Machine-learning prediction and multi-objective optimization of concrete sulfate resistance
Sulfate attack progressively deteriorates concrete in marine, saline-soil, and sulfate-rich environments. This study developed an interpretable machine-learning and multi-objective optimization framework for sulfate-resistance-oriented concrete design. A literature-based dataset containing 744 records from 21 publications and 18 input features was compiled; 549 records were retained after outlier screening. Support vector regression, random forest, gradient boosting, and XGBoost models were evaluated using Bayesian hyperparameter optimization and random 5-fold cross-validation. XGBoost achieved the best performance ( R²=0.96, RMSE = 0.0429, and MAE = 0.0280). SHAP-based interpretability analysis revealed that at the same concentration, magnesium ions erode concrete more severely than sodium ions; water-reducing agent and sand content correlate positively with durability, while water-to-binder ratio and coarse aggregate amount correlate negatively. XGBoost was coupled with NSGA-II to optimize durability, cost, and carbon emissions. A balanced Pareto solution achieved a corrosion-resistance coefficient of 1.24, a cost of 312.24 CNY/m³, and carbon emissions 226.00 kgCO₂/m³.