ABSTRACT In this study, a physics‐informed and experimentally anchored machine learning framework is developed to optimize perovskite solar cell (PSC) architectures. A hybrid dataset is constructed by integrating depth‐resolved optical generation profiles under AM1.5G illumination with a rigorously curated experimental PSC database, ensuring one‐to‐one correspondence between simulated optical descriptors and real device configurations. To preserve physical causality and prevent information leakage, a leakage‐safe two‐stage modeling strategy is implemented, in which intermediate electrical parameters (Voc, Jsc, and FF) are first predicted and subsequently propagated to the final PCE model. The Random Forest framework achieves high predictive accuracy and stability under grouped cross‐validation. SHAP analysis reveals nonlinear and interaction‐dominated effects of absorber and transport‐layer thicknesses, as well as bandgap. Sensitivity analysis and density‐based performance mapping (KDE and heatmaps) identify statistically dense high‐efficiency regimes rather than simple monotonic trends. Nearest‐neighbor validation confirms that top‐ranked predicted architectures lie within experimentally realizable structural neighborhoods. An evidence‐based design matrix defines optimal ranges for absorber (∼450–550 nm), HTL (∼180–200 nm), ETL (∼20–75 nm), and bandgap (∼1.55–1.60 eV). The framework remains robust under ±10% perturbations and provides an interpretable pathway for data‐driven PSC design.
SARVELAT et al. (Sun,) studied this question.