Machine Learning Stability and Bandgaps of Lead‐Free Perovskites for Photovoltaics

Abstract Compositional engineering of perovskites has enabled the precise control of material properties required for their envisioned applications in photovoltaics. However, challenges remain to address efficiency, stability, and toxicity simultaneously. Mixed lead‐free and inorganic perovskites have recently demonstrated potential for resolving such issues but their composition space is gigantic, making it difficult to discover promising candidates even using high‐throughput methods. A machine learning approach employing a generalized element‐agnostic fingerprint is shown to rapidly and accurately predict key properties using a new database of 344 perovskites generated with density functional theory. Bandgap, formation energy, and convex hull distance are predicted using validation subsets to within 146 meV, 15 meV per atom, and 11 meV per atom, respectively. The resulting model is used to predict trends in entirely different chemical spaces, and perform rapid composition and configuration space sampling without the need for expensive ab initio simulations.