Reliability-aware and interpretable machine learning for compressive strength prediction of fiber-reinforced concrete

Reliable prediction of fiber-reinforced concrete (FRC) compressive strength is central to data-driven mixture design, where mixture components interact nonlinearly. This study develops a protocol-aware machine-learning framework using 521 steel-fiber records for internal modeling and 271 basalt-fiber records for external evaluation. Twenty regressors are benchmarked on 12 raw inputs augmented with six engineered descriptors; Boruta selection retains 14 predictors. Across 30 repeated holdout experiments, ExtraTrees and CatBoost emerge as statistically indistinguishable top performers (R² = 0.94 – 0.96, RMSE = 3.9 – 4.2 MPa); CatBoost is selected for its balance of accuracy and interpretability. SHAP, ALE, permutation importance, and LIME identify water content, water–cement ratio, and coarse aggregate as principal drivers. Zero-shot external evaluation shows a sharp performance drop (R² = 0.52), while 30% target-domain calibration restores strong accuracy (R² = 0.83). Bayesian intervals achieve 92% coverage. The framework enables reliable, interpretable, uncertainty-aware FRC strength screening.