Machine learning-based modeling of double perovskites via multi-output prediction of bandgap and dielectric properties

Double perovskites are promising candidates for next-generation photonics, optoelectronics, and photovoltaics owing to their excellent optical and electronic properties, low-cost fabrication, and reduced toxicity relative to lead-based perovskites. Systematic first-principles exploration is hindered by the vast compositional and structural diversity of these systems. To address this, we propose an integrated machine learning (ML)–density functional theory (DFT) framework to identify efficient double perovskite materials. We develop a multi-output regression model that simultaneously predicts bandgap and dielectric constant across a large compositional space, leveraging intrinsic correlations between these properties to improve physically informed predictions. Five regression algorithms—Random Forest, Gradient Boost, XGBoost, CatBoost, and AdaBoost—are evaluated, with XGBoost performing best, achieving a combined coefficient of determination (R 2 ) of over 0.90. The XGBoost model attains (R 2 ) values of 0.69 and 0.95 for bandgap and dielectric constant, with mean squared errors of 0.2 and 0.4 and root mean squared errors of 0.4 and 0.6, respectively. SHAPley analysis reveals key descriptors and compositional trends, while DFT validation confirms physical reliability with errors below 18 %. This ML–DFT approach accelerates discovery and rational design of stable, high-performance, and environmentally benign optoelectronic materials.