A hybrid machine learning and SCAPS 1D expedition for optimization of lead-free BeSiP₂ perovskite solar cells

Abstract

The chalcopyrite-type semiconductor BeSiP₂ has recently emerged as a promising lead-free absorber candidate for thin-film photovoltaics due to its direct bandgap, strong visible-light absorption, and high carrier mobility. However, a systematic understanding of its device-level behavior and optimization strategy remains limited. In this study, a comprehensive physics–data hybrid framework is developed to investigate the FTO/NiO/BeSiP₂/TiO₂/C–Cu solar cell, integrating SCAPS-1D simulations with machine learning (ML) based prediction and feature interpretation. The numerical analysis reveals that device performance is highly sensitive to transport-layer thicknesses, contact resistances, temperature, and illumination intensity. An optimized structure achieves a power conversion efficiency (PCE) of 22.94%, Voc = 0.89 V, Jsc = 31.0 mA cm⁻², and FF = 83%. The results indicate that high shunt resistance (Rsh) and suitable absorber thickness (~800 nm) maximize efficiency by suppressing recombination and enhancing charge transport. Ensemble learning models XGBoost and Random Forest were trained on simulation data, achieving R² > 0.99 with minimal prediction error. Both models identified Rsh, illumination intensity, and absorber thickness as dominant contributors to η, aligning closely with device physics. The proposed SCAPS–ML hybrid methodology enables accurate prediction and interpretability of photovoltaic trends, offering a scalable pathway for optimizing lead-free BeSiP₂-based solar cells and other emerging thin-film technologies.