Machine-Learning-Driven Exploration of HTL Effects in Mg₃SbBr₃ Perovskite Solar Cells for Ultra-High Efficiency

Abstract

The urgent pursuit of lead-free, highly efficient, and thermally stable perovskite solar cells (PSCs) has driven the exploration of novel inorganic absorbers. In this work, a comprehensive computational investigation is conducted on Mg₃SbBr₃-based PSCs by integrating SCAPS-1D device simulations with machine learning (ML) approaches to achieve a predictive understanding of performance-limiting factors and optimization routes. The SCAPS-1D simulation results reveal that the device structure FTO/TiO₂/Mg 3 SbBr 3 /CNTS/Au achieves a maximum power conversion efficiency (PCE) of 35.06%, with V OC = 1.02 V, J SC = 42.607 mA cm⁻², and FF = 80.68%, outperforming configurations using PbS-TBAI and Spiro-OMeTAD as HTLs. The inclusion of CNTS significantly reduces interface recombination, enhances hole mobility, and maintains superior thermal stability under varying illumination intensities and operating temperatures. A dataset derived from the SCAPS simulations was employed to train four ML regression models for predicting key photovoltaic parameters. Among them, CatBoost exhibited the highest predictive accuracy with R² values exceeding 0.99 across PCE, FF, J SC , and V OC . SHAP (Shapley Additive Explanations) analysis provided quantitative interpretability, identifying defect density, conduction band, and electron affinity as the dominant parameters influencing efficiency. These findings not only validate Mg₃SbBr₃ as a highly promising lead-free absorber but also demonstrate that coupling physics-based simulations with explainable machine learning significantly accelerates material optimization. This integrative approach offers a scalable route for designing next-generation, environmentally sustainable PSCs with theoretical efficiencies surpassing 35% and improved operational stability.