Reinforced concrete (RC) columns are critical load-bearing members in engineering structures. Under explosive loading, they may sustain severe damage and potentially trigger the collapse of the entire structure. Therefore, rapid and accurate assessment of the damage state of RC columns subjected to explosive loading is essential for ensuring structural safety and implementing effective protective measures. Accordingly, a data-driven machine learning (ML) model was proposed to predict the damage indices of RC columns under explosive loading. To train the model, a comprehensive database comprising 3133 samples was established by integrating 259 samples collected from the literature and 2874 supplementary samples generated through numerical simulations. The database includes both RC columns with uniformly spaced stirrups and those with densified stirrups satisfying seismic design requirements. Eleven key parameters were selected as input features, with the RC column damage index taken as the output. Six ML models were employed to predict the damage indices under explosive loading, and their predictive performance was comparatively evaluated using four regression metrics. The results indicate that the tabular prior-data fitted network (TabPFN) achieved the best predictive accuracy and generalization performance, with a coefficient of determination of 0.989 on the test set and a mean absolute error and a root mean square error of 0.018 and 0.030, respectively. The interpretability of the TabPFN model was further analyzed using SHAP (Shapley additive explanations). The results show that charge mass, standoff distance, and column cross-sectional depth are the key features governing the damage severity of RC columns. Notably, the volumetric stirrup reinforcement ratio contributes more significantly to improving the blast resistance of RC columns than the longitudinal reinforcement ratio. Finally, comparison with independent finite element simulation results confirmed that the TabPFN model exhibits good predictive stability as a surrogate model under unseen loading conditions. In terms of computational efficiency, the proposed model achieves a single-prediction inference time of approximately 0.5 s, representing an efficiency improvement of nearly four orders of magnitude compared with conventional numerical simulation methods. The established model enables rapid surrogate prediction of RC column damage under explosive loading and provides guidance for preliminary parameter analysis in blast-resistant structural optimization design.