I-AdsorbML: Indlela I-AI I-Transforming I-Catalyst Discovery for Clean Energy and Chemicals

Umbhali: Janice Lan Aini Palizhati Muhammed Shuaibi Brandon M. Wood Brook Wander Abhishek Das Matt Uyttendaele Umbhali: Lawrence Zitnick Zachary W. Ulissi Umbhali: Nami Nami I-Aini Palizhati U-Mohammed Shuaibi U-Brandon M. Wood Ukuhlobisa Abhishek Das Ukulungiswa Umbhali: Lawrence Zitnick U-Zachary W. Ulissi abstract I-catalysis ye-computational ibhizinisi ibhizinisi ibhizinisi ibhizinisi emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo emangalisayo Ukuhlolwa Ukucubungula ama-catalyst eyenziwe nge-catalyst eyenziwe ngama-catalyst eyenziwe ngama-catalyst eyenziwe nge-catalyst eyenziwe ngama-catalyst eyenziwe ngama-catalyst eyenziwe ngama-catalyst eyenziwe ngama-catalyst eyenziwe ngama-catalyst eyenziwe ngama-catalyst eyenziwe ngama-catalyst eyenziwe ngama-catalyst eyenziwe ngama-catalyst eyenziwe ngama-catalyst eyenziwe ngama-catalyst eyenziwe ngama-catalyst. , , , Ngokusho ukubaluleka kwekhwalithi design indawo, izindlela ephumelelayo ukucubungula zihlanganisa kakhulu , , , I-catalysis ye-computational inikeza umthamo yokubonisa inani elikhulu lwezinto ukuze kuhlanganise izifundo ze-time- ne-cost-intensive. 1 2 3 4 4 5 6 7 Umphumela ebalulekile kwezinto eziningi zokuqala zokufinyelela i-catalyst heterogeneous kuyinto ukucubungula ama-energy adsorption. I-energy adsorption kuyinto amandla ehlanganisiwe ne-molecule, noma adsorbate, ekusebenzeni ne-catalyst surface. I-adsorbates ikakhulukazi zihlanganisa izinyathelo ezihlukahlukene, noma ama-intermediates, ku-reaction pathway (isib. *CHO ku-CO2 ukunciphisa). I-energy adsorption ikhiqizwa ngokufinyelela i-adsorbate-surface configuration okunciphisa amandla jikelele le-structure. Ngakho-ke, i-energy adsorption iyona-energy ephakeme jikelele phakathi zonke izindawo ze-adsorbate nezimo. L Ukubonisa ukuthi izinga le-adsorption ye-reaction intermediates inokukwazi ukuba ama-descriptors eqinile ezihambelana neziphumela ze-experimental njenge-activity noma i-selectivity. , , , , Ukukwazi ukuhlaziywa izakhiwo ze-catalyst kusuka ku-principles eduze isisombululo sokusebenza kwe-catalyst screening , . 8 9 10 11 12 13 1 14 Ukubonisa amandla adsorption inikeza izinzuzo eziningana. Kukhona izindawo eziningi zokusebenza ezingenalutho ye-adsorbate ku-surface, futhi ngezinye izindawo zokuxhumana kukhona izindlela ezininzi zokufaka i-adsorbate (bheka phansi emzimbeni e-Fig. ). Uma i-adsorbate ibekwe emaphakathi ye-catalyst, ama-athomu e-adsorbate kanye nama-athomu e-surface zihlanganisa nabanye. Ukubonisa amandla e-adsorbate ye-configuration ye-adsorbate-surface eyodwa, izindawo ze-athomu kufanele zihlukaniswe kuze kube nokufinyelela emoyeni yendawo. I-Density Functional Theory (DFT) , , I-DFT iyisisombululo se-single-point calculation lapho i-output yi-system energy ne-per-atomic forces. I-relaxation ke iveza i-optimization ye-local lapho i-per-atomic forces zihlanganiswe ngokuvamile nge-DFT futhi isetshenziselwa ukuhlaziywa kwezindawo ze-atomic nge-algorithm ye-optimization (isib. i-conjugate gradient). ) kuze kube lokuphumelela energy minimum yendawo. Ukuze uthole global minimum, a strategy for sampling adsorbate surface configurations futhi / noma ubuchwepheshe ezifana minima jumping , Ukuphumelela ama-barriers energy ngesikhathi sokuphumelela kuyadingeka. 1 15 16 17 18 19 20 Okokuqala, isixazululo se-adsorbate ne-surface isixazulwe, bese izinhlelo eziningi zihlanganiswa ngokuvamile kanye / noma ngempumelelo. Kule isixazululo, izinhlelo ze-DFT zihlanganiswa kanye nezinhlelo zihlanganiswa ngokuvumelana nezinzuzo zofilimu ezigcwele zokubonisa amandla ze-adsorption (isib. desorption, dissociation, surface mismatch). I-energy engaphansi ku-all configurations iyatholakala njenge-adsorption energy. amandla Adsorption (Δ I-adsorbent-surface energy (I-adsorbent-surface energy) iyahambisana ne-adsorbent energy ( sys) ukunciphisa amandla yokuhlanza isifinyezo (isib. i-plate) ( ibhodi) kanye nokushisa kwe-gas phase adsorbates noma izinhlobo yokufaka ( gas), njengezifundowe ngu Chanussot et al. futhi zihlanganiswe ku- (Ukuza) , . E E E E Ukusekela ulwazi Yini) 2 4 I-adsorbate ebusweni-ukudluliselwa kufuneka ukunakekelwa izakhiwo ezithile izakhiwo ezivamile ukuze amandla lokusabela kwabo kubhalwe kahle nangokuthi. Omunye isibonelo yokudluliselwa kuyinto i-adsorbate akufanele ukudluliselwa, i.e., ukufuduka, kusuka ku-surface ku-ultimate ukwakhiwa (i-Fig. Ngaphezu kwalokho, uma i-adsorbate ine-athomu eminingi, kufanele akufanele ukuxazululwa noma ukuxazululwa ku-adsorbates eminingi ngoba akufanele ukunemba amandla ye-adsorbate ye-molecule ye-interest , Ngaphezu kwalokho, uma i-adsorbate inikeza izinguquko ezithakazelisayo emzimbeni kuqhathaniswa ne-pure surface, Ukubuyekezwa ukuthi uma isakhiwo esilinganisile ukuphazamiseka omunye lezi zimpendulo, kungekho kuqinisekisa ukuthi ukucindezeleka kwenziwe ngempumelelo; lezi iziphumo zihlanganisa kodwa zihlanganisa ama-energy adsorption ezingenalutho noma ezingenalutho njengezinguquko. 1 19 21 E Ukucaciswa kwe-global optimum adsorbate-surface configuration kubhalwe ngokuvamile ku-expert intuition noma eminyakeni eminyakeni eyi-heuristic approaches. I-intuition ne-trial and error ingasetshenziselwa izinhlelo ezimbalwa ezinzima kodwa akufanele ekubunjweni kwezinhlelo eziningi. I-heuristics eyenziwe ngokuvamile isekelwe ku-surface symmetry , . Lezi zindlela zisetshenziswe ngempumelelo emifundo emva-descriptor-based , , , , , Ngaphezu kwalokho, i-graph-based method has been used to identify unique adsorbate-surface configurations Nangona kunjalo, njengoba ukucindezeleka kwama-surfaces ne-adsorbates, ukucindezeleka ukufumana isakhiwo se-adsorbate-surface enhle kakhulu. Lokhu kubalulekile ikakhulukazi lapho i-adsorbate iyahambelana ne-adsorbate, enezinhlangano eziningana nezinhlangano zayo, ngakho-ke kukhona izinga eziningi zokusebenza ze-liberation emkhakheni. 