Аннотация
The prediction of compressive strength of ultra-high-performance concrete (UHPC) is essential for optimizing mix ratios and supporting engineering applications. Traditional testing methods are costly and time-consuming, while empirical formulas often lack generalizability, limiting their effectiveness for precise design requirements. This study aims to establish a high-precision and interpretable UHPC compressive strength prediction and interpretation framework. Utilizing 595 valid mix ratio-strength datasets (the data were split into training and test sets at an 8:2 ratio with five-fold cross-validation used for model evaluation), we compared the performance of six ensemble algorithms: XGBoost, CatBoost, random forest (RF), gradient boosting decision tree (GBDT), AdaBoost, and LightGBM (LGBM). Bayesian hyperparameter optimization was used to identify the optimal parameter combination, and combined SHAP-SALib analyses were employed to achieve explicit model interpretability. Results indicate that the Bayesian-optimized CatBoost model performs best (test set: R 2 = 0.951, RMSE = 6.56 MPa, MAPE = 3.10%), followed by XGBoost (test set: R 2 = 0.925, RMSE = 8.067 MPa, MAPE = 3.573%). Both models demonstrate significantly superior generalization stability compared to the others. SHAP analysis identifies the key influencing factors in descending order: age and steel fiber. SALib analysis confirms that the combined contribution of water-binder ratio, age and Silica fume is substantial, with significant multi-factor interaction effects. This study establishes a high-precision and interpretable UHPC compressive strength prediction model and provides data-driven guidance for mix proportion design.