A data fusion approach to optimize compositional stability of halide perovskites

•Physics-informed machine learning enables accelerated search of stable perovskites•Closed-loop Bayesian optimization takes the human out decision-making loop•>17-fold higher stability achieved within a combinatorial space CsxMAyFA1?x?yPbI3•Improved perovskite thin-film translates into enhanced solar cell reliability Despite recent intensive efforts to improve environmental halide materials for energy harvesting and conversion, traditional trial-and-error explorations face bottlenecks in navigation vast chemical compositional spaces. We develop closed-loop framework that seamlessly marries data from first-principle calculations high-throughput experimentation single algorithm. This us achieve rapid CsxMAyFA1?x?yPbI3 perovskites while taking loop. envision this fusion approach is generalizable directly tackle challenges designing multinary materials, we hope our successful showcase on will encourage researchers other fields incorporate knowledge physics algorithms, applying hybrid models guide discovery high-dimensional Search resource-efficient spaces an outstanding challenge creating environmentally semiconductors. demonstrate physics-constrained sequential subsequently identify most alloyed organic-inorganic perovskites. fuse degradation tests phase thermodynamics end-to-end algorithm using probabilistic constraints. By sampling just 1.8% discretized (MA, methylammonium; FA, formamidinium) space, centered at Cs0.17MA0.03FA0.80PbI3 show minimal optical change under increased temperature, moisture, illumination with >17-fold improvement over MAPbI3. The thin films have 3-fold improved compared state-of-the-art multi-halide Cs0.05(MA0.17FA0.83)0.95Pb(I0.83Br0.17)3, translating without compromising conversion efficiency. Synchrotron-based X-ray scattering validates suppression decomposition minority formation fewer elements maximum 8% MA. anticipate can be extended wide range systems. instability limits their usage optoelectronics, such as cells, light emitters, lasers, photodetectors.1Boyd C.C. Cheacharoen R. Leijtens T. McGehee M.D. Understanding mechanisms improving photovoltaics.Chem. Rev. 2019; 119: 3418-3451https://doi.org/10.1021/acs.chemrev.8b00336Crossref PubMed Scopus (575) Google Scholar Compositional engineering is, date, one effective methods presence heat, humidity, sacrificing optoelectronic performance.2Jeon N.J. Noh J.H. Yang W.S. Kim Y.C. Ryu S. Seo J. Seok S.I. high-performance cells.Nature. 2015; 517: 476-480https://doi.org/10.1038/nature14133Crossref (4521) fact has led research spaces, AxByC1?x?yPb(IzBr1?z)3.3Saliba M. Polyelemental, multicomponent semiconductor libraries through screening.Adv. Energy Mater. 9: 1803754https://doi.org/10.1002/aenm.201803754Crossref (50) However, only small fraction been experimentally explored, part due prohibitively expensive brute force synthesis. paucity resulting inhibits generalization across diverse structural requiring each begin ab initio experimental investigations.4Khenkin M.V. Katz E.A. Abate A. Bardizza G. Berry J.J. Brabec C. Brunetti F. Bulovi? V. Burlingame Q. Di Carlo et al.Consensus statement assessment reporting photovoltaics based ISOS procedures.Nat. Energy. 2020; 5: 35-49https://doi.org/10.1038/s41560-019-0529-5Crossref (300) similar those faced by communities, including heterogeneous catalysts, battery electrodes, high-entropy metal alloys magnetic materials.5Rohr B. Stein H.S. Guevarra D. Wang Y. Haber J.A. Aykol Suram S.K. Gregoire J.M. Benchmarking acceleration learning.Chem. Sci. 11: 2696-2706https://doi.org/10.1039/c9sc05999gCrossref (26) Scholar, 6Zhang W. Mao Li Chen Z. Guo Phosphorus-based alloy advanced potassium-ion anode.J. Am. Chem. Soc. 2017; 139: 3316-3319https://doi.org/10.1021/jacs.6b12185Crossref (583) 7George E.P. Raabe Ritchie R.O. High-entropy alloys.Nat. 4: 515-534https://doi.org/10.1038/s41578-019-0121-4Crossref (799) field several others require new tools navigate these efficiently locate optima extract scientific insights.8Zhong Tran K. Min Dinh C.T. De Luna P. Yu Rasouli A.S. Brodersen al.Accelerated CO2 electrocatalysts active learning.Nature. 