22 23 9 10 24 25 26 27 28 Nangona i-DFT inikeza umthamo wokubala amandla nama-atomic kanye nama-energies, kuyinto enhle ekhompyutha. ( 3) nge inombolo ye-electron. Ukubuyekeza isakhiwo esisodwa ye-adsorbate-ulwandle nge-DFT ephelele ukunciphisa kungabangela ~ 24 amahora ukucubungula , Njengoba izakhiwo eziningi zihlanganisa ngokuvamile ukuze uthole amandla adsorption, zonke izibalo DFT zihlanganisa kungabangela izinsuku noma ngisho izinsuku. I-hypothetically, uma umuntu wahlanganisa i-brutto-force 100,000 izinto kusuka ku-Materials Project database for CO2 Reduction Reaction (CO2RR) usebenzisa 5 adsorbate descriptors, ~90 surface/material, futhi ~100 sites/surface, ngeke kufuneka ~4.5 billion CPU-day yekhompyutha, ingozi engatholakalayo ngisho supercomputers ezinkulu ehlabathini. Ukuze ukunciphisa kakhulu ikhompyutha esifunekayo, isixazululo esithakazelisayo kuyinto ukukhuthaza search ye-lowest energy adsorbate-surface configurations with machine-learned potential. O N 2 29 30 Ngaphezu kwalokho, amathuba ye-machine learning (ML) yokubuyisa amandla nama-atomic kanye nama-energy zibonisa ukuthuthukiswa okuphakeme kulinganiselwe kwama-benchmarks ezivamile ngenkathi i-orders ye-magnitude engaphansi kwe-DFT , , , , , , Nangona ukucaciswa kwe-ML ku-Open Catalyst 2020 Dataset (OC20) ye-dataset eningi futhi eningi kubaluleke ku-0.3 eV for energy estimation, ukucaciswa kwe-0.1 eV kunokufuneka ukucaciswa okucacileyo. Ukusabela ukuthi indlela ye-hybrid eyenza i-DFT ne-ML potentials kunokufinyelela ukucaciswa okuphakeme ngenkathi ukugcina ukusebenza. 2 31 32 33 34 35 36 37 Ukubuyekeza ukusebenza kwezindlela ezintsha yokufaka amasethi amancane-energy adsorbate-surface kuyinto ezinzima ngaphandle idatha yokuqinisekisa standardized. Kuyinto ezivamile ukuthi izindlela ezintsha zihlanganisa eminingi izinhlelo, okuvimbela generalization ukuhlola , , , , . Ngaphandle kwe-OC20 (1 ukunakekelwa ) "energies adsorption", kungekho isampula isizukulwane ezininzi ngamunye adsorbate-ulwandle isixazululo okufakiwe ukuthi isizukulwane eyodwa eyenziwe akuyona engathengi global. Lokhu kwenza OC20 a dataset engathengi yokufinyelela isizukulwane esincane ukuxhumana amandla Ukuze uthole ukwelashwa lokhu, sinikeza i-Open Catalyst 2020-Dense Dataset (OC20-Dense). I-OC20-Dense ibandakanya izixazululo ezimbili — isixazululo se-validation kanye ne-test set. Isixazululo se-validation isetshenziselwa ukuthuthukiswa; futhi isixazululo se-test for reporting performance. Yonke isixazululo iqukethe izixazululo ezingu-1000 ezizodwa ze-adsorbate-surface kusuka ku-validation ne-test sets ye-OC20 dataset. Akukho idatha kusuka ku-OC20-Dense isetshenziselwa ukuqeqeshwa. Ukuhlola ukuxazululwa kwe-approximation yethu, sinikeza izixazululo ezingu-250 kusuka ku-in-domain ( 19 28 38 39 40 O M 2 I-Hybrid approach to estimating adsorption energies takes advantage of the strengths of both ML potentials and DFT. Thina sampling a large number of potential adsorbate configurations using both heuristic and random strategies and performing relaxations using ML potentials. I- *k-* energy relaxed engcono kunokuthuthukiswa ngokusebenzisa i-single-point DFT calculations noma nge-full DFT relaxations. Ngokusebenzisa le ndlela, i-compromise efanele kungenziwa phakathi ukucacisa nokusetshenziswa. Ukucwaninga okuphakeme ku-energy adsorbate-surface configuration nge-improvement ye-initial structure generation kanye ne-global optimization strategies , , , , , , U-Peterson kwenziwa i-minima jumping method futhi yakhelwe indlela yobuchwepheshe yehlabathi enikeza idatha ye-adsorbate usebenzisa i-minima jumping eqinile. Nokho, indlela yakhelwe ngokuphelele ku-DFT yokusebenza ukubuyekeza, okwenza kube lula ngokucwaninga. Ngaphezu kwalokho, uJung et al. Ukusetshenziselwa ukufundisa ukusebenza lapho inqubo ye-Gaussian isetshenziselwa ukulungiselela i-simulations ye-minimal jumping ezinzima. Izakhiwo ezakhiwe nge-simulations zayo zibonwa yi-DFT futhi zithunyelwe ngempumelelo ku-training set kuze kube nokuphumelela kwe-model convergence. I-model e-trained isetshenziselwa ngempumelelo ze-minimum jumping ezinzima, i-subset ifakwe nge-DFT, futhi amandla lokugqibela le-adsorption iyatholakala. Sinikeze ukuthi izinhlelo zokuqala zokusetshenziswa kwe-machine learning yokushesha le nkqubo zihlanganisa amamodeli ezisebenzayo ngamunye ye-adsorbate / i-catalyst, okuvimbele ukusetshenziswa okwengeziwe , I-one possibility yokwandisa kakhulu i-versatility ye-methods ngokuvamile ukunciphisa i-cost ye-human kanye ne-computational kuyinto ukusetshenziswa kwe-generalizable machine learning potential yokushintshwa kwe-low-energy adsorbate-surface configurations. 19 21 28 38 39 40 41 19 21 42 43 I-Contribution of this work kuyinto ezintathu: Thumela i-AdsorbML algorithm ukucacisa amandla ye-adsorption ngaphansi kwe-spectrum ye-precision-efficiency trade-offs. Thuthukisa i-Open Catalyst 2020-Dense Dataset (OC20-Dense) ukuze benchmark umsebenzi yokufunda amandla adsorption. Thina benchmark literature Graph Neural Network (GNN) amamodeli on OC20-Dense usebenzisa i-AdsorbML algorithm eyakhiwe; ukucacisa amamodeli amabili ezinhle ezisebenzayo izicelo zokusebenza zokusikrini. imiphumela OC20-Dense Ukuhlolwa Ukuhlola izindlela zokucwaninga ama-adsorption energy, sinikeza i-Open Catalyst 2020-Dense Dataset (OC20-Dense) enikezela ngokugcwele amandla lokucubungula kwegama lokucubungula ngokuvamile ngokuvamile ukuhlola izinhlelo ezingu-adsorbate eyodwa-systems surface. Yonke OC20-Dense isixazululo iqukethe ~1000 i-adsorbate-source eyodwa ihlanganisa 74 i-adsorbates, 800+ izakhiwo ze-inorganic bulk crystal, futhi i-total of 80,000 + i-heuristically ne-random-generated configurations. A ukubuyekezwa kwezinye izixazululo zitholakala ku-Table . I-dataset eyadingeka ~4 million CPU-hrs yekhompyutha ekupheleni. Ukubuyekezwa okwengeziwe ku-OC20-Dense kungenziwa ku-"Methods". 