Interpretable Machine Learning Analysis of Perovskite Solar Cells Based on SCAPS-1D Simulations: Insights into Performance-Limiting Mechanisms

Abstract

This work presents an interpretable machine learning framework to investigate the factors influencing the efficiency of perovskite solar cells. A dataset of approximately 5,000 simulations was generated using SCAPS-1D simulations by varying key parameters, including bandgap, defect density, carrier mobility, doping concentration, and compositional features. Among the evaluated models, Random Forest achieved the best predictive performance. SHAP-based analysis was employed to provide both global and local interpretability. The results indicate that bandgap and defect density are the primary determinants of device performance, particularly in low-efficiency regimes. In contrast, interface energy alignment and compositional variables play a secondary role under near-optimal conditions. Moreover, charge carrier mobility and dopant density also influence performance in degraded devices.