SHAP-based interpretable machine learning for predicting the oxide double perovskite band gaps

Abstract Oxide double perovskite has excellent optical properties. In recent years, using machine learning (ML) to study the provskite properties has become a hot topic, however, the corresponding model and characteristic explanation have been missed. We propose a ML-based method to interprete the feature importance of band gap in the oxide double perovskite from the global level by using multiple models and to validate the result from the local level. Four different ML algorithms Random Forest (R 2 :0.883, RMSE:0.535), Gradient Boosting Regression (R 2 :0.835, RMSE:0.639), Decission Tree Regression (R 2 :0.801, RMSE:0.712), and eXtreme Gradient Boosting (R 2 :0.888, RMSE: 0.528) are used, which have good predictive performance. The results suggest that the average number of valence electrons in the d- orbitals (mean NdValence), the maximum melting point (maximum MeltingT), and the range of unfilled valence electrons(range NUnfilled) of the constituent elements in the material have the greatest influence on the band gap in the oxide double perovskite. It has a positive effect on the band gap when the mean NdValence is less than 1.5, and a negative effect when the mean NdValence is greater than 1.5. When the range of NUnfilled is less than or equal to 6, it negatively affects the band gap. In addition, by using models to explain specific materials, we found that for small band gap and large band gap materials, the features have negative and positive effects, respectively, the result is further supported by the calculated SHAP value of different range of band gaps. Our method provides valuable insights for the application of oxide double perovskites in the field of photovoltaics.