Composition engineering guided experimental fabrication of Cs(1-n)AnPb(1-m)BmX3 via machine learning for high-efficiency solar cells

Abstract

Fine-tuning the composition of cesium-lead halide to enhance the power conversion efficiency of solar cells is a challenging task. Machine learning is used to accelerate the screening of promising photovoltaic materials from 5,376 candidate structures within seconds. The predicted results show that B-site regulation has a greater impact on the electronic structure and optoelectronic properties, primarily due to its significant contribution at the band edge. CsPb_0.75Cu_0.25Br₃ and CsPb_0.75Cu_0.25I₃ are quickly screened and exhibit high spectral limited maximum efficiencies of 24.51 % and 32.46 %, respectively. SHAP analysis of feature importance reveals that both atomic mass and atomic number significantly influence bandgap prediction, impacting results on both global and local sample levels. The DFT calculations prove that Cu²⁺ introduces s and d orbitals at the band edge, creating additional channels for carrier transport and enhancing the density of states. This work provides guidance for the experimental study on composition engineering of perovskite materials.