1 Thola imiphumela eminingi ye-GNN eyenziwe ngempumelelo ku-OC20 ukuze ukulawula ukusebenza kwamodeli eyenziwe ku-OC20-Dense. Lezi zihlanganisa i-SchNet Ngena ngemvume + , Ngena ngemvume Imininingwane , GemNet-OC-MD , GemNet-OC-MD-Large , SCN-MD-Large , futhi eSCN-MD-Large where MD corresponds to training on OC20 and its accompanying ab initio Molecular Dynamics (MD) dataset. Amamodeli asebenziselwe njenge-part of this work; amamodeli asebenziselwe asebenziselwe ngqo ukususwa kwezinto ezidlulileyo ezivumelekile futhi kungatholakala ku Of the models, (e)SCN-MD-Large futhi GemNet-OC-MD-Large zihlanganisa ngempumelelo phakathi kwe-OC20 kanye ne-Open Catalyst 2022 Dataset (OC22). Ukuhlola okungenani lezi zendwendwe zihlanganisa i-OC20-Dense kuyoba kubalulekile ukunikezela ukuthi izivakashi ze-OC20 zihlanganisa izivakashi ezingavamile ezilandelayo ezifana ne-one eyenziwa lapha. 31 32 33 44 34 34 34 35 45 https://github.com/Open-Catalyst-Project/ocp/blob/main/MODELS.md Ideally, the ground truth for OC20-Dense would be the minimum relaxed energy over all possible configurations for each adsorbate-surface system. Since the number of possible configurations is combinatorial, the community has developed heuristic approaches to adsorbate placement on a catalyst surface , Uma ukulawula kuphela izakhiwo ze-heuristic, sincoma lokhu njenge-DFT-Heuristic-Only (DFT-Heur). Ukongeza indawo yokufaka, sinamathela izindawo ezivamile emaphaketheni ngempumelelo nge-adsorbat esekelwe kumazwe ngamunye ngamazwe nge-rotating ngempumelelo ngokulandelana axis futhi lula wobble emhlabeni Waze isixazululo. Uma ukuguqulwa phakathi kwe-heuristic kanye ne-random configurations, sincoma lokhu njenge-DFT-Heuristic-Only (DFT-Heur). Nakuba i-computer-more expensive, le benchmark inikeza ukubuyekeza okuhlobene kwe-configurations futhi ukuguqulwa okuhlobene kwe-adsorption energy kunokusebenzisa kuphela i-heuristic configurations, isisekelo esiyinhlanganisela esebenzayo kumphakathi. Izinzuzo ezingaphezu kuka-benchmarks ezimbili zingatholakala ku-"Methods". 22 23 z x y ML ukunakekelwa Thina ukubuyekeza ukuthi izinga le-ML izibuyekezo zingathola amandla le-adsorption ku-threshold of the DFT energy minimum, noma ngaphansi. Nakuba i-ML ephelele ye-DFT iyakwazi kuphela ukuhlangabezana ne-DFT, ama-errors ezincinane eziningana nezinhlangano ze-optimizer zihlanganisa i-noise ku-relaxations futhi kuholele ku-configurations engaphansi. Ukuze wonke imodeli, ukuguqulwa kwenziwa ku-set esifanayo ye-adsorbate configurations. I-configurations yokuqala ifakwe ngokusekelwe izindlela ze-heuristic ezisetyenzisiwe ezisetyenziswa emakheni. , Ukusebenza kwe-ML-driven relaxations ku-all initial configurations; izinhlelo ezingatholakala ukulayishwa kwe-adsorption energy ngenxa ye-constrictions zezimboni zithunyelwe, kuhlanganise i-dissociation, i-desorption, ne-surface mismatch. Ukuxhumana okuhlobene kwe-constrictions ye-relaxation ingatholakala ku-"Methods". 46 22 23 Ukubuyekeza ukusebenza, sincoma ukubuyekeza amandla adsorption ku-tolerance engaphansi kwe-acceptable (0.1 eV kule umsebenzi). , , ) noma engaphansi kwe-DFT adsorption energy e-OC20-Dense. Qaphela ukuthi ama-general truth adsorption energies e-OC20-Dense kukhona isixazululo ephakeme, njengoba kungenzeka ukuthi isixazululo isixazululo isixazululo isixazululo engaphansi kungenziwa. Lapho ukulawula ama-energy adsorption eyenziwe nge-ML, imiphumela kufuneka ifakwe ngokusebenzisa i-single-point DFT calculation, njengoba i-metric ye-evaluation ngaphandle kwesixazululo engaphansi kungenziwa kalula ngokuvamile ama-energy amancane (bheka Ukubuyekeza ngokuphathelene i-ML, sincoma amandla lokuphuma kwe-ML efanele uma kufinyelela ngaphakathi kwe-0.1 eV ye-DFT adsorption energy noma engaphansi, futhi ukucaciswa kwe-DFT ye-single-point ye-ML isakhiwo esithathwe ku-0.1 eV ye-ML energy esithathwe. Lokhu kuncike ukuthi i-ML yokubuyekeza akuyona kuphela amandla okunciphisa okunciphisa kodwa akuyona ngokuphathelene. Iziphumo zithunyelwe ku-Table , lapho amamodeli top OC20 kuhlanganise eSCN-MD-Large kanye GemNet-OC-MD-Large zithunyelwe izinga lokuphumula ze-56,52% kanye ne-48.03%, ngokuvamile. I-Energy MAE phakathi kwe-ML ne-DFT adsorption energies iyathunyelwe futhi ku-Table , uqhagamshelana kakuhle nezinga lokuphumula kanye ne-OC20 Ukuhlobisa 2 37 46 Yini) 2 2 Waze Nangona izimo zokusebenza zokusebenza zangaphambili , izinga okusezingeni eliphezulu zokusebenza zokusebenza ngokuvamile kubalulekile. Ngokuvamile ephakeme lapho izidakamizwa ezinhle zithembisa ukuhlolwa okusheshayo noma ngisho ukuhlanganiswa kwe-experimental, isivinini sokuphumelela e-50% kungabangela ukuchithwa isikhathi esikhulu kanye nezinsizakalo ukuhlola i-positives ezihambayo. Njengoba ukuthuthukiswa kwama-modeli iyaqhubeka ukwandisa i-metric, le ncwadi kusiza izindlela ze-ML+DFT ezihambayo ukuze ukwandisa izinga lokuphumelela ngokushesha kwezibalo ezengeziwe. 37 Ukulungiselela Ukuhlobisa Thina uvumela I-algorithm ukusetshenziswa kwe-ML yokushesha inqubo yokufaka kwe-adsorbate (I-Fig. ). Kule imodeli, sincoma izindlela ezimbili ezihlanganisa i-ML ezilandelayo izibalo ze-DFT ukucacisa amandla le-adsorption. Sincoma ukuthi lokhu izindlela kuyinto jikelele futhi ingasetshenziswa ne-algorithm ye-initial configuration. Ukuhlobisa 2 Fig. 2: The algorithm. AdsorbML Ukuhlobisa ikhaya Ukuqhathaniswa kokuqala kubhalwe ngokusebenzisa izindlela ze-heuristic ne-random. Ukuqhathaniswa kwe-ML ikhiqizwa ku-GPU kanye nesigaba se-lowest to highest energy. izinhlelo zithunyelwe ku-DFT ngenxa yokubuyekezwa kwe-single-point (SP) noma ukuguqulwa kwe-full (RX) kusuka ku-ML-relaxed isakhiwo. I-system that does not satisfy constraints is filtered at every stage a relaxation is performed. The minimum is taken across all DFT outputs for the final adsorption energy. k Kuhlolwe kulabo, ingxenye yokuqala kuyinto ukukhiqiza ukujabulela kwe-ML. Nokho, ngaphandle kokufaka i-minimum ye-energy e-ML, sinikeza izinhlelo ngokuhambisana ne-minimum kuya ku-highest energy. izinjini ezincinane ze-energy zihlanganiswa futhi (1) izibalo ze-DFT ye-single-point zitholakala ku-structures