Abstract: This study combines high-fidelity numerical analysis techniques and machine learning methods to investigate and predict the damage modes of Ultra-high toughness cementitious composite (UHTCC) panels under near-range explosion loads with a span of 600 mm and a TNT equivalent of up to 1 kg. An explosion response database of 300 samples of UHTCC panels is established based on the fluid-structure interaction algorithm, and both existing empirical prediction methods and interpretable machine learning methods are applied for analysis. The results indicate that the existing McVay and Morishita methods can only capture breach mode, while the UFC 3-340-02 method cannot predict any damage modes. The established machine learning model based on the eXtreme Gradient Boosting (XGBoost) algorithm performs excellently in stratified 10-fold cross-validation and can realize rapid prediction of near-field explosion damage of UHTCC panels. The feature importance under various damage modes is different, and the tensile-to-compressive strength ratio of the material plays a more significant role in spalling and bending damage. Based on this, material and structural designs can be optimized to improve the explosion resistance of protective structures.