Machine learning-assisted optimization of eco-friendly Sb₂S₃ thin-film solar cells: a sustainable SCAPS-1D and SHAP-based design framework

• Machine learning framework optimizes Sb₂S₃ solar cell design. • SHAP analysis reveals absorber defect density as key performance factor. • Critical defect threshold of 10¹⁶ cm⁻³ identified for high efficiency. • Optimized device achieves 17.69% PCE, a 79% relative increase. Antimony sulfide (Sb₂S₃) is a promising eco-friendly absorber for thin-film solar cells, yet its performance is limited by complex defect physics. This study introduces an interpretable machine learning (ML) framework to efficiently optimize and understand the TiO₂/Sb₂S₃/CuI device architecture. We generated a dataset of 3,621 unique device configurations using SCAPS-1D simulations and trained an Extreme Gradient Boosting (XGBoost) model to predict photovoltaic performance, achieving exceptional accuracy for power conversion efficiency (PCE, R² = 0.997). Crucially, SHapley Additive exPlanations (SHAP) analysis was employed to decode the model, moving beyond simple prediction to reveal quantitative physical insights. The analysis established a clear performance hierarchy, identifying the Sb₂S₃ bulk defect density as the dominant factor, with a critical threshold of 10¹⁶ cm⁻³ above which performance degrades catastrophically. It also uncovered conditional optimization rules, such as defect tolerance in thicker absorbers (>1.0 µm) and the necessity of a high-quality absorber to benefit from transport layer doping. Guided by these ML-derived rules, we designed an optimized device with a theoretically enhanced PCE of 17.69%, a significant improvement over the baseline. This work provides a sustainable, data-driven blueprint for the targeted development of high-performance, eco-friendly photovoltaics.