Multi-objective concrete mix design for sulfate attack resistance using interpretable machine learning predictions

Sulfate attack is a common cause of concrete deterioration, leading to expansion, cracking, and reduced performance. Understanding and predicting the concrete expansion under sulfate attack is crucial for the durable design. In this study, 2445 expansion measurements of mortar and concrete under various sulfate attack conditions are used to build and compare predictive models using four machine learning algorithms: Kernel Ridge Regression (KRR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Multi-Layer Perceptron (MLP). The MLP model shows the highest prediction accuracy. To enhance interpretability, the Shapley Additive Explanations (SHAP) method was applied to the MLP model, identifying key variables governing expansion behavior. Furthermore, a multi-objective optimization framework was established by integrating the MLP model with the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS), simultaneously optimizing expansion, global warming potential (GWP), and life cycle cost (LCC). Under a 5% sulfate concentration, the performance-prioritized mix design derived from the weighting scheme (Expansion: Strength: GWP: LCC = 3:3:1:1) yielded an expansion of 0.017% after 365 days, a compressive strength of 65 MPa at 28days, a GWP of 355 kg CO₂-eq/m³, and an LCC of 406 CNY/m³. This result highlights that prioritizing mechanical performance and sulfate resistance leads to improved durability and strength, while also achieving relatively lower environmental and economic costs.