Rapid discovery of narrow bandgap oxide double perovskites using machine learning

It is urgent to discover new functional materials quickly, but experimental research is a huge challenge to search for target materials from the vast chemical space. Here, we propose a two-step machine learning strategy to accelerate the discovery of the photovoltaic oxide double perovskites. According to the leave-one-out crossvalidation results, the support vector classification (SVC) and support vector regression (SVR) methods are selected to establish the perovskite classification model and the bandgap regression model from three classification algorithms and three regression algorithms, respectively. The models perform well in cross-validation and independent test set validation, indicating their excellent predictive ability. The prediction accuracy of the SVC classifier on the test set reaches 0.968. For the SVR model, the bandgap correlation coefficient in the test set is 0.919. The SVC classifier filters out the candidates of perovskite structures from enormous virtual samples. Then the bandgaps of candidate perovskites are predicted by the SVR regression model. Successfully, 60 promising oxide double perovskites for photovoltaic applications are screened out from 6529 virtual samples. Especially 19 perovskites with bandgap values between 1.25 eV and 1.45 eV are close to the ideal bandgap value (1.34 eV). Further data analysis shows that Fe, Ni, Sc and Co occupying B '-site and Bi, Ta, Nb, Sb, V, and Mn occupying B '' site are most likely to form narrow-bandgap oxide double perovskites. This work provides an effective approach for the design and discovery of new oxide double perovskites via machine learning techniques.