Predicting photovoltaic output parameters of perovskite solar cells using explainable machine learning with physics-informed and simulation-derived descriptors

The predictive modeling of perovskite solar cell performance is challenged by the strong coupling between structural, electronic, optical, and recombination-related processes. In this study, a simulation-driven and physics-informed machine learning framework is proposed to predict the key photovoltaic performance parameters, namely open-circuit voltage (V oc ), short-circuit current density (J sc ), fill factor (FF), and power conversion efficiency (PCE), using a comprehensive drift–diffusion simulation dataset. The model is constructed using 29 physically meaningful input descriptors that explicitly encode device geometry, charge transport, optical generation, recombination pathways, and electrical loss mechanisms, enabling a physically rich and systematic representation of perovskite solar cell operation. Six regression algorithms, namely Linear Regression, K-Nearest Neighbors, Support Vector Regression, Multilayer Perceptron, Random Forest, and Extreme Gradient Boosting, are evaluated using five-fold cross-validation. Among these, XGBoost consistently delivers the best predictive performance across all target parameters, achieving low prediction errors and strong correlations with the reference data (PCE: RMSE = 0.012, r = 0.951). Beyond numerical accuracy, SHapley Additive exPlanations (SHAP) are employed to provide transparency into the learned relationships. The analysis reveals that V oc is primarily governed by recombination-related mechanisms, J sc is dominated by photogeneration and transport processes, FF is strongly influenced by resistive losses, and PCE emerges as a coupled outcome of these interacting effects. The strong agreement between the SHAP results and known device physics shows that the model captures physically meaningful relationships rather than spurious correlations. Overall, this work demonstrates that combining physics-based drift–diffusion simulations with explainable machine learning enables accurate performance prediction, offering a robust pathway for data-driven analysis and optimization of perovskite solar cells. • Physics-informed machine learning is applied to predict key PSC performance parameters. • A drift-diffusion simulation dataset with 29 physical descriptors is employed. • XGBoost achieves the highest accuracy and stability for Voc, Jsc, FF, and PCE. • SHAP analysis reveals physically consistent mechanisms governing device performance. • The framework enables accurate and interpretable optimization of perovskite solar cells.