Interpretable Machine Learning Model for Evaluating Flexural Strength of Ultra High-Performance Concrete

Abstract

Ultra-high-performance concrete (UHPC) mix design remains experimentally expensive because many ingredients interact nonlinearly to govern flexural behavior. An interpretable machine-learning pipeline was developed to predict UHPC flexural strength from literature-derived mixes (317 observations, 14 input variables). Nine regression models were screened under a rigorous Monte-Carlo protocol (1,000 random 70/30 splits). Tree-based boosting dominated: on a representative split, CatBoost achieved R²test=0.928, RMSE = 1.980 MPa, MAE = 1.454 MPa, MAPE = 8.386%, with XGBoost close behind; Random Forest and Gradient Boosting formed a reliable second tier, while linear, SVR, and KNN underfit. Global and local interpretability (SHAP and PDP) revealed a stable hierarchy of drivers: steel fiber content and curing time were strongly beneficial; coarse aggregate content was deleterious and nearly monotonic; water became increasingly harmful at high dosages; superplasticizer exhibited an interior “sweet spot”; cement and silica fume were favorable (silica fume above ~100 kg/m³); sand was weakly positive; limestone powder was near-neutral. Guided by mean|SHAP|, the feature set was reduced from 14 to 9 variables with only a modest trade-off (R²test =0.916, RMSE = 2.143 MPa, MAE = 1.522 MPa, MAPE = 9.07%). External verification on an independent dataset confirmed generalization and preserved the correct nonlinear response to steel fibers in the practical 0-2% range. A lightweight GUI operationalizes the nine-input model, enabling rapid “what-if” exploration and reducing measurement burden by 36%. The results deliver both accuracy and transparency, distilling actionable rules for UHPC tailored to flexure-critical applications: prioritize steel fibers and adequate curing, cap coarse aggregate, and maintain water/superplasticizer within stable windows while using cement and silica fume to tune the matrix.