Explainable physics-informed machine learning for prediction of power conversion efficiency in ABX3 perovskite solar cells

The accurate prediction of power conversion efficiency (PCE) in perovskite solar cells is essential for accelerating material discovery and device optimization. Although machine learning (ML) models have demonstrated strong predictive capabilities, most existing approaches rely primarily on data-driven correlations and often provide limited physical interpretability. In this study, we propose an explainable physics-informed machine learning (PiML) framework based on Gradient Boosting (GB) for predicting the PCE of ABX₃ perovskite solar cells using experimental data. The proposed framework incorporates two physics-derived descriptors, namely the Goldschmidt tolerance factor and ionic mismatch, into both the feature representation and a physics-guided weighting mechanism, enabling domain knowledge to be integrated directly into the learning process. The proposed PiML-GB model achieved competitive predictive performance (R² = 0.930, MAE = 1.231%, and RMSE = 1.729%), comparable to both the baseline GB model and previously reported studies. Ablation analysis confirmed that the physics-derived descriptors and the physics-guided weighting strategy provide complementary contributions to the overall framework, while bootstrap confidence intervals and statistical tests demonstrated that the proposed model maintains predictive performance without significant degradation. Furthermore, quantitative physical consistency analysis, together with SHapley Additive exPlanations (SHAP), showed that the proposed framework exhibits stronger alignment with established perovskite structure–property relationships by assigning greater importance to physically meaningful descriptors, particularly ionic mismatch and the Goldschmidt tolerance factor. Overall, the principal contribution of the proposed PiML framework is not a substantial increase in predictive accuracy but the integration of physically meaningful knowledge that enhances physical consistency, interpretability, and scientific reliability while preserving competitive predictive performance.