The chloride surface concentration (C s ) is a critical boundary parameter for evaluating the corrosion risk of reinforcement in marine concrete structures. This study develops an optimized, interpretable machine-learning framework for predicting C s using 642 field-exposure records from the literature. Three tree-based models, Gradient Boosting (GB), the Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting (XGB), were optimized using a structured grid search and compared with two ensemble methods, namely Stacking and Voting. The optimized LightGBM model was selected as the representative model because it provided a favorable balance between independent testing accuracy, model simplicity, computational efficiency, and interpretability. For the representative independent testing set, LightGBM achieved a strong predictive performance, with a root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²) of 0.139, 0.099, 0.274, and 0.878, respectively. To improve interpretability, Shapley additive explanations, partial dependence, and individual conditional-expectation analyses were used to identify and explain the dominant variables controlling C s . The exposure type, fine-aggregate content, exposure time, temperature, cement content, and environmental chloride concentration were the most influential factors, with nonlinear and sample-dependent effects. The uncertainty was further assessed using 100 repeated random train-test splits, yielding average values of 0.1757, 0.1190, 0.7906, and 0.3858 for RMSE, MAE, R 2 , and MAPE, respectively. These findings indicate that the optimized LightGBM provides satisfactory predictive accuracy, a physically meaningful interpretation, and practical preliminary support for service-life assessment.
Nguyen et al. (Mon,) studied this question.