Machine Learning-Driven Compositional Engineering of Intrinsically Stable Mixed Halide Perovskites for Photovoltaic and Optoelectronic Applications

The design of stable, efficient perovskite materials is crucial for advancing photovoltaic and optoelectronic applications. This study leverages machine learning (ML) to accelerate compositional engineering within the expansive chemical space of Hybrid Organic-Inorganic Perovskites (HOIPs). We focus on two primary objectives: (a) the exploratory analysis and prediction of optoelectronic (bandgap) and intrinsic stability (defect-formation energies) properties across an open-source dataset of 1,044 HOIPs. Using a pretrained universal graph-neural network force-field, CHGNet (Crystal Hamiltonian Graph Neural Network), we extended this dataset by calculating defect-formation energies at the A-, B-, and X-sites, essential for assessing intrinsic stability. Interpretability methods, including SHAP and feature importance from tree-based models, reveal physical mechanisms behind ML predictions, enhancing model transparency and reliability; (b) designing intrinsically stable, optoelectronically optimized mixed halide perovskites through multi-objective optimization. We integrate ML-predicted property models with a genetic algorithm to pinpoint compositions balancing optimal optoelectronic performance and long-term stability, providing actionable insights for next-generation material discovery.