Prediction of Chloride Diffusion in Concrete Based on 10-Year Marine Exposure and Machine Learning

Abstract:

To predict the chloride diffusion behavior of high-performance concrete in marine environments， Five machine learning models—light gradient boosting machine（LightGBM）， categorical boosting（CatBoost）， extreme gradient boosting（XGBoost）， random forest（RF）， and artificial neural network（ANN）—were developed using a dataset from a 10-year in-situ marine exposure test in Zhanjiang city. The results show that among the five machine learning models， CatBoost demonstrates superior robustness and accuracy， achieving the highest determination coefficient（0.914 3） and the lowest error（mean absolute error， root mean square error， and mean absolute percentage error）. Shapley additive explanations（