Highly versatile and accurate machine learning methods for predicting perovskite properties

Abstract

The determination of band gaps in unidentified materials has substantial importance for photovoltaic applications. In this research we used machine learning techniques to predict the band gap of perovskite materials within an acceptable margin of error. We developed a model to predict the band gaps of inorganic perovskites using machine learning algorithms. Our methodology utilizes a comprehensive dataset of 3720 ABX3-type perovskites and 2660 A₂B(I)B(II)X₆-type double perovskites, encompassing key properties such as band gap and formation energy. The features include nearly 300 descriptors generated by Matminer python. We applied six machine learning models, including XGBoost. The most effective model, XGBoost, demonstrated a notable R² coefficient of 0.873 and a root mean square error (RMSE) of 0.5868 eV. Lastly, we conducted SHAP (SHapley Additive exPlanation) analysis to identify the most influential descriptors. The findings indicate that a higher formation energy, a significant proportion of transition metals, and a large number of d orbital valence electrons contribute to the formation of narrow bandgap perovskites. Conversely, a substantial number of f-orbital electrons and electronegativity differences between elements tend to result in wide-bandgap perovskites. This comprehensive analysis not only offers insights into the fundamental factors influencing the band gap of perovskite materials but also underscores the potential of machine learning in expediting materials research.