Data-driven design of tunable band gaps in perovskite oxides via machine learning

초록

Band-gap prediction in perovskite oxides is crucial for optoelectronic applications (e.g., photovoltaics and photocatalysis), because it dictates light absorption, charge-carrier process, and device efficiency. To address the limitations of perovskites, such as their often wide and fixed band gaps that limit visible-light absorption, researchers have increasingly relied on compositional substitutions and doping, which introduce significant diversity but complicate predictions. However, challenges arise from their vast compositional diversity, strong electron correlations, and limitations in density functional theory, which often underestimates band gaps due to self-interaction errors and high computational costs. Machine learning addresses these by enabling rapid, data-driven predictions for complex compositions. In this study, we develop a machine learning framework that integrates band-gap prediction with literature-based validation for perovskite oxides. A curated descriptor dataset is used to benchmark three ensemble models (Random Forest, LightGBM, XGBoost), with Random Forest achieving the best performance (R-2 = 0.907, MAE = 0.177 eV, RMSE = 0.220 eV). To assess generalization beyond the training composition list, we further perform an out-of-sample "blind validation" by excluding oxides (BiFeO3, LaFeO3, and CaTiO3). Literature UV-vis/Tauc's plot report band gaps of 2.17, 2.20, and 3.00 eV, which agree with machine learning predictions within 0.02-0.08 eV. Building on this validated predictive capability, we demonstrate a target-driven screening strategy that prioritizes candidates within a photovoltaic-relevant band-gap window. Overall, the proposed framework provides an efficient route for data-driven band-gap screening of perovskite oxides, while reliability tags (e.g., uncertainty quantification, phase-stability assessment warnings) represent important next steps toward fully closed-loop inverse design.

키워드