Interpretable modelling of chloride ingress behaviour in seawater-exposed concrete

The durability of concrete in marine environments is a critical concern due to the aggressive ingress of chlorides, which accelerates the corrosion of reinforcement and compromises structural integrity. This study presents a data-driven framework for predicting chloride concentration (Cl ̄) in seawater-exposed concrete using advanced machine learning (ML) techniques. A curated dataset of experimental records was developed, encompassing a range of binder compositions, admixtures, aggregate proportions, and environmental exposure conditions. Three supervised ML models, Multi-Layer Perceptron (MLP), Gradient Boosting Regressor (GBR), and Extreme Gradient Boosting (XGBoost), were trained and evaluated to forecast Cl ̄ with high accuracy. Model performance was assessed using statistical metrics, with the MLP model achieving the highest accuracy (R² = 0.92), followed by GBR and XGBoost. Shapley Additive exPlanations (SHAP)-based feature contribution analysis revealed that annual mean temperature, exposure time, and water content significantly influenced chloride ingress. The results affirm the potential of ML models in supporting performance-based mix design and long-term durability assessments for marine concrete. Furthermore, to enhance the practical applicability of the proposed framework, a user-oriented graphical interface was developed, enabling rapid prediction of chloride concentration based on input material and environmental parameters without requiring expertise in machine learning This research offers a practical and scalable approach to chloride prediction, enabling engineers to optimise material selection and improve structural resilience under coastal conditions.