Explainable and MOWCA-optimized machine learning for compressive strength prediction and counterfactual mixture design of GGBS concrete

Abstract Accurate prediction of compressive strength in ground granulated blast furnace slag (GGBS) concrete is challenging due to complex nonlinear interactions among curing age, water content, and binder proportions. This study develops a comprehensive explainable machine learning framework integrating the Modified Water Cycle Algorithm (MOWCA), interpretable modelling, and counterfactual mixture design for GGBS concrete. Five algorithms: Support Vector Machine, Random Forest, AdaBoost, LightGBM, and CatBoost, were systematically optimized using MOWCA, a single-objective metaheuristic minimizing a composite fitness score (RMSE + MAE + (1 − R²)) via 10-fold cross-validation, and trained on 796 mixtures compiled from published experimental data (70:30 train-test split). CatBoost-MOWCA achieved the best predictive accuracy (R² = 0.989, RMSE = 2.225 MPa, MAE = 1.156 MPa, MAPE = 2.826%), with 92.05%, 97.91%, and 99.58% of predictions falling within ± 10%, ± 20%, and ± 30% error margins, respectively. SHAP analysis and partial dependence plots identified curing age, water-to-binder ratio, and water-to-cement ratio as the dominant strength-governing parameters, revealing pronounced nonlinear effects: strength plateauing beyond 90–120 days of curing, substantial reduction when the water-to-binder ratio exceeds 0.60, and strengths exceeding 80 MPa at ratios below 0.35. Counterfactual explanations translated these relationships into actionable, schedule-aware mixture-design strategies, identifying minimal parameter adjustments required to achieve target strength levels. The optimized model was deployed as an open-source graphical user interface for real-time prediction and mix design support. Relative to the most closely related prior study using the same dataset, the proposed framework improved testing R² from 0.979 to 0.989, while extending interpretability beyond conventional feature importance to include partial dependence and counterfactual analysis, providing a transparent, deployable framework for sustainable concrete mixture design.