Interpretable XGBoost-Based Prediction and Mix Optimization of UHPC Compressive Strength

The mechanical properties of ultra-high performance concrete (UHPC) are nonlinearly coupled by multiple mixing parameters. Traditional test methods cannot fully reveal the parameter action mechanism, and existing machine learning prediction models generally have the "black box" defect. In this paper, 626 groups of UHPC test samples are integrated, the isolation forest algorithm is adopted to clean abnormal data, and an XGBoost compressive strength prediction model is constructed after parameter optimization by Optuna algorithm. Combined with OAT local sensitivity, Sobol global variance decomposition and SHAP interpretable theory, the independent effects and interactive coupling laws of each mixing ratio parameter are quantified. The results show that the coefficient of determination R² of the model on test set reaches 0.9682 with excellent prediction accuracy. Curing duration X22 acts as the core parameter governing compressive strength; its sensitivity rises continuously with variable perturbations, and it exhibits significant positive interactions with components including silica fume and sand content. Sand content X10, steel fiber dosage X19 and silica fume content X7 are secondary critical variables. X10 serves as the global core interaction node. Both steel fiber and silica fume present a nonlinear threshold characteristic for compressive strength: inhibition at low dosage and promotion at high dosage. Parameters such as coarse aggregate content, aggregate particle size and specimen width exert negligible independent influences. The findings reveal the influence laws of key mixture proportion parameters on the compressive strength of UHPC, which can provide references for the mixture design and engineering application of UHPC.