efanayo (ML+SP) noma (2) izincazelo ze-DFT zitholakala ku-structures ML-relaxed (ML+RX). I-strategy yokuqala kusetshenziselwa ukufinyelela ukucaciswa kwamanzi we-energy ye-ML-predicted relaxed structure, lapho i-second kusetshenziselwa i-ML njenge-pre-optimizer nge-DFT ekupheleni ukucaciswa. Ngokuvumelana ne-ML-predicted relaxed structure izinhlelo zokusebenza okuphansi, sinikeza imodeli nge izinzuzo ukufinyelela amandla adsorption ngokuphathelene. Njengoba thina ukwandisa , ukucubungula kwe-DFT ngaphezulu, kodwa kuqhathaniswa ne-DFT ephelele, sincoma ukunciphisa kakhulu. I-adsorption energy ye-system eyodwa ikhiqizwa ngokusebenzisa i-minimum ye-best DFT izilinganiso ezilandelayo. k k k k k Kwi-strategies ezimbili, i-energies ye-ML isetshenziselwa kuphela ukuhlaziywa kwezinhlobonhlobo, kanye nokuqinisekisa lokugqibela kwamanzi kusuka ku-DFT yekhompyutha. Nakuba ku-computational, kuyoba kuhle ukuvuseleleka ngokuphelele ku-ML, ukusetshenziswa kwe-DFT ikakhulukazi ukwandisa ukucacisa futhi inikeza isinyathelo sokuqinisekisa ukuze sinikeze iningi ukunakekelwa kwama-adsorption energy predictions ethu. Ukucubungula Umthamo lethu kuyinto ukufumana amandla adsorption efana noma engcono kunazo ukusetshenziswa kwe-DFT kuphela ku-OC20-Dense. I-metric etholakalayo ukucubungula lokhu umsebenzi yi-success rate, okuyinto i-percent of OC20-Dense systems where our ML+DFT adsorption energy is within 0.1 eV or lower than the DFT adsorption energy. A validation of the ML energy is not included in these experiments since all final adsorption energies will come from at least a single DFT call, ensinguquko zonke izinga zihlanganisa. Enye metric we track is the speedup compared to the DFT-Heur+Rand baseline. Speedup is assessed as the ratio of DFT electronic steps used by DFT-HeurRand to the propos Ukubuyekeza isisekelo esivamile se-DFT-Heur esebenzisa kuphela ukucubungula kwe-DFT, isivinini se-success yi-87.76% iyatholakala ngesivinini se-1.81×. Yini Ukubuyekezwa kwe-SP Iziphumo zokusebenzisa izibuyekezo single-point phezu izimo ML-ukuguqulwa zihlanganiswe ku Fig. . i-eSCN-MD-Large kanye ne-GemNet-OC-MD-Large zithunyelwe izinga lokuphumula we-86% + = 5 nge-eSCN-MD-Large enikezela zonke amamodeli nge-success rate ye-88.27%, okungenani engcono kunesibuyekezo se-DFT-Heur. Amamodeli abaningi kuhlanganise i-SchNet ne-DimeNet++ zihlanganisa kakhulu nge-success metrics okungenani ne-3.13% kanye ne-7.99%, ngokuvamile; okuvumela ukuba izakhiwo ezihlangene ezihambayo ezihambile kakhulu. I-speedups iyatholakala kakhulu kuzo zonke amamodeli, ezingu-1400× kuya ku-1500× for = 5, izindinganiso ze-magnitude eziningana ne-DFT-Heur baseline. Ngokuphathelene, i-eSCN-MD-Large ne-GemNet-OC-MD-Large zihlanganisa ukuhlaziywa kwama-1384× kanye ne-1388×, ngokulandelana. Uma isivinini kuyinto ebalulekile kakhulu, ukuhlaziywa kwama-6817× kungatholakala nge = 1 ngokuvamile ukugcina izinga lokuphumula we-82% ye-eSCN-MD-Large. Ngokuphumelela okwengeziwe, = 3, izinga lokuphumula ze-87.36% kanye ne-84.43% zitholakala ku-eSCN-MD-Large kanye ne-GemNet-OC-MD-Large ngenkathi zihlanganisa izinga lokuphumula ze-2296× kanye ne-2299×, ngokuvamile. , ama-minimum-energy-binding sites yezinhlelo eziningana zihlanganiswa njengokuthi zihlanganiswa nge-ML+SP phakathi kwama-models ezahlukene. 3 k k k k 4 Ukuze wonke imodeli, isivinini se-DFT ne-success rate esebenzayo ifakwe ku-ML+RX ne-ML+SP phakathi kwezinye izinzuzo ezinhle. . I-system ibhekwa ngempumelelo uma amandla wokubacindezelwa okungenani ku-0.1 eV kusuka ku-DFT minimum, noma ngaphansi. Zonke izinga lokuphumelela kanye nezinga lokuphumelela zihlanganisa ku-Random+Heuristic DFT. I-DFT ye-heuristic ibonisa njenge-common community baseline. I-top-right corner ibonisa indawo efanelekayo-maximizing speedup kanye nezinga lokuphumelela. I-point eyenziwe ku-teal ibheka ne-optional ephilayo ebizwa ku-abstract-a 87.36% izinga lokuphumelela kanye ne-2290x speedup. Inani elifanayo ye-OC20-Dense validation set ingatholakala ku- . k Yini I-ML-relaxed structures has energy calculated with a DFT single-point, ML+SP. A variety of systems are shown including those where ML finds lower, higher, and comparable adsorption energies to DFT. Qaphela ukuthi ezinye izakhiwo ku-system third and fourth are symmetrically equivalent, futhi ukuthi i-SchNet induces a large surface reconstruction in the third system resulting in the extremely large DFT energy (10.31 eV). I-ML+RX ye-RX While single-point evaluations offer a fast evaluation of ML structures, performance is heavily reliant on the accuracy of the predicted relaxed structure. This is particularly apparent when evaluating the max per-atom force norm of ML-relaxed structures with DFT. SchNet and DimeNet++ have on average a max force, , of 2.00 eV/Å and 1.21 eV/Å, respectively, further supporting the challenge these models face in obtaining valid relaxed structures. On the other hand, models like GemNet-OC-MD-Large and eSCN-MD-Large have an average of 0.21 eV/Å futhi 0.15 eV/Å, ngokulinganayo. Nangona lezi zibonelelo kakhulu ku-valid relaxed structures (isib. ≤ 0.05 eV/Å), these results suggest that there is still room for further optimization. Results on DFT relaxations from ML-relaxed states are plotted in Fig. . eSCN-MD-Large and GemNet-OC-MD-Large outperform all models at all izinga, nge 90.60% kanye 91.61% izinga lokuphumula = 5, ngokuvamile. Ngokusho izindleko ezengeziwe ze-DFT ezihlaziywa nge-refining relaxations, ama-speedups zihlukile ngokushesha. = 5, we see speedups of 215× and 172× for eSCN-MD-Large and GemNet-OC-MD-Large, respectively. Both SchNet and DimeNet++ see much smaller speedups at 42× and 55×, respectively. The much smaller speedups associated with SchNet and DimeNet++ suggest that a larger number of DFT steps is necessary to relax the previously unfavorable configurations generated by the models. Conversely, eSCN-MD-Large’s much larger speedup can be attributed to the near relaxed states (average ~0.15 eV/Å) it