Machine learning-guided optimization of lead-free perovskite solar cells: predicting PCE with high accuracy

This study investigates the influence of absorber layer thickness and acceptor doping concentration on the photovoltaic performance of CsSn₀.₅Ge₀.₅I₃-based lead-free perovskite solar cells (PSCs), complemented by machine learning (ML) for predictive modeling of power conversion efficiency (PCE). The simulation results indicate that an optimal thickness of 0.8 µm enhances light absorption and charge extraction, yielding a short-circuit current density (JSC) of 27.36 mA/cm², while the overall maximum values obtained from the plots are 28.53 mA/cm² for JSC, 1.219 V for VOC, 89.84% for FF, and 31.29% for PCE. Increased acceptor doping enhances the built-in electric field but also introduces recombination losses beyond 10¹⁸ cm⁻³. To accelerate device optimization, three ML algorithms-Neural Network (NN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost)-were trained on a comprehensive dataset comprising 1000 samples with variations in thickness, doping concentration, and bulk defect density. Performance evaluation using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and coefficient of determination (R²) indicated that XGBoost achieved the highest predictive accuracy (R² = 0.997), followed by RF (R² = 0.995), while NN underperformed comparatively. Shapley Additive Explanations (SHAP) analysis further identified absorber thickness and acceptor concentration as dominant predictors of PCE. These findings underscore the importance of balanced physical design and the utility of ML in enhancing PSC optimization workflows.