Predicting and analyzing stability in perovskite solar cells: Insights from machine learning models and SHAP analysis

Despite significant advances in the efficiency of perovskite solar cells (PSCs) over the past decade, stability remains a critical challenge for commercialization. This study analyzes T80, T90 data, and aging curves from over 3000 PSC devices using machine learning algorithms such as random forest, XGBoost, and neural networks. Among these algorithms, the XGBoost model demonstrates remarkable performance in predicting T80 and T90, with a coefficient of determination of 0.9984 and 85% accuracy. The trained neural network model is able to predict long-time aging curves exceeding 3000 h with a prediction error of less than 10%. Furthermore, we use the SHAP tool to explain the established models and analyze the factors influencing PSC stability. Our findings reveal a correlation between stability and factors such as perovskite composition, bandgap, thickness, test temperature, and humidity conditions. Additionally, we find that highly stable PSCs should have FA, MA and Cs content between 0.6 and 0.8, 0-0.5, and 0-0.1, respectively, with an optimal FA:MA:Cs ratio of 0.75:0.19:0.06. Disregarding efficiency considerations, pure Sn or Pb PSCs exhibit good stability, with an optimal Br/I ratio of 0.25/0.75. This study presents a detailed method for analyzing and predicting PSC stability, providing guidance for developing high-efficiency and stable PSCs.