Accelerating stability of ABX3 perovskites analysis with machine learning

Abstract

Perovskites have been widely utilized in the fields of photoelectrochemical (PEC) and photovoltaic (PV) due to their exceptional characteristics. However, the search for stable perovskite materials among thousands of perovskite materials still poses a significant challenge. In this study, we developed an optimal model for predicting the thermodynamic stability of perovskites using machine learning (ML) based on a dataset of 2877 ABX3 (X=O、F、Cl、Br、I) perovskites. Both four classification and four regression ML algorithms were employed and evaluated using the five-fold cross-validation approach. For classification models to distinguish stable perovskites, the Gradient Boosting Classification (GBC) exhibits the highest accuracy and AUC values of 0.872 and 0.939, respectively. For regression models to predict Ehull values, the eXtreme Gradient Boosting Regression (XGBR) shows the best performance, with RMSE of 0.108 and R2 of 0.93. Furthermore, further model validation suggests that combining both models can obtain a more accurate predictions on the perovskite stability. Subsequently, analysis of hidden structure-properties trends reveals a strong dependence of perovskite stability on the elements occupying the A-site. Finally, 23 and 18 stable perovskite compounds with suitable bandgap for PEC and PV applications were also screened, respectively. Our research demonstrates the enormous potential of ML in accelerating the analysis of stability in ABX3 perovskites.