Discovering key factors determining perovskite bandgap under data scarcity inspired by knowledge distillation

Metal halide perovskites (MHPs) are promising candidates for next-generation optoelectronic applications due to their tunable compositions and exceptional optoelectronic properties. However, the compositional and structural diversity presents challenges for the optimization of their properties by traditional experimental and computational methods. Machine learning (ML) provides a data-driven approach for rapidly predicting material properties, yet in data-scarce MHP systems its predictive power and model explainability is limited. Herein, a small data ML framework tailored for MHP materials is developed, inspired by knowledge distillation to address this limitation. By generating a large set of inorganic MHP compositions via ionic substitution and utilizing the optimized unstructured learning model CrabNet_s, bandgap values are predicted with high precision. Additionally, physics-based chemical descriptors are designed, with CrabNet_s serving as a teacher model to train simpler student models, improving interpretability, while SHAP analysis identifies the contribution of each descriptor. B-site (the octahedrally-coordinated metal center in perovskite structures) electronegativity of inorganic MHPs is uncovered as the key compositional factor influencing bandgap values, enabling precise bandgap tuning by adjusting elemental ratios. This study enhances the application of ML in designing MHPs under small data conditions, offering a promising strategy for accelerating their development in optoelectronic applications.