581: 178-183https://doi.org/10.1038/s41586-020-2242-8Crossref (247) 9Kumar Bossert H. McDonald Chatzidimitriou Ardagh M.A. Pang Lee Tsapatsis Abdelrahman O.A. Dauenhauer P.J. Catalysis-in-a-Box: robotic screening catalytic time COVID-19 beyond.Matter. 3: 805-823https://doi.org/10.1016/j.matt.2020.06.025Abstract Full Text PDF (8) 10Burger Maffettone P.M. Gusev V.V. Aitchison C.M. Bai X. Alston B.M. Clowes al.A mobile chemist.Nature. 583: 237-241https://doi.org/10.1038/s41586-020-2442-2Crossref (191) 11Cole A design-to-device pipeline data-driven discovery.Acc. Res. 53: 599-610https://doi.org/10.1021/acs.accounts.9b00470Crossref (27) 12Gómez-Bombarelli Aguilera-Iparraguirre Hirzel T.D. Duvenaud Maclaurin Blood-Forsythe Chae Einzinger Ha D.G. Wu al.Design efficient molecular organic light-emitting diodes virtual approach.Nat. 2016; 15: 1120-1127https://doi.org/10.1038/nmat4717Crossref (465) 13Kirman Johnston Kuntz D.A. Askerka Gao Todorovi? Ma Privé G.G. Sargent E.H. Machine-learning-accelerated crystallization.Matter. 2: 938-947https://doi.org/10.1016/j.matt.2020.02.012Abstract (38) 14Li Najeeb Alves L. Sherman Parrilla P.C. Pendleton I.M. Zeller Schrier Norquist A.J. Chan E. Robot-accelerated investigation (RAPID): 1. Inverse Temperature Crystallization.ChemRxiv. https://doi.org/10.26434/CHEMRXIV.10013090.V1Crossref Machine learning-based approaches (e.g., [BO]) emerged explore variable “closed-loop” fashion, whereby outcome round informs next intervention. BO, which attracted increasing attention developments self-driving laboratories various science, recently successfully directed hole-transport materials,15MacLeod B.P. Parlane F.G.L. Morrissey Häse Roch L.M. Dettelbach K.E. Moreira Yunker L.P.E. Rooney M.B. Deeth J.R. al.Self-driving laboratory materials.Sci. Adv. 6: eaaz8867https://doi.org/10.1126/sciadv.aaz8867Crossref (102) piezoelectric oxides,16Xue Balachandran P.V. Yuan Hu Qian Dougherty E.R. Lookman Accelerated BaTiO3-based piezoelectrics vertical morphotropic boundary learning.Proc. Natl. Acad. U S 113: 13301-13306https://doi.org/10.1073/pnas.1607412113Crossref (84) photocatalysts.10Burger Within combined liquid synthesis microcrystal crystallization.17Epps R.W. Bowen M.S. Volk A.A. Abdel-Latif Han Reyes K.G. Amassian Abolhasani Artificial chemist: autonomous quantum dot bot.Adv. 32: 2001626https://doi.org/10.1002/adma.202001626Crossref (70) 18Gu Tang Langner Duchstein Zhao Levchuk I. Kalancha Stubhan Hauch Egelhaaf H.J. al.Robot-based antisolvents lead perovskites.Joule. 1806-1822https://doi.org/10.1016/j.joule.2020.06.013Abstract 19Higgins Valleti S.M. Ziatdinov Kalinin Ahmadi Chemical robotics enabled exploration photoluminescent behavior via learning.ChemRxiv. https://doi.org/10.26434/CHEMRXIV.12436079.V1Crossref 20Choubisa Ryczko Voznyy O. Mills Tamblyn Crystal site feature embedding large spaces.Matter. 433-448https://doi.org/10.1016/j.matt.2020.04.016Abstract (15) model-free statistical shows limitations principled guidance domain expertise, because it learn everything scratch. Recent situ experiments independently revealed insights fundamental composition-dependent alloys; however, merging computational selective compositions policy entire remains challenge.3Saliba State-of-the-art two-step theoretical hard constraint before shortlisted are limited inefficiencies arising from: (1) high-performing systems often too sparse experimentation, (2) discrepancies between calculation assumptions non-thermodynamic equilibria decreases accuracy.21Curtarolo Hart G.L. Nardelli Mingo N. Sanvito Levy highway design.Nat. 2013; 12: 191-201https://doi.org/10.1038/nmat3568Crossref (1110) Scholar,22Jesper Jacobsson Correa-Baena J.