achieves in its predictions. With = 1, ama-speedups ka-1064× zitholakala nangokuthumela isivinini se-84.13% ye-eSCN-MD-Large. Ngezinye isivumelwano esebenzayo, = 3, success rates of 89.28% and 89.59% are attainable for eSCN-MD-Large and GemNet-OC-MD-Large while maintaining speedups of 356× and 288×, respectively. Ukubuyekezwa fmax fmax 3 k k k fmax k k Iziphumo zihlanganisa i-spectrum ye-accuracy kanye ne-efficiency-compromise eyenziwa lapho ukhethe isinyathelo. Ukuze amamodeli bethu, iziphumo zethu ze-ML+SP zihlanganisa cishe amaxesha amahora angu-8 kunama-ML+RX kanye nokunciphisa kuphela izinga lokuphumula kwama-success rates (i-3-4%), okuvumela ukuba i-composite iyatholakala. Lezi zihlanganisa kunzima kakhulu kumodeli ezingaphumile. In Table , sici ukuguqulwa kwezibuyekezo ezinhle kakhulu, ngama-parity, noma kakhulu engaphansi kwe-found truth, lapho kakhulu engcono / engaphansi kubaluleke ku-lower/higher than 0.1 eV of the DFT adsorption energy. Phakathi nezinhlangano ezimbini, sincoma ukuthi amamodeli amancane akufanele ukufumana ama-minima ezinhle kakhulu. = 5 ML+RX, i-eSCN-MD-Large ibonise i-9.10% yama-system nge-minima ezinzima kakhulu, kuqhathaniswa ne-DimeNet++ ibonise i-15.57%. Ngaphezu kwalokho, uma i-eSCN-MD-Large ibonise amamodeli e-ML+SP, ibonise ukuguqulwa okungenani ne-ML+RX; isiphumo se-model ebonakalayo ku-minima esikhulu yendawo ukuthi ukunciphisa kwe-DFT esilandelayo kunezinzima okungenani. Lokhu kwakhona ibonise ukuthi uhlobo elinye le-noise emodeli ingasiza ekufuna i-minima ezingcono. I-set ephelele ye-tabulated results ye-ML+SP kanye ne-ML+RX ama-exper I-OC20-Dense isikhungo se-validation. 3 k Yini Distribution splits Ngaphezu kwalokho, sincoma izinga lokuphumula phakathi kwezinhlayiyana ezahlukile ze-dataset. OC20-Dense isampula ngokulinganiselwe ezine izinhlayiyana ze-OC20 – ID, OOD-Adsorbate, OOD-Catalyst, futhi OOD-Both. Ngezinye amamodeli ethu ezinhle, sincoma ukuthi ukusebenza lihlale ngokulinganiselwe phakathi kwezinhlayiyana ezahlukile (i-Fig. ). Lokhu kubonisa ukuthi ngezicelo kuhlanganise ama-adsorbates noma izindwangu ezingatholakala ku-OC20, i-AdsorbML inikeza imiphumela enhle futhi enhle. Nakuba siphinde iziphumo zihlanganisa nge-OC20 lapho i-ID iyahambisana ne-OOD, lokhu akuyona okungagunyaziwe lapha. eSCN-MD-Large, i-ML+SP ku = 5, ivumela i-86.00% ku-ID ngenkathi i-88.35% i-success rate ku-OOD-Both, nge imibuzo efanayo ku-ML+RX. Sinikeza lokhu kokuhlanganisa ubukhulu obukhulu obukhulu be-sample ngalinye (250). I-set ephelele ye-results ingatholakala ku . 5 k SI Fig. 5: ML+SP success rate at = 5 across the different subsplits of the OC20-Dense test set and several baseline models. k k Izimodeli eziphezulu zokusebenza zibonisa izinguquko ezingenalutho phakathi kwezinhlayiyana ezahlukile ezahlukile, okuvumela ukusebenza okuhle sokuqinisekisa ku-out-of-domain adsorbates kanye ne-catalysts ezingatholakala ku-OC20 training dataset. Ukuhlolwa kwe-Configuration Ngezinye iziphumo eziyinhloko, sincoma ukusebenza kokusebenzisa kuphela izakhiwo ze-heuristic noma kuphela ze-ML ngempumelelo ku-OC20-Dense validation set. Iziphumo zithunyelwe ku-SCN-MD-Large, ngenxa ye-ML+SP strategy. Ku = 5, lapho kusetshenziselwa ukuguqulwa kwe-random kuphela, ukunciphisa imiphumela kusuka ku-87.77% kuya ku-82.94%. Ngaphezu kwalokho, uma kusetshenziselwa kuphela ukuguqulwa kwe-heuristic, ukunciphisa imiphumela kakhulu ku-62.18%. Lokhu kubonisa ukuthi ukuguqulwa kwe-random kungase abe umphumela emikhulu. Iziphumo ezengeziwe zingatholakala ku . k Yini Ukuhlobisa Thina ukubonisa lokhu umsebenzi njengezinyathelo ebalulekile kodwa yokuqala ukunciphisa i-computational cost of DFT for not only catalysis applications, but computational chemistry more broadly. inikeza i-spectrum ye-accuracy kanye ne-efficiency trade-offs enokukhetha ngokuvumelana ne-application ne-computational resources ezine. Ngokwesibonelo, uma sincoma ukucubungula inani elikhulu le-catalysts ye-CO2 reduction reaction, ngokuvumelana ne-budget ye-computing ebonakalayo, singathola i-ML+SP ku = 2 for a 85% success rate while screening ~3400× more materials than would have been possible with DFT alone. On the other hand, if depth of study is more important, ML+RX is a good alternative as the structures are fully optimized with DFT and the computational speedup comes from reducing the total number of relaxation steps required. In this scenario, the ML potential serves as an efficient pre-optimization step. Even though ML models comprise a small portion of the overall compute (see Ukuze uthole okwengeziwe, sincoma ukuthi izidingo zayo zithunyelwe ngakumbi njengoba izinzuzo ezininzi zithunyelwe ku-efficiency ye-inference emkhakheni. Ukuhlobisa k Yini One observation that merits additional studies is that ML models found much better minima between 5 and 15% of the time, depending on the efficiency trade-offs (Table ). Uma amamodeli yethu ye-ML iyona ephelele, akukwazi ukuthola izibonelo ezinzima ezinzima ezinzima; Nokho, i-noise ephilayo emithonjeni ye-force predictions evumela amamodeli e-ML ukuhlola izindawo ezingafuneki ze-potencial energy surface. Ukuhlola ukuthi kungeningi i-noise ephilayo ne-explicit , impact ML relaxations and downstream tasks such as success rate is an important area of future research. 3 46 47 Another natural extension to this work is focusing on alternative methods of global optimization and initial configuration generation. Here, we focused on accelerating brute-force approaches to finding the global minimum by enumerating initial adsorbate-surface configurations. However, there are likely to be much more efficient approaches to global optimization such as minima hopping I-Optimization ye-Limited , I-Bayesian Optimization, noma indlela eyenziwe ngqo. Kubaluleke ukubonisa ukuthi ngenkathi ukucubungula kwelanga elikhulu kunezindlela ezivamile ze-heuristic, akuyona enhle futhi enhle. Siphinde ukuthi ukwandisa inani lwezinhlangano ezingenalutho ngaphezu kwama-sampled kunezinguquko, njengoba ukuguqulwa kwezinga lokuphumelela kusuka ku-heuristic + 80% ye-DFT ezingenalutho kuya ku-heuristic + 100% ye-DFT ezingenalutho kuphela ku-1.6% (bheka i-DFT ye-80% ye-heuristic). Ukuze uthole okwengeziwe). Uma ukucubungula izinguquko ezininzi ze-ML kusebenza ngokuvamile, ukubuyekeza kanjani ukulawula izakhiwo ezimbonini kungabangela ukunambitheka kanye nokusetshenziswa. Sihlola lokhu ngokushesha ku , lapho ukunciphisa izinhlelo ezinama-energy eziningana ne-ML ziphumela izinzuzo ezincinane. 