Explainable and uncertainty-calibrated stacked ensemble for predicting compressive strength of blended cementitious concrete

Accurate and transparent prediction of concrete compressive strength can support rapid screening of candidate mixture designs. This study develops an explainable, uncertainty-aware workflow that combines domain-informed features, including the water-to-binder ratio ( w / b ), superplasticizer-to-binder ratio ( SP / b ), and supplementary cementitious material fractions, with a stacked ensemble tuned using hybrid genetic algorithm–particle swarm optimisation (GA–PSO). Random Forest, XGBoost, and a shallow feed-forward neural network are combined through an Elastic–Net meta-regressor trained using out-of-fold base-model predictions. The analysis uses the public Yeh concrete compressive-strength dataset, comprising 1030 original records and 1005 records retained after documented data-integrity screening. Model development used age-stratified 10-fold cross-validation within an 80% modelling partition, while a disjoint 20% hold-out partition was excluded from model development and used for final evaluation. The stacked ensemble achieved an internal cross-validation RMSE of 2.485 MPa, MAE of 1.201 MPa, and R 2 = 0.941. On the untouched hold-out set, the corresponding values were 2.62 MPa, 1.27 MPa, and 0.936. Relative to GA–PSO-tuned XGBoost, the hold-out RMSE decreased from 5.05 to 2.62 MPa. Paired bootstrap resampling of the hold-out observations gave a 95% RMSE interval of [2.33, 2.96] MPa for the stack compared with [4.44, 5.66] MPa for tuned XGBoost. MC-dropout and residual-bootstrap uncertainty intervals showed near-nominal empirical coverage, while SHAP and Morris analyses consistently identified w / b , curing age, cement content, and SP / b as influential predictors. These findings establish strong internal predictive performance on the historical Yeh dataset; external validation on contemporary and project-specific concrete data is required before practical engineering use.