Predictive modeling and discovery of double perovskite oxides for photovoltaics

Discovery of potential oxide perovskites for solar photovoltaic applications remains a key challenge because of the size of the compositional space and the computational cost of first-principal methods. A multi-phase machine learning framework has been developed in this study for the prediction of electronic band gaps following a systematic feature selection and hierarchical integration of classification and regression models. By employing XGBoost classification model and weighted voting ensemble regression model which demonstrated superior predictive performance, 592 charge neutral and geometrically formable oxide perovskite compositions were identified within the photovoltaic relevant band gap range of 1.28–1.62 eV. This set was subsequently refined to 38 compositions upon the application of elemental sustainability and manufacturability constraints. These 38 structures are presented as band-gap compatible candidates prioritized for future validation.