Rationalizing Perovskite Data for Machine Learning and Materials Design.

Machine learning has been recently used for novel perovskite designs owing to the availability of a large amount of perovskite formability data. Trustworthy results should be based on the valid and reliable data that can reveal the nature of materials as much as possible. In this study, a procedure has been developed to identify the formability of perovskites for all of the compounds with the stoichiometry of ABX3 and (A'A″)(B'B'')X6 that exist in experiments and are stored in the Materials Projects database. Our results have enriched the data of perovskite formability to a large extent and corrected the possible errors of previous data in ABO3 compounds. Furthermore, machine learning with a multiple models approach has identified the A2B'B″O6 compounds that have suspicious formability results in the current experimental data. Therefore, further experimental validation experiments are called for. This work paves a way for cleaning perovskite formability data for reliable machine-learning work in future.