Interpretable machine learning-assisted screening of perovskite oxides

Abstract

Perovskite oxides are extensively utilized in energy storage and conversion. However, they are conventionally screened via time-consuming and cost-intensive experimental approaches and density functional theory. Herein, interpretable machine learning is applied to identify perovskite oxides from virtual perovskite-type combinations by constructing classification and regression models to predict their thermodynamic stability and energy above the convex hull (E_h), respectively, and interpreting the models using SHapley Additive exPlanations. The highest occupied molecular orbital energy and the elastic modulus of the B-site elements of perovskite oxides are the top two features for stability prediction, whereas the Stability Label and features involving the elastic modulus and ionic radius are crucial for E_h regression. A classification model, which displays an accuracy of 0.919, precision of 0.937, F1-score of 0.932, and recall of 0.935, screens 682 143 stable perovskite oxides from 1 126 668 virtual perovskite-type combinations. The E_h values of the predicted stable perovskites are forecasted by a regression model with a coefficient of determination of 0.916, and root mean square error of 24.2 meV atom⁻¹. Good agreement is observed between the regression model predicted and density functional theory-calculated E_h values.