Advanced prediction of perovskite stability for solar energy using machine learning

Abstract

In this work, we delve into the realm of perovskite materials with a comprehensive analysis on its structural and thermodynamic stability. Employing a machine learning approach, our study focuses on three important features for stability prediction such as formation energy (E_f), energy above hull (E_hull), and tolerance factor (TF). These features act as key indicators, allowing us to understand the intricate balance of energy and thermodynamic stability in perovskite structures for solar energy applications. We achieve this by training machine learning models on datasets generated computationally using DFT. Understanding the structural prediction of perovskite materials (ABX₃, ABO₃, ABO₂X and ABOX₂), whether thermodynamically stable or unstable, is critical for assessing their suitability for photovoltaic or photocatalytic applications. This study examines 14,199 mixed perovskite halides, oxides, and oxynitrides in order to determine the relationship between the aforementioned parameters and perovskite material composition. When compared to other machine learning models, using the ExtraTrees regression algorithm results in a higher accuracy of approximately 93.6 %, 94.75 %, and 98.41 % in predicting E_f, E_hull, and TF, respectively. The proposed method not only predicts E_f, E_hull, and TF, but it also aids in the discovery of new materials. We are particularly interested in ABO₃ and ABO₂N compositions from this perovskite family. We have come up with 306 stable perovskite oxides and 311 stable oxynitrides using our prediction. Among these, we discovered 45 novel compositions of perovskite oxynitrides (ABO₂N) and two novel compositions of perovskite oxides (ABO₃) that are energetically, thermodynamically, and structurally stable which need experimental validation further. Our prediction represents a robust, quick, and cost-effective strategy for illuminating new avenues in materials science and improving the understanding of the structural and thermodynamic behavior of perovskite materials. Furthermore, we present feature ranking, correlation, and display feature importance graphs and SHapley Additive Explanations (SHAP) relevant to structural stability prediction.