In data science we trust: Machine learning for stable halide perovskites

The exploration of the compositional space halide perovskites is prohibitively costly using traditional experimental research strategies. Sun and colleagues demonstrated a physics-informed, machine-learning-guided, approach that overturns this limitation enabling rapid efficient maximization stability in mixed-cation perovskites. In perovskite photovoltaics, experimentalism king. Thousands groups across globe daily apply tried-and-true process fabrication, characterization, optimization to develop some most highly performing solar devices date. However, every new development requires enormous cumulative costs time, materials, equipment, facilities, workforce. Costs are only exacerbated by complex stoichiometries state-of-the-art Mixed-cation perovskites, specifically those incorporate Cs+, methylammonium (MA+), formamidinium (FA+) cations into their structure, form vast that, with study, exceptionally expensive explore. Given need drastically improve perovskites’ allow commercialization, difficulty navigating particularly limiting for both current compositions discovery ones. Many studies have navigated mixed MA+, FA+ (CsMAFA) aim improving film stability, but these works typically sampled an exceedingly small number potential compositions.1Jesper Jacobsson T. Correa-Baena J.P. Pazoki M. Saliba Schenk K. Grätzel Hagfeldt A. Exploration lead halogen high efficiency cells.Energy Environ. Sci. 2016; 9: 1706-1724https://doi.org/10.1039/c6ee00030dCrossref Scopus (485) Google Scholar Other sought model perovskite’s structure density functional theory (DFT) calculations, such simulations limited intensive processing requirements often affected low reliability from broad assumptions.2Saidaminov M.I. Kim J. Jain Quintero-Bermudez R. Tan H. Long G. F. Johnston Zhao Y. Voznyy O. et al.Suppression atomic vacancies via incorporation isovalent ions increase cells ambient air.Nat. Energy. 2018; 3: 648-654https://doi.org/10.1038/s41560-018-0192-2Crossref (361) An ideal system would require minimal input results computational locate stable composition little error. Such lacks feasibility human analytical can be rapidly implemented application machine learning algorithms. Machine-learning-based frameworks seen success analysis materials electrocatalysis, optoelectronics, superconductors.3Khabushev E.M. Krasnikov D.V. Zaremba O.T. Tsapenko A.P. Goldt A.E. Nasibulin A.G. Machine Learning Tailoring Optoelectronic Properties Single-Walled Carbon Nanotube Films.J. Phys. Chem. Lett. 2019; 10: 6962-6966https://doi.org/10.1021/acs.jpclett.9b02777Crossref PubMed (27) Scholar, 4Li Z. Achenie L.E.K. Xin Adaptive Strategy Accelerating Discovery Perovskite Electrocatalysts.ACS Catal. 2020; 4377-4384https://doi.org/10.1021/acscatal.9b05248Crossref 5Dung Le Noumeir Quach H.L. J.H. H.M. Critical temperature prediction superconductor: A variational bayesian neural network approach.IEEE Transactions on Applied Superconductivity. 30: 1-5Google These commonly operate basis closed-loop sequential optimization, where conditions trial used advise further trials, iteratively, until achievement target goal conditions. While cycles utilizing purely statistical means group data, lack generalizability larger datasets. Physics-based constraints employed as co-guideline algorithms wider contexts, massively expediting design experiments, results. issue Matter, al.6Sun S. Tiihonen Oviedo Liu Thapa Hartono N.T.P. Goyal Heumueller Batali C. al.A data fusion optimize perovskites.Matter. 2021; 4 (this issue): 1305-1322https://doi.org/10.1016/j.matt.2021.01.008Abstract Full Text PDF (30) applied physics-informed Bayesian (BO) algorithm characterization explore CsxMAyFA1-x-yPbI3 perovkskite. Without use learning, optimizing required thousands trials slightly varying cation ratios relied evaluation determine trends relating possibly introducing errors bias analysis. With co-workers’ approach, were located three 214 total films. “instability index,” parameter quantify degradation over was set minimization target, BO between tested future examine. entirely unconstrained trend determination within dataset, space, settling false minima or maxima stability. Integrating physics-based constraint grouping allowed improved developed mitigated minima. co-workers utilized predicted Gibbs free energy phase mixing (?Gmix) many act BO’s theorical constraint, which they extracted DFT-calculation structure. de-mixing inactive phases ?-CsPbI3 FAPbI3, DFT-calculated ?Gmix composition, represents one mechanism time. Chemical also occur mainly attributed unstable nature MA+ cation. This chemical instability holistically reflected indexes fabricated Combining calculated DFT results, mechanisms incorporated (Figure 1A). Across based guided algorithm, honed distinct regions CsMAFA combinatorial (Figures 1B 1C). smallest optimized index, denoted Region I, achieved more than 17 times greater iodide 3.5-times Cs0.05(MA0.17FA0.83)0.95Pb(I0.83Br0.17)3. insights study films function notable, impressive result lays reduction work enabled machine-learning algorithm. fabrication just 1.7% possible sample variations, created able predict entire superior widely referenced Goldschmidt tolerance factor 1D). solely examines possibilities space. Anions Br? Cl?, included highest devices, worth adding Benefiting significant effort required, other could studied, majority remain unexplored, even higher photovoltaic performances.7Sun Q. Yin W.J. Wei S.H. Searching cell genome techniques high-throughput calculations.J. Mater. C Opt. Electron. Devices. 8: 12012-12035https://doi.org/10.1039/d0tc02231dCrossref (15) Furthermore, while targets time through tuning, it not application. parameters describing film, including quantum yield, grain-size, defect matched suitable similar fashion likely mirror done devices,8Brandt R.E. Kurchin R.C. Steinmann V. Kitchaev D. Roat Levcenco Ceder Unold Buonassisi Rapid Photovoltaic Device Characterization Parameter Estimation.Joule. 2017; 1: 843-856https://doi.org/10.1016/j.joule.2017.10.001Abstract (33) where, unlike computationally calculation, single measurement enables draw correlations device properties. becomes increasingly useful examination higher-dimensional spaces, CsxMAyFA1-x-ySnrPb1-r(IpBrqCl1-p-q)3, inordinately difficult. Rather niche usage, et. schema readily adapted dramatically reduce cost associated material optimization. Their emblematic coming age research, far beyond what guidance enable. Here, we may soon begin see science-guided become exception, norm. perovskitesSun al.MatterFebruary 1, 2021In BriefData combines first-principle calculations experimentation end-to-end framework, allowing accelerated search alloyed without intervention. Full-Text Open Archive