-P. Pazoki Saliba Schenk Grätzel Hagfeldt Exploration mixed halogen high efficiency cells.Energy Environ. 1706-1724https://doi.org/10.1039/C6EE00030DCrossref lack physics-informed iterative hinders ultimate goal stability. Here, introduce both Gibbs free mixing (?Gmix) density functional theory (DFT) calculation23Schelhas L.T. Christians Goyal Kairys Harvey S.P. D.H. Stone K.H. Luther Zhu al.Insights operational processing layers.Energy 1341-1348https://doi.org/10.1039/c8ee03051kCrossref (77) quantified aging every decision BO making. apply optimize iodide suffer severe heat moisture-induced five-element CsxMAyFA1?x?yPbI3. Under multiplex stress air, identified overperforming MAPbI3 starting point 17-fold reference composition (Cs0.05(MA0.17FA0.83)0.95Pb(I0.83Br0.17)3) three rounds, results found transferable device DFT here serves constrain not chemically, but also structurally ?-perovskite alloys. To search, constructed batch (Figure 1). In promising suggested acquisition function, expected improvement, EI(?), balances exploitation regions high-uncertainty space. As key contribution, ?Gmix function “composition selection” step, providing additional information effectively multi-cation thermodynamically relative single-cation counterparts (Figures 1A 1B). define “instability index” (Ic), figure merit optimizing round, consists steps selection, “film synthesis,” quantification,” minimize value. Our makes use surrogate ML model, Gaussian process (GP) regression,24SheffieldML/GPyOpt: Process Optimization Using GPy n.d..https://github.com/SheffieldML/GPyOptGoogle estimate value uncertainty Ic non-explored (see Experimental procedures). (one BO), 28 spin-coated samples 1C) examined parallel chamber 85% humidity (RH) 85°C air S1). 0.15 Sun visible applied enable automatic image capture 5 min RGB camera (~200 ?m resolution). Photoactive phases exhibit band gap ~1.5 eV, whereas main products hot humid conditions, PbI2 (2.27 eV)25Zhu X.H. Wei Z.R. Jin Y.R. Xiang A.P. Growth characterization crystal used gamma ray detectors.Cryst. Technol. 2007; 42: 456-459https://doi.org/10.1002/crat.200610847Crossref (71) ?-CsPbI3 (2.82 eV),26Hu Liu Ji Miao Qiu Zhang Bismuth incorporation stabilized ?-CsPbI3 fully inorganic cells.ACS Lett. 2219-2227https://doi.org/10.1021/acsenergylett.7b00508Crossref (349) or ?-FAPbI3 (2.43 eV)27Masi Echeverría-Arrondo Salim K.M.M. Ngo T.T. Mendez P.F. López-Fraguas Macias-Pinilla D.F. Planelles Climente J.I. Mora-Sero Chemi-structural stabilization formamidinium embedded dots.ACS : 418-427https://doi.org/10.1021/acsenergylett.9b02450Crossref (44) deteriorated photophysical properties S2). shown Figure 1D, hence color-based metric proxy macroscopic evolution high-band-gap, non-perovskite Videos S1 S2).28Hashmi S.G. Tiihonen Martineau Ozkan Vivo Kaunisto Ulla Zakeeruddin Long term processed inkjet infiltrated carbon-based printed cells intense ultra-violet soaking.J. 4797-4802https://doi.org/10.1039/c6ta10605fCrossref (60) 29Hashmi Dar M.I. Sarikka Paltakari Lund P.D. Air carbon reproducibility.Adv. 1600183https://doi.org/10.1002/admt.201600183Crossref (104) 30Stoddard R.J. Dunlap-Shohl W.A. Qiao Meng Kau W.F. Hillhouse H.W. Forecasting decay performance transmittance reflected dark-field imaging.ACS 946-954https://doi.org/10.1021/acsenergylett.0c00164Crossref index (Ic) integrated color unencapsulated film test duration Complementary direct band-gap measurements after UV-vis spectroscopy listed S12.Ic(?)=?c={R,G,B}?0minT|c(t,?)?c(0,?)|dt,(Equation 1) where ?=(x,y,1?x?y), t time, c area-averaged, color-calibrated red, green, blue pixel values sample. cutoff was set T = 7,000 observed divergence most- least-stable S3). workflow systematic against varying nominal compositions, ?, (x, y limit two decimal places) (Tables https://www.cell.com/cms/asset/b8d56803-5bb1-4c11-8a89-20768dc0abcd/mmc3.mp4Loading ... Download .mp4 (0.15 MB) Help files Video S1. Optical changes initialization Round 0 https://www.cell.com/cms/asset/74c8d777-8b82-4d42-bed7-e64ea6fcb67d/mmc4.mp4Loading (0.16 S2. 