Comprehensive Structural Descriptors Enable Machine Learning Prediction of Perovskite Band Characteristics

Organic-inorganic hybrid and all-inorganic perovskites are promising candidates for next-generation optoelectronics. Accurate prediction of band characteristics is essential for guiding the rational design of high-performance perovskite devices. Although machine learning (ML) has been applied in this field, the soft lattices of halide perovskites exhibit significant structural distortions that are often difficult for conventional ML descriptors to capture. Here, we introduce a set of physically meaningful structural descriptors that directly quantify key distortion modes: octahedral distortion, cation displacement, and segregation index. The resulting model achieves high accuracy for prediction of bandgap, VBM, and CBM positions of two representative Br-alloyed lead halide perovskite systems (Cs_0.25FA_0.75PbI_xBr_3-x and CsPbI_xBr_3-x), underscoring the importance of structural information in ML-driven electronic-property prediction. Notably, we identified an unexpected inverse correlation between the standard deviation of Pb-X-Pb bond angles and the bandgap, which is associated with an enhanced contribution of Br to the VBM under higher structural order.