Abstract: An interpretable blast fragmentation size prediction model based on Bayesian optimization （BO）-XGBoost was built to address the problems of low prediction accuracy and poor generalization in blast fragmentation size prediction. Firstly, the dataset was preprocessed by adopting Spearman correlation analysis and random forest （RF）-based feature selection. Subsequently, the BO algorithm was employed to optimize the hyperparameters of the XGBoost model. Finally, SHAP analysis was conducted to interpret the model’s decision-making process. The results demonstrate that the BO−XGBoost model yields the highest prediction accuracy, and the determination coefficient （R²）, root mean square error （RMSE）, mean absolute error （MAE）, mean absolute percentage error （MAPE）, and explained variance （EV） on the test set are 0.961 0, 0.027 6, 0.026 3, 2.703 1, and 0.971 4 respectively, outperforming XGBoost, CatBoost, LightGBM, and AdaBoost models. During the prediction of the average blast fragmentation size, the elastic modulus contributes most significantly to the BO−XGBoost model’s prediction. A graphical user interface （GUI） for average blast fragmentation size based on the BO−XGBoost model was developed, and the interface featured a simple layout and easy operation, thereby enhancing the prediction efficiency of average fragmentation size in field applications. In a practical application at a mine, the model achieves an average absolute error of 0.03 m in average fragmentation size prediction, meeting on-site accuracy requirements and demonstrating favorable generalization capability and applicability. This study provides a solid foundation for the development of intelligent blasting techniques in large open-pit mining.