20 19 21 SI Yini Nangona amamodeli asebenzayo ezifana ne-GemNet-OC ne-eSCN-MD-Large zibonisa izinga lokuphumula kakhulu ku-OC20-Dense, izinga lokuphumula kwe-ML ngaphandle kwezinye i-DFT ezilandelayo akuyona engcono kakhulu kwezicelo zokusebenza (I-Table ). In order for future modeling work to address this challenge, there are a number of observations worth highlighting. First, there is a positive correlation between success rate on OC20-Dense and both the Ukuhlobisa okuhlobisa ( ) OC20 umsebenzi. Ngakho, esisekelwe ukunciphisa and I-metric ingasetshenziswa njenge-proxy lapho ibhekwa amamodeli ku-OC20. Enye ibheka ebalulekile ekuthuthukiseni imodeli iyona i-validation set ye-OC20-Dense kuyinto i-subset ye-OC20 validation set; Ngenxa yalokho, idatha ye-OC20 validation akufanele isetshenziswe ekuthuthukiseni kwe-OC20-Dense. Ekugcineni, kuncike kakhulu ukuthi imiphumela eyenziwa ku-OC20-Dense validation set iyatholakala ngokusebenzisa ukucubungula kwe-DFT ye-single-point ngoba i-success rate metric ingasetshenziswe ngokubonisa kuphela amandla amancane. Lokhu kungenziwa nge ~1000 ukucubungula we-single-point. Ngaphandle kokuqalisa kwe-OC20-Dense test 2 S2EF Waze IS2RE Waze Ukuthuthukiswa okusheshayo ku-datasets kanye ne-machine learning ye-chemistry ivumela amamodeli ukufinyelela kwindawo lapho zithuthukisa futhi ukwandisa izibalo ze-DFT. Imiphumela yethu ibonise ukuthi amamodeli e-ML e-state-of-the-art asebenzayo kuphela ukulungisa izibalo ze-DFT ze-catalysis kodwa akwazi ukuhlaziywa okungcono kwezakhiwo ezidingekayo ze-global, njenge-adsorption energies. Nakuba amamodeli ezisetshenziselwa lokhu umsebenzi zihlukile kakhulu kumathalisi we-adsorbate-surface, izinhlelo ze-fine-tuning angasiza ukufinyelela izicelo kwezinye izinhlelo zamakhemikhali kuhlanganise izinhlelo ze-metal-organic ne-zeolites . Similarly, the models used in this work were trained on a consistent level of DFT theory (revised Perdew–Burke–Ernzerhof, no spin polarization), generalizing to other functionals and levels of theory could also be enabled with fine-tuning or other training strategies. Given the timeline of ML model development, these results would not have been possible even a couple of years ago. We anticipate this work will accelerate the large-scale exploration of complex adsorbate-surface configurations for a broad range of chemistries and applications. Generalizing these results to more diverse materials and molecules without reliance on DFT is a significant community challenge moving forward. 29 Methods I-Open Catalyst 2020-Dense Dataset (OC20-Dense) Ukubuyekezwa kwe-adsorption energy estimates kudinga i-ground truth dataset enikezela ngokugqithiselwe i-set ye-potencial adsorption configurations. Nakuba i-OC20 i-calculated adsorption energies (1 ukunakekelwa ) izinhlelo, ama-energies akufanele akufanayo ne-minimum ye-adsorbate-surface combination. Ngokuvamile, ngenxa ye-catalyst surface eyodwa, i-OC20 ibonise zonke izindawo ezivamile ze-adsorption kodwa kuphela ibonise i-adsorbate ebonakalayo kwi-site eyenziwe ngempumelelo ku-configuration eyodwa. Imisebenzi eyenziwa yi-OC20 ivumela ukuthuthukiswa kwamakhasimende amancane kakhulu eyenziwa ku-catalysis , , , , Nokho, izicelo ezifana Ukubambisana kakuhle, akuyona engcono lapho ukulawula ukusebenza njengoba amamodeli zithunyelwe lapho ukhangela ubuningi obuningi, okungenani-energy – imiphumela enhle kakhulu. Njengomphumela ekwelapheni umsebenzi OC20, sinikeza OC20-Dense ukuhlola ukusebenza amamodeli ukuze uthole amandla adsorption. O M 34 35 47 48 49 Waze I-OC20-Dense yakhelwe ukuhlangabezana ngokushesha amandla wokubambisana kwe-adsorption ye-adsorbate ye-surface. Ukuze uthole lokhu, isampula enhle ye-adsorption yokuqala kuyadingeka. I-OC20-Dense iqukethe izixazululo ezimbini — i-validation kanye ne-test set. Ukuze yonke isixazululo, i-1000 i-adsorbate eyodwa ye-surface e-validation / test set ye-OC20 ngokubhekiselele isampula. Isampula eyodwa isetshenziselwa ngamunye isixazululo (i-ID, i-OOD-Adsorbate, i-OOD-Catalyst, i-OOD-Both) ukuhlola ukuxhumana kwamamodeli kulokhu. Ukuze yonke isixazululo se Ukubuyekezwa Ngokuvumelana ne-adsorbate kanye ne-surface, i-Pymatgen ibonisa zonke izindawo ezivamile, i-adsorbate isetshenziselwa indawo, kanye ne-rotating e-z-axis eyenziwe ngempumelelo ku-z-axis. and Ukusetyenziswa kwe-axis ku-adsorbate. Nakuba izindlela ze-heuristic zihlanganisa izindlela ezinhle, zihlanganisa indawo yokufinyelela okungagunyaziwe ngaphandle kokubili ukuthi i-energy ephakeme efanelekayo iyatholakala. Ukuze kusetshenziswe lokhu, sinikeze ngempumelelo izindawo ze-M ku-surface bese zihlanganisa i-adsorbate ku-top of the selected site. Kule umsebenzi, i-M = 100 isetshenziswe futhi i-rotating ye-random isetshenziselwa i-adsorbate ngezindlela efanayo. Kwi-strateji ezimbili, sinikeze izakhiwo ezingenalutho — adsorbates asekelwe ku-plate futhi / noma zihlanganiswa kakhulu ku-surface. Izakhiwo ze-DFT z . 