SHAP-guided Machine Learning for Interpretable Band Gap Prediction and Inverse Design in ABX₃ Perovskites
TL;DR
This work reports positive demonstration of inverse design of high confidence ABX3 candidates with 1.5–2.5 eV separations, such as the sub-optimized discovery of RbSnS3, which provides a scalable way towards rational discovery of perovskites.
Abstract
The (ABX₃) perovskites form the basis of the future of optoelectronics, but the limiting DFT calculations remain the bottleneck to high-throughput density screening. Our presented explainable machine learning (ML) framework, based on SHapley Additive exPlanations (SHAP), attains a mean absolute error (MAE) of 0.2644 eV in predicting band gaps and revealing physics-consistent drivers (lattice volume and coordination). Combining SHAP-informed feature engineering with gradient-boosted trees with validation on Materials Project (MP) data (that includes locating stable and novel candidates) will bridge the gap between accuracy and interpretability. Compared to literature results of Broad Learning Systems, interpretable GBRT + symbolic regression, attention-based networks, and Conv2D-SVM (Fourier descriptors), our framework is chemically useful, reproducible, and balanced with high interpretability. We report positive demonstration of inverse design of high confidence ABX3 candidates with 1.5–2.5 eV separations, such as the sub-optimized discovery of RbSnS3, which provides a scalable way towards rational discovery of perovskites. Uncertainty quantification through quantile regression and conformal prediction allows the framework to flag out-of-domain predictions automatically, as demonstrated by the NaAlF₃ case.