Interpretable Machine Learning for Bandgap Prediction and Descriptor‐Guided Design Rules of Phosphates

The vast chemical diversity of crystalline phosphates and the high cost of first‐principles calculations hinder rapid discovery of wide‐bandgap materials. Here, we develop an interpretable machine‐learning framework for phosphate bandgap ( E g ) prediction and descriptor‐guided design using 474 structures from the NOEMD database. A Bayesian‐optimized Categorical Boosting (CatBoost) regressor achieves high accuracy on an independent test set ( R 2 = 0.94, MAE = 0.18 eV). SHapley Additive exPlanations reveal two governing, physically meaningful descriptors—the standard deviation of d ‐valence electron counts (), and the standard deviation of atomic density ()—that dominate E g regulation across diverse compositions and structures. and are negatively associated with E g , linking electronic‐configuration heterogeneity and packing‐density inhomogeneity to bandgap narrowing. We translate these descriptor‐property relationships into actionable screening and design guidelines. This work provides an interpretable strategy to accelerate the discovery of phosphate‐based wide‐bandgap crystals for ultraviolet/deep‐ultraviolet and related optoelectronic applications.