Interpretable prediction of chloride ingress in marine concrete using gene and multi-expression programming

Chloride penetration presents a major durability challenge for concrete structures located in coastal and marine settings. This research applies linear regression and advanced symbolic machine learning methods, specifically Gene Expression Programming (GEP) and Multi Expression Programming (MEP), to model and predict the non-steady state chloride migration coefficient in concrete systems containing supplementary cementitious materials. A comprehensive database from a wide range of concrete types and mix designs, was assembled from experimental studies. Twenty-two input features, reflecting material composition and environmental exposure conditions, were used to develop two predictive models. Model performance was evaluated using statistical metrics, Taylor diagrams, and residual error analysis. The MEP model outperformed both linear regression and GEP by providing more accurate predictions, reflected in its higher R² values and lower associated error rates. Furthermore, SHapley Additive exPlanations analysis provided transparent insights into the influence of key variables, with water-to-binder ratio, slag, and curing conditions identified as critical features. The results highlight the potential of MEP as a reliable, interpretable, and cost-effective alternative to conventional testing methods for assessing chloride migration. This approach supports the development of durable concrete and aligns with performance-based design practices for sustainable marine infrastructure.