A unified machine learning framework for band gap prediction in lead-free ABX₃ and double perovskites

We present a machine learning framework for accurately predicting the band gaps of lead-free halide perovskites, including both ABX₃ and double perovskite A₂B′B″X₆ structures. Unlike many previous studies that rely on computationally expensive DFT calculations, our approach employs readily accessible electronic and mass-related features extracted directly using Python libraries. A dataset of 1016 compounds was used to train four algorithms: CatBoost, XGBoost, Random Forest, and Multi-Layer Perceptron with hyperparameters optimized via Optuna. Among these, CatBoost achieved the highest accuracy, with R² = 98.6% and RMSE = 0.151 eV. SHAP analysis revealed that B-site electronegativity is the key factor governing the band gap. Remarkably, when evaluated independently, the double perovskite subset achieved R² = 99.6% with the neural network model, ranking among the highest reported for this material class. These results demonstrate that combining simple, easily obtained descriptors with advanced machine learning enables rapid and accurate band gap prediction across diverse perovskite families, providing a practical and efficient alternative to conventional DFT-based methods. This framework paves the way for the accelerated design of stable, lead-free perovskite absorbers for next-generation photovoltaic applications.