Machine learning prediction of chloride convection-zone depth in marine tidal concrete with model explainability and experimental validation

Reliable prediction of chloride convection-zone depth is important for service-life assessment of tidal-zone concrete, but most data-driven chloride studies focus on surface concentration or diffusion-related indicators. This study proposes a data-driven prediction and explainability framework for chloride convection-zone depth. A database of 624 experimental measurements with 25 material, curing, exposure, and loading features was compiled from published studies. Seven machine-learning algorithms were evaluated under a unified training and testing procedure, and the regression outputs were further assessed for identifying measurable convection-zone occurrence using a 0.1 mm (mm) threshold. Shapley additive explanations (SHAP) and partial dependence plots (PDPs) were used to analyze model-learned feature effects, and independent drying-wetting experiments were conducted for validation. Extreme gradient boosting (XGBoost) achieved the best held-out regression performance, with coefficient of determination (R 2 ) = 0.958, root mean squared error = 0.446 mm, and mean absolute error = 0.262 mm. Light gradient boosting machine (LightGBM) also showed stable performance, with R 2 = 0.936. The classification-oriented assessment produced area under the receiver operating characteristic curve values above 0.98 for all models, and the classification conclusion remained stable under threshold-sensitivity analysis. SHAP and PDP results showed that environmental factors, including dry-wet ratio, exposure time, cycle period, and chloride ion concentration, made the largest contributions to model predictions. Material factors such as water-to-binder ratio mainly acted as secondary modifiers. The validation experiments supported the predicted two-stage response to dry-wet ratio. The framework provides a transparent sample-level prediction tool within the represented feature domain.