Strength Prediction in UHPC with XGBoost Model and Shapley Algorithm Interpretation

This study presents a comprehensive investigation into predicting the compressive strength of ultra-high performance concrete (UHPC) by combining advanced machine learning techniques with model interpretability methods. The XGBoost regression model is employed to capture complex, nonlinear relationships between multiple mixture parameters-including cement, silica fume, water, superplasticizer, and aggregate content-and the resulting UHPC strength. Extensive experimental data are used to train and validate the model, and the results demonstrate that XGBoost achieves excellent predictive accuracy, high robustness, and strong generalization performance compared to conventional regression approaches. To enhance interpretability and provide insights into the contribution of individual factors, the Shapley additive explanation (SHAP) algorithm is applied. The analysis reveals that the interaction between silica fume and cement content has a particularly significant impact on the predicted strength, emphasizing the importance of optimizing their proportions for mixture design. Furthermore, the SHAP heatmap indicates that only a small subset of samples exhibits Shapley values below the mean, suggesting that the dataset contains relatively few high-quality samples and highlighting areas for potential improvement in raw material selection. Through detailed SHAP-based analysis, the optimal range of silica fume dosage is identified as 0-320 kg, providing practical guidance for formulating UHPC with superior performance. In addition, error metrics and residual analysis confirm that the XGBoost model effectively captures the underlying data patterns while minimizing overfitting, reinforcing its suitability for engineering applications. The combined approach not only validates the predictive capability of XGBoost for UHPC strength estimation but also demonstrates the value of interpretable machine learning in revealing critical feature interactions and guiding practical material optimization. These findings offer a data-driven framework for improving UHPC design, supporting more efficient and reliable construction practices, and promoting the broader application of advanced computational methods in concrete technology.