End-to-End Prediction of Chloride Concentration Profiles in Marine Concrete Using Machine Learning

For durability design of concrete structures exposed to marine environments, the relationship between chloride concentration and penetration depth is critical as it determines the theoretical onset of rebar corrosion and thickness of the protective layer. However, current codes across the world typically rely on deterministic environmental category and design service life of concrete structure to obtain the value of thickness. To address this issue, the authors proposed an end-to-end machine learning method that directly predicted chloride concentration profiles from input variables, without estimating intermediate parameters, e.g., surface chloride concentration and chloride diffusion coefficient. Field exposure data was collected and compiled from 63 published studies, covering 44 exposure sites across 16 countries, yielding a total of 6365 data samples representing 701 chloride concentration profiles. Six end-to-end machine learning models were developed, among which tabular prior-data fitted network (TabPFN) performed best within the constructed dataset, achieving a 0.858 ± 0.025 for testing R2. SHapley Additive exPlanations (SHAP) analysis identified exposure depth and exposure duration as the dominant input features, and the spatial and temporal evolution of chloride concentration formed the principal basis for model prediction. Through parametric investigation and external validation, it was found that the TabPFN model can reasonably describe the attenuation of chloride concentration with depth and effectively capture the influences of different factors on chloride diffusion within concrete. With the designed GUI, the authors believed that it can be a convenient tool for the rapid prediction of chloride concentration, providing useful insights for subsequent durability assessment.