Heterogeneity-aware stacked machine learning for predicting high-performance concrete compressive strength

Accurate prediction of compressive strength (CS) in high-performance concrete (HPC) is essential for optimizing mix design and ensuring structural reliability. Unlike conventional concrete, HPC incorporates low water-to-binder ratios, supplementary cementitious materials, and chemical admixtures that introduce stronger nonlinear interactions and greater mix-design variability, increasing modeling complexity. Traditional empirical models often struggle to capture these coupled effects. This study develops a heterogeneity-aware machine learning (ML) framework based on stacked ensemble modeling to enhance prediction accuracy, robustness, and interpretability. A dataset of 1,525 HPC mix designs compiled from five independent sources was used, incorporating eight mix and curing variables as predictors of CS. Sixteen regression algorithms, including tree-based models, kernel methods, and neural networks, were implemented as base learners. Their out-of-fold predictions trained meta-learners, with Multiple Linear Regression (MLR), Elastic Net Regression (ENR), Support Vector Regression (SVR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) evaluated as alternatives. In 10-fold cross-validation, all meta-model configurations achieved high predictive accuracy (R2 = 0.97–0.98). To evaluate generalization, heterogeneity-aware grouped cross-validation was performed by holding out one entire dataset at a time. Results revealed that models trained on mixed-source data can overestimate generalization and deteriorate when predicting unseen sources, a limitation not commonly evaluated in prior HPC studies. While stacking improved robustness, its accuracy was comparable to the best single model but offered enhanced stability across heterogeneous datasets and reduced systematic error. SHAP analysis confirmed the dominant influence of XGBoost and GB while identifying key material parameters governing CS. The proposed framework supports practical engineering decision-making, including mix optimization and early-stage strength assessment, and offers a scalable, interpretable approach for data-driven HPC prediction.