Machine learning insights into band gap properties in halide-based perovskites

Abstract

Halide perovskites show great promise for applications in optoelectronic devices. The lead-free perovskites are attracting increasing interest due to their low toxicity and motivate the exploration of alternative compositions and structures, including A$_2$BX$_6$, A$_2$BB$^\prime$X$_6$, A$_3$B$_2$X$_9$, and A$_4$BX$_6$. Accurate predictions of a wide range of band gap energies are important for designing new materials. It is also desired to generate a direct relationship between the structural and elemental descriptors and the band gap energies. In this work, we develop machine learning models to predict band gap energies across various types of halide perovskites based on atomic and structural properties. Algorithms including ensemble tree-based methods, random forest regression (RFR), gradient boosted regression trees (GBRT), and extreme gradient boosting (XGB) showed strong predictive accuracy. We also analyzed feature importance to identify key descriptors, including B-site and X-site elemental properties, as well as the number of A- and B-site atoms, as primary factors influencing band gap energies. These results improve our understanding of the ML models and provide guidance for designing new halide perovskite materials.