3 Due polymorphic nature, identical crystallized different behaviors, making essential evaluate stabilities any optimization.1Boyd end members study consist cubic ?-FA/MAPbI3 temperature.31Kim Eom Y.-S. Hong K.-H. Thermodynamics perovskites: highly materials.Chem. 4265-4272https://doi.org/10.1021/acs.chemmater.0c00893Crossref (9) Phase de-mixing during leads are, therefore, captured Ic. Nevertheless, soon desirable causes deterioration electronic perovskite.32Knight Herz Preventing segregation mixed-halide perspective.Energy 13: 2024https://doi.org/10.1039/d0ee00788aCrossref Schelhas al.23Schelhas demonstrated predict tendency ?-CsxMAyA1?x?yPbI3 (Gmix) polymorphs APbI3 (A Cs, MA, FA) (G0) given temperature. mixing, loop 2A). allows ?- ?-phase non-degraded considered thus enabling reduce probability formation. Data refers techniques map more datasets coming related distinct distributions. case, relate ?Gmix(?) Ic(?). streams account modeled thermodynamic measured thermal-moisture instability, respectively. Hence, inadequate combine equivalent include prior following BO.33Herbol H.C. Poloczek Clancy Cost-effective discovery: multiple sources.Mater. Horiz. 7: 2113https://doi.org/10.1039/D0MH00062KCrossref Scholar,34Doan H.A. Agarwal Counihan M.J. Rodríguez-López Moore J.S. Assary R.S. Quantum chemistry-informed accelerate design sustainable storage 6338https://doi.org/10.1021/acs.chemmater.0c00768Crossref (14) data-fused according Equation 2:P(?Gmix(?),?DFT)=11+e??Gmix(?)/?DFT,(Equation 2) P(?Gmix(?),?DFT) logistic cumulative distribution modeling ?DFT parameter calibrated control smoothness boundaries unstable forming soft 2A procedures details). Given cost complexity systems, first regress 85 DFT-modeled 47 binary (29 MAFA CsFA 12 CsMA computed present work same methods) quasi-ternary CsxMAyA1?x?yPbI3 auxiliary GP model defines ?Gmix(?). visualizes P(?Gmix(?),?DFT)?[0,1], low suggest (?Gmix>> 0) (?Gmix << 0). inspired unknown proposed Gelbart al.35Gelbart Snoek Adams R.P. constraints.Uncertain. Artif. Intell. Proc. 30th Conf. UAI. 2014; 2014: 250-259Google developing instead boundary, able discount predicted go rather than completely exclude unfavorable regions. accounts inherent predictions, accuracy, scarcity threshold calibration). adapt optimizations loop, thereby achieving sample-efficient being exact S4 S5). integrate formulation, weigh obtain DFT-weighted EIC(?),as illustrated 2A. Traditional EI(?) utilizes indicates potential Cs-poor Cs-rich regions, EIC(?) reduces energetically despite Ic: subsequent rounds converge among 2A, S6, S7). Comparisons weighting teacher-student Figures S8 S9, that, fusion, continues instability. 2B demonstrates sequentially identifies tests. Iterative landscape (posterior mean Ic, Ic(?), uncertainty) presented S3 S4. 2C reveals decrease Rounds 0–3. converges S5 convergence conditions) optimal region bounded 8%–29% <14% 68%–92% FA. identification global optimum lying FA-rich, Cs- MA-poor consistent reports FA-rich superior MA-rich less volatile Cs enhance moisture resistance.36Saliba Matsui J.Y. Domanski J.P. Nazeeruddin M.K. Tress Cesium-containing triple cation cells: stability, reproducibility efficiency.Energy 1989-1997https://doi.org/10.1039/C5EE03874JCrossref Interestingly, local near Cs0.26MA0.36FA0.38PbI3, sampled four validated non-intuitive reproducible. ability rapidly success major advantage automated strategies leveraging intuition alone. Further validation interest discussed subsection. quasi-ternary-phase subdivided minimum achievable resolution (1% composition). yields 5,151 possible singular, binary, ternary were converging (i.e., 94 unique 112 0–3, see Supplemental Three seven representative performed validate trend, Table S10–S13. find overall non-linear. quantify de