23 22 x y SI I-OC20-Dense test set iqukethe i-989 i-adsorbate + i-surface combinations eyenziwe nge-74 i-adsorbates ne-837 i-bulks. Ngemva kokufaka kwama-sampling, i-56,282 i-heuristic ne-49,432 i-configurations ye-random zihlanganisa nge-DFT. Ngokuvamile, kukhona i-56 i-heuristic ne-50 i-random configurations ngamakhasimende (ngokuthi i-M = 100 i-random sites ziye ziye ziye ziye zithunyelwe ngesikhathi sokufilishwa). Ngokuvamile, ~4 million amahora ye-computing ziye zithunyelwa ekukhiqizeni i-dataset. Zonke i-DFT zihlanganiswa ziye zihlanganiswe nge (VASP) , , , . Ukubuyekezwa kwe-DFT settings kanye ne-details kungenziwa ku- . Vienna Ab initio Simulation Package 50 51 52 53 Yini Ukuhlolwa kwama-metric To sufficiently track progress, we propose two primary metrics—success rate and DFT speedup. is the proportion of systems in which a strategy returns energy that is within , noma ngaphansi kwe-DFT adsorption energy. A margin of = 0.1 eV iyatholakala njengoko isixazululo esithakazelisayo ukujabulela inani elincinane le-error ukuze kube lula. , Ukunciphisa le threshold ukunciphisa ukucaciswa kuyinto isinyathelo esithathayo uma amamodeli + amasu zihlanganisa. Nakuba izinga eliphezulu lokuphumula zingatholakala nge-DFT ukucaciswa, thina usebenzisa as a means to evaluate efficiency. Speedup is measured as the ratio of DFT electronic, or self-consistency (SC), steps used by DFT-Heur+Rand and the proposed strategy. Electronic steps are used as we have seen them correlate better with DFT compute time than the number of ionic, or relaxation, steps. DFT calculations that failed or resulted in invalid structures were included in speedup evaluation as they still represent realized costs in screening. We chose not to include compute time in this metric as results are often hardware-dependent and can make comparing results unreliable. ML relaxation costs are excluded from this metric as hardware variance along with CPU+GPU timings make it nontrivial to normalize. While ML timings are typically negligible compared to the DFT calculations, a more detailed analysis of ML timings can be found in the . Metrics are reported against the rigorous ground truth—DFT-Heur+Rand, and compared to a community heuristic practice—DFT-Heur. Formally, metrics are defined in Equs. ( (NgoLwesithathu) Ngathi Success rate σ σ 2 37 DFT speedup Yini 2 3 where kuyinto adsorbate-surface uhlelo, Ukubalwa kwezinhlelo ezizodwa, 1 kuyinto umsebenzi we-indicator, kuyinto isitimela esethelwe, is the number of self-consistency, or electronic steps, and is the minimum energy across all configurations of that particular system. Ukuze amabili amametrics, ephakeme kuyinto engcono. i N NSCsteps Izinzuzo ze-Relax It is possible that some of the adsorbed surface configurations we consider can relax to a state that are necessary to discard in our analysis. For this work, we considered three such scenarios: (1) desorption, (2) dissociation, and (3) significant adsorbed-induced surface changes. Desorption, the adsorbed molecule not binding to the surface, kungenziwa kakhulu ongaphakeme ngoba izinhlelo desorbed zokusebenza ngokuvamile high energy. Nokho, kungenziwa ezisebenzayo ukucacisa lapho omunye izakhiwo zihlanganisa ziye ngokuvamile dissorbed to the surface. Dissociation, ukucaciswa kwe-adsorbed molecule ku-athomu ezahlukene noma ama-molecule, kuyinto problematic ngenxa yokwenza adsorbent energy engaphakanyayo akuyona ngokuvumelana ne-athomu ephezulu . SI Ukufinyelela Data The full open dataset is provided at . https://github.com/Open-Catalyst-Project/AdsorbML Code availability Zonke ikhowudi ezokuxhumana kusebenza ku . https://github.com/Open-Catalyst-Project/AdsorbML Ukuhlobisa Nørskov, J. K., Studt, F., Abild-Pedersen, F. & Bligaard, T. (John Wiley & Sons, 2014). Fundamental Concepts in Heterogeneous Catalysis Chanussot, L. et al. I-Open Catalyst 2020 (oc20) Dataset kanye ne-Community Challenges. I-ACS Catal. 11, i-6059-6072 (2021). Isihloko se-CAS Google Scholar Dumesic, J. A., Huber, G. W. & Boudart, M. Izinzuzo ze-catalysis ye-heterogeneous (I-Wiley Online Library, 2008). Zitnick, C. L. et al. Isikhathi esifundweni se-electrocatalyst ukusetshenziswa kwe-machine learning yokuhweba kwamandla eyenziwe ngamanani. I-preprint ku-https://arxiv.org/abs/2010.09435 (2020). Choudhary, K. et al. Izinzuzo ezidlulile kanye nezicelo ze-deep learning methods ku-material science. NPJ Comput. Mater. 8, 59 (2022). Isihloko se-Google Scholar Wen, T., Zhang, L., Wang, H., Weinan, E. & Srolovitz, D. J. Deep potentials for materials science. , 022601 (2022). Mater. Futures 1 Article Google Scholar Wei, J. et al. Machine Learning in Material Science. InfoMat 1, 338–358 (2019). Isihloko CAS Google Scholar Ulissi, Z. W., Medford, A. J., Bligaard, T. & Nørskov, J. K. Ukusabela ukubaluleka kwenethiwekhi ye-surface reaction ngokusebenzisa ukubaluleka kwezilinganiso ze-machine learning kanye nezilinganiso ze-dft. Nat. Commun. 8, 1-7 (2017). Umbhalo we-Google Scholar U-Tran, K. & Ulissi, Z. W. Ukufundiswa okuqhubekayo phakathi kwe-intermetallics ukuqondisa ukukhululwa kwe-electrocatalysts yokunciphisa i-CO2 ne-h2 evolution. Nat. Catal. 1, 696–703 (2018). Isitolo se-CAS Google Scholar Zhong, M. et al. Ukukhishwa okusheshayo kwe-co2 electrocatalysts usebenzisa isifundo se-machine esebenzayo. Nature 581, 178-183 (2020). Article CAS Google Scholar Liu, X. et al. Ukuphathelela imibuzo ye-electrochemical carbon dioxide reduction rates. Nat. Common 8, 1-7 (2017). Google Scholar Nørskov, J. K. et al. Izinzuzo ze-exchange current for hydrogen evolution. J. Electrochem. Soc. 152, J23 (2005). Wan, X. et al. Ukukhishwa kwe-machine-learning-assisted ye-high-efficient entropy alloy catalysts for the oxygen reduction reaction. Patterns 3, 100553 (2022). Seh, Z. W. et al. Ukuhlanganiswa kwe-theory ne-experiment ku-electrocatalysis: Ukuhlobisa ku-design yezinto. Isayensi 355, eaad4998 (2017). Hohenberg, P. & Kohn, W. Inhomogeneous electron gas. Phys. Rev. 136, B864 (1964). Kohn, W. & Sham, L. J. I-Equations ye-self-consistent kuhlanganise imiphumela ye-exchange ne-correlation. Physical Review 140, A1133 (1965). Sholl, D. S. & Steckel, J. A. Density Functional Theory: A Practical Introduction (John Wiley & Sons, 2022). Teukolsky, S. A., Flannery, B. P., Press, W. & Vetterling, W. Izimpendulo ze-Numerical ku-C. SMR 693, 59-70 (1992). Peterson, A. A. Global optimization of adsorbate-surface structures while preserving molecular identity. , 40–53 (2014). Top. Catal. 57 Goedecker, S. Minima jumping: indlela efanelekayo yokufunda i-global minimum ye-potencial energy surface of complex molecular systems. J. Chem. Phys. 120, 9911–9917 (2004). Jung, H., Sauerland, L., Stocker, S., Reuter, K. & Margraf, J. T. I-machine-learning-driven global optimization ye-surface adsorbate geometries. NPJ Comput. Mater. 9, 114 (2023). Ong, S. P. et al. I-Python Materials Genomics (i-pymatgen): isibuyekezo se-Python enhle, esisodwa se-open-source yokuhlola izinto. I-Comput. Mater. Sci. 68, 314-319 (2013). I-Boes, J. R., Mamun, O., Winther, K. & Bligaard, T. Graph isikhokelo se-theory ku-high-throughput surface adsorption structure generation. J. Phys. Chem. A 123, 2281–2285 (2019). Andersson, M. P. et al. Ukuphakamisa kwe-screening ye-catalysis ye-heterogenic: i-catalysts ye-pareto-optimal ye-methanation. J. Catal. 239, 501-506 (2006). I-Bligaard, T. et al. I-Brønsted-Evans-Polanyi relationship ne-vulcano curve ku-catalysis enhle. J. Catal. 224, 206–217 (2004). Studt, F. et al. Identification of non-precious metal alloy catalysts for selective hydrogenation of acetylene. , 1320–1322 (2008). Science 320 Nilekar, A. U., Sasaki, K., Farberow, C. A., Adzic, R. R. & Mavrikakis, M. I-Mixed-metal Pt monolayer electrocatalysts nge-CO tolerances ezintsha. J. Am. Chem. Soc. 133, 18574–18576 (2011). Deshpande, S., Maxson, T. & Greeley, J. Graph ubuchwepheshe ukucacisa izakhiwo ze-multidentate ne-high-coverage adsorbates for heterogeneous catalysis. NPJ Comput. Mater. 6, 1-6 (2020). Tran, R. et al. The open catalyst 2022 (oc22) dataset and challenges for oxide electrocatalysts. , 3066–3084 (2023). ACS Catal. 13 Jain, A. et al. Commentary: The materials project: a material genom approach to accelerating materials innovation. APL Mater. 1, 011002 (2013). I-Schütt, K. et al. Schnet: inethiwekhi ye-neural ye-convolutional ye-continuous-filter ye-modelling ye-quantum interactions. In Advances in Neural Information Processing Systems (eds Guyon, I. et al.) 991-1001 (I-Neural Information Processing Systems Foundation, Inc. (NeurIPS) 2017). Gasteiger, J., Groß, J. & Günnemann, S. Directional message passing for molecular graphs. in (ICLR, 2020). International Conference on Learning Representations (ICLR) Gasteiger, J., Giri, S., Margraf, J. T. & Günnemann, S. Fast futhi uncertainty-aware umyalezo directional ukuhambisa ama-molecule non-equilibrium. Preprint at https://arxiv.org/abs/2011.14115 (2020). I-Gasteiger, J. et al. GemNet-OC: Ukuphuhlisa i-graph neural networks yezinhlobonhlobo ezingu-molecular. I-Trans. Mach. Learn. Res. https://openreview.net/forum?id=u8tvSxm4Bs (2022). Zitnick, C. L. et al. Izincwajana ze-spherical ze-modelling ye-atomic interactions. Adv. Neural Inf. Process. Syst. 35, 8054–8067 (2022). Chmiela, S. et al. Machine learning of accurate energy-conserving molecular force fields. , e1603015 (2017). Sci. Adv. 3 Kolluru, A. et al. Open challenges in developing generalizable large-scale machine-learning models for catalyst discovery. , 8572–8581 (2022). ACS Catal. 12 Chang, C. & Medford, A. J. Ukusetshenziswa kwe-density functional tight binding kanye ne-machine learning ukulawula ukunambitheka kwama-biomass intermediates ku-Rh(111). J. Phys. Chem. C. 125, 18210-18216 (2021). Chan, L., Hutchison, G. R. & Morris, G. M. Bayesian optimization for conformer generation. J. Cheminform. 11, 32 (2019). Fang, L., Makkonen, E., Todorović, M., Rinke, P. & Chen, X. Ukusebenza kwe-amino acid conformer search nge-Bayesian optimization. J. Chem. I-Theory Comput. 17, 1955-1966 (2021). Xu, W., Reuter, K. & Andersen, M. Ukubuyekeza imithombo yokubambisana ye-adsorbates ehlanganisiwe usebenzisa umshini yokufunda nge-physics-inspired graph representation. Nat. Comput. Sci. 2, 443-450 (2022). Ulissi, Z. W. et al. Machine-learning methods enable exhaustive searches for active bimetallic facets and reveal active site motifs for co2 reduction. , 6600–6608 (2017). ACS Catal. 7 Ghanekar, P. G., Deshpande, S. & Greeley, J. I-Adsorbate i-chemical environment-based machine learning framework for heterogeneous catalysis. Nat. I-Commun. 13, 1-12 (2022). Schütt, K., Unke, O. & Gastegger, M. Equivariant message passing for the prediction of tensorial properties and molecular spectra. in 9377–9388 (PMLR, 2021). International Conference on Machine Learning U-S. Passaro noC. L. Zitnick, Ukunciphisa i-SO(3) i-convolutions ku-SO(2) ngoba i-Equivalant GNN efanelekayo. ku-Proceedings of the 40th International Conference on MachineLearning, Proceedings of Machine Learning Research, Vol. 202 (eds Krause, A. et al.) 27420–27438 (PMLR, 2023). Schaarschmidt, M. et al. Learned force fields are ready for ground state catalyst discovery. Preprint at (2022). https://arxiv.org/abs/2209.12466 Godwin, J. et al. I-Simple gnn regularization for 3d molecular property prediction futhi ngaphezulu. ku-International Conference on Learning Representations (ICLR) (ICLR*,* 2021). Ying, C. et al. Ngaba ama-transformers ngempumelelo kakhulu ekubunjweni kwe-graph? Adv. Neural Inf. Process. Syst. 34, 28877–28888 (2021). Shuaibi, M. et al. Ukuguqulwa kwe-invariant graph neural networks usebenzisa i-spin convolutions. Preprint at https://arxiv.org/abs/2106.09575 (2021). Kresse, G. & Hafner, J. Ab initio molecular-dynamics simulation of the liquid-metal-amorphous-semiconductor transition in germanium. , 14251–14269 (1994). Phys. Rev. B 49 Kresse, G. & Furthmüller, J. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. , 11169–11186 (1996). Phys. Rev. B 54 Kresse, G. & Joubert, D. From ultrasoft pseudopotentials to the projector augmented-wave method. , 1758 (1999). Phys. Rev. B 59 Kresse, G. & Furthmüller, J. Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set. , 15–50 (1996). Comp. Mater. Sci. 6 Okuzenzakalelayo iyatholakala ngaphansi kwe-CC by 4.0 Deed (i-Attribution 4.0 International) isivumelwano. Okuzenzakalelayo iyatholakala ngaphansi kwe-CC by 4.0 Deed (i-Attribution 4.0 International) isivumelwano.