Machine learning-driven optimization of mixed-cation ABX3 perovskites for enhanced photovoltaic performance and stability

Abstract

We present a machine learning guided design approach for mixed-cation ABX₃ perovskites with monovalent copper (Cu⁺) as the A-site cation candidate for high photovoltaic performance and stability. First, we generated a dataset comprised of various perovskite compositions with materials properties such as electronic structure, carrier transport, and structural stability. Random forest predictors trained to accurately predict (R²>0.85) metrics of power conversion efficiency (PCE) and stability. A feature importance analysis indicated that the dominant predictors were bandgap (for PCE) and octahedral distortion (for stability). Our models recognised compositions with optimal ratios of Cu⁺/Cs⁺ mixing at the A-site, allowing maximum performance and stability. The most-performing compositions involved Cu⁺ content in the 4–8 unit range, especially when with Cs⁺ and Rb⁺. Orbital overlap between Cu⁺ and the MCl₆ octahedron was found to significantly enhance carrier mobility and reduce decomposition energy. Multi-objective optimization identified promising candidates with optimal properties, e.g., Cu8Pb1Sr3Cl24Br24I24 with high stability (0.92) and yet competitive PCE (0.25%). These results provide a foundation for designing stable, high-performance Cu⁺-based perovskites and demonstrate the success of machine learning in accelerating the discovery of new photovoltaic materials.