Residual-corrected XGBoost for interpretable prediction of oxide double perovskite band gaps

The band gap is a key screening parameter for oxide double perovskites, but prediction from composition alone remains difficult. We combine a capacity-limited XGBoost model with an SVR residual learner. The SVR uses out-of-fold residuals and receives the base prediction as an additional input. We test the model on 1306 rocksalt-ordered AA'BB'O₆ entries with GLLB-SC reference gaps. Across ten repeated 80/20 splits, mean test R 2 rises from 0.943 for Strong XGBoost to 0.958 for PA-WXGB-SVR. RMSE falls from 0.382 to 0.327 eV, and MAE from 0.279 to 0.235 eV. For true E g < 2 eV, MAE falls from 0.451 to 0.346 eV and recall rises from 0.612 to 0.725. We also examine the 0.93–1.61 eV range relevant to ideal single-junction solar cells. Recall rises from 0.358 to 0.533 and precision from 0.502 to 0.576. The mean number of selected entries also rises from 9.8 to 12.9. When both models select 20 candidates ranked by distance from 1.34 eV, mean target hits rise from 8.4 to 8.8. SHAP analysis identifies unfilled valence states and d-orbital occupation as important descriptors for the base model. All errors and screening hits are measured against GLLB-SC labels.