Rapid assessment of building vulnerability following an explosion is a prerequisite for crisis management and emergency decision-making. In this study, we developed a physics-guided, CatBoost-based framework with a regime-switching strategy to estimate damage percentage and instantaneous structural resilience. The dataset comprised 432 samples and 11 key variables, including blast parameters, structural geometry, and shielding conditions. The proposed method partitions records into general and specialized branches and produces the final output by integrating Ridge Residual Correction and Affine Calibration. The results showed that the proposed model achieved high predictive accuracy, with a mean absolute error (MAE) of 7.13, a root mean square error (RMSE) of 10.07, and an R^2of 0.872. Although XGBoost exhibited a marginally lower overall MAE of 7.00, paired comparative analysis indicated that this difference was not statistically significant; meanwhile, in direct-impact scenarios, the proposed model performed substantially better, reducing MAE from 11.21 to 8.19. Finally, SHAP and Accumulated Local Effects (ALE) analyses identified scaled distance as one of the main damage drivers, and the safety-oriented Mondrian interval achieved 93% coverage, supporting the model’s uncertainty quantification. These findings confirm the model’s utility as a reliable tool for rapid assessment of instantaneous resilience under critical conditions.