasses.Furthermore,forY3Tl5S12wecanseethattheY(d)bandsareabovetheTl(s)-S(s)displayingmoredisperseanti-bondingmixing,anresultinginverylowelectroneffectivemasses(m^{*}_\text{e}=0.19).Forchalcogenideswithunbalancedoxidationstates,suchasY3Ge5S12,similarfeaturescanalsobeobserved(seeFig. 1(c)).BetweentheemptyY(d)-S(p)hybridizationstatesoftype(iii)andthefullS(p)dominatedtype(i)bandsarethestatesfromGe(s)andS(p)anti-bondingoftype(ii).Asdiscussedabovetheyarepartiallyoccupied,sothesystemismetallic.Itispossiblethatthesystemmightre-establishabalancedoxidationstatebycreatingGevacanciesthusbecomingsemiconducting,thoughadetailedinvestigationofsuchapossibilityisbeyondthescopeofthepresentpaper.

(a) K3In5F12

(b) K3In5I12

(c) Mg3Rh5H12

(d) Y3Rh5H12

Figure 4: CalculatedmBJelectronicbandstructuresforselectedhalideandhydridegarnets.TheFermilevelissetatzero.

SomerepresentativesofthebandstructuresfornitridesgarnetsareshowninFig. 3.TheoxidationstatecompensatedCa3W5N12issemiconducting.Theanti-bondingp-dbandsformingtheCBseparatealsointotwomanifolds:thelowerpartismainlyconstructedfromthetungstenonthetetrahedralsites,andtheupperpartmainlyfromtheoctahedralW.TheVBare,asusual,mostlycomposedofN(p)states.Forchargeunbalancednitrides(suchasCa3Re5N12)thathavemorevalenceelectronsfromRecomparedtoW,theFermilevelpassesthroughthepartiallyoccupiedp-danti-bondingbandsandthesystembecomesmetallic.However,theseparationbetweenp-dbondingandanti-bondingbandscanstillbeseenaswellasthedoublemanifoldsofp-danti-bondingstates.WhenachemicalelementwithlessvalenceelectronsreplacesW,forexample,inLa3Nb5N12(seeFig. S1inSI),theminorityspinchannelofp-dbondingstatesispartiallyempty,andthesystemactuallybecomesahalf-metal.Forhalides,weshowthebandstructureofK3In5F12andK3In5I12asexamples.Apparently,theformercompoundhasunbalanceoxidationstates,howeveritissemiconducting.TheBadernetchargesoftheInatoctahedralandtetrahedralsitesare+1and+2.4,respectively.TheInintheB-siteisthereforeinthe+1and+3oxidationstates,reachingchargecompensationasdiscussedabove.ThetopofthevalencebandsaremainlycomposedofIn(s)-F(p)anti-bondingstates,whichseparatefromthelowervalancebands,formedbylocalizedF(p)-In(p)bondingstates,byaround4 eV.Unlikeinmostsulfidesandnitrides,whered-statesfromAatomsdominatethebottomoftheconductionbands,inK3In5F12thebottomoftheCBismainlyformedbyIn(p)-F(p)anti-bondinghybridizedstates.ThisisbecausetheK(s)-F(p)anti-bondingstateshaveamuchhigherenergythand-panti-bondingstatesinthosesulfidesandnitrides.WhenreplacingFwithI,theBadernetchargesfortheInatomsinoctahedralandtetrahedralsitesarerespectively+0.7and+1.0,showingthatthe+1/+3oxidationstatesisnotpossibleinK3In5I12,andthesystembecomesmetallic.ThetopvalencebandformedbyIn(p)-I(s)anti-bondingstatesseparatesintwomanifolds.TheupperpartcomesfromtheInatomsintetrahedralsitesandthelowerbelongstotheoctahedralones.ThiscanbepartiallyexplainedbythefactthattheIn–Ibondlengthislargerinthelatter.InFig 4weshowthebandstructureofselectedexamplesofhydridegarnets.ForbothMg3Rh5H12andY3Rh5H12theseparationbetweenH(s)-Rh(d)mixedbandsandcationRh(d)dominatedbandsarestillpresent,similartothesituationsinotheruncompensatedgarnets.Thesetwosystemsarealsobothchargeuncompensated,the“extra”electronspartiallyfillingtheRh(d)dominatedbands,leadingtometallicsystems.Anotherwaytore-establishcompensationmightbetoforcemoreHatomstooccupytheinterstitialsites,butagain,weleaveexplorationofthispossibilitytofutureinvestigations.

IV Conclusions

Weperformedamachine-learningassistedhigh-throughputinvestigationofternarygarnets.Weconcentratedinnon-oxides(thathavebeenstudiedpreviously)andinferromagneticorparamageticcompounds.Wefindawealthofsystemsontheconvexhull(i.e.,thermodynamicallystable)orclosetoit.Thisincludeschalcogenides(withthestabilitydecreasingfromStoTe),nitrides,halides,hydrides,etc.Wealsofoundotherpossibilities,suchas“inverted”garnets,butthesewereslighltylessstablethantheconventionalphase.Thematerialstendtobesemiconducting/insulatingwhenthecompositionischargecompensated,otherwiseweobtainmetallicground-states.Thelatteronescouldbeespeciallyrelevantas,tooutknowledge,nogarnetsconductingatroomtemperatureareknow.Someofthemetallicgarnetsevenhavelatticeconstantsthataresuitabletocreatehetero-structureswithYIG.Afewchalcogenidegarnets,inparticularthesulfides,arethermodynamicstable,andarestraightforwardgeneralizationsofcommonoxidegarnets.Band-gapsareasexpectedconsiderablysmallerforthesulfides,anddecreasefurtheracrosstheperiodicgroup.Thisopensupthepossibilitytoengineerthebandgapofgarnetsbyanionicalloying,fromtheextremeultavioletoftheoxidephasetoUV-Aregimeorevenintothevisible.Wepredictseveralnitridesystemsthathaveinterestingelectronicpropertiesduetothepresenceoftransitionmetalsinveryhighchargestates.Inviewoftherecentsynthesisoftwoexoticnitrideperovskites 10.1126/science.abm3466; 10.1002/anie.202108759thatwerepredicted Sarmiento_P_rez_2015; Flores_Livas_2019withamethodsimilartotheoneusedinthispaper,weareconfidentthatalsonitridegarnetsareaccessibleexperimentally.Finally,wefindafewsemiconductinghalideswherethechemicalelementoccupyingtheoctahedralandthedodecahedralsiteisintwodifferentchargestates.Aboveall,webelievethatourworkprovesthatanexhaustivesurveyoftheternary,andperhapsalsoofthequaternary,spaceofmaterialsisnowaccessibletohigh-throughputstudies,evenforlargeandcomplexunitcells.Thisismadepossiblebymachinelearningmethods,thatalreadyachievedanoutstandingmaturityintheshorttimesincetheirfirstappearance,andthatarereachinganunprecedentedaccuracy.Weexpectthesemethodstofurtheracceleratethediscoveryofnewmaterialswithexceptionalproperties.

V Acknowledgements

TheauthorsgratefullyacknowledgetheGaussCentreforSupercomputinge.V.(www.gauss-centre.eu)forfundingthisprojectbyprovidingcomputingtimeontheGCSSupercomputerSuperMUC-NGattheLeibnizSupercomputingCentreundertheprojectpn25co.