From Explainable Machine Learning to Physics-Guided Design Rules for Inorganic Phosphors

The rational design of inorganic phosphors with targeted emission wavelengths is challenging because luminescence depends on strongly coupled structural, electronic, and excitation-dependent factors. Machine-learning studies in this area often prioritize predictive accuracy without systematically evaluating descriptor reproducibility and physical interpretability. Here, we develop a physics-guided explainable machine-learning framework that combines multi-method feature selection, descriptor stability analysis, and physicochemical interpretation for emission-wavelength prediction. A curated dataset of 2327 single-dopant phosphor compositions, initially represented by 125 physicochemical descriptors, was refined to 49 physically meaningful descriptors. Twelve complementary feature-selection methods were evaluated using eight stability metrics and 10-fold cross-validation. Pearson correlation and ANOVA-F independently converged on the same seven-descriptor subset with complete 10/10 fold-level reproducibility. The resulting Random Forest model achieved a mean cross-validated R² = 0.9308, MAE = 7.78 nm, and RMSE = 22.55 nm. The seven-descriptor model retained essentially the same predictive performance as the best fully reproducible 24-descriptor model while reducing descriptor dimensionality by about 71% relative to that model and by about 86% relative to the full physically meaningful candidate space. On the independent test set, RF and GB achieved R² values of 0.9192 and 0.9023, respectively. SHAP analysis, descriptor ablation, interaction analysis, and complementary feature-importance methods consistently identified the same descriptor hierarchy. The ionic radius of the emission-center site was the dominant predictor, followed by excitation source and dopant valency. These results suggest four practical considerations for phosphor screening, namely activator-site compatibility, dopant oxidation state, host chemical environment, and excitation conditions. The results show that descriptor selection should consider not only predictive accuracy but also reproducibility, compactness, generalizability, and physical interpretability, as confirmed by learning-curve analysis.