Machine-learning driven approach for exploration of properties of antimony chalcogenide perovskite based double absorber with back surface field layer

Abstract

To address the toxicity and stability concerns of lead-based perovskites, this study investigates a lead-free, antimony chalcogenide-based double-absorber solar cell with the structure WSe₂/Sb₂S₃/Sb₂Se₃/WS₂. Numerical simulations were performed using SCAPS-1D, followed by machine learning-based efficiency prediction using Support Vector Regression (SVR), Random Forest (RF), Stacked SVR + RF, and Extreme Gradient Boosting (XGBoost). The optimized configuration, with 0.2 µm Sb₂S₃ (shallow acceptor density: 10¹⁶ cm⁻³), 0.8 µm Sb₂Se₃ (shallow donor density: 10¹⁹ cm⁻³), achieved a power conversion efficiency (PCE) of 28.39%, V_OC of 0.97 V, J_SC of 33.32 mA/cm², and fill factor of 87.91%. All layers were modelled with a bulk defect density of 10¹⁵ cm⁻³ and an interfacial defect density of 10¹⁰ cm⁻². Among the ML models, XGBoost demonstrated the best performance with an MSE of approximately 0.003 and R² of 0.9996. SHAP analysis identified Sb₂Se₃ donor concentration as the most impactful feature, while Sb₂S₃ thickness had the least effect. This study showcases the potential of combining SCAPS-1D simulation with interpretable ML models for accelerated design and optimization of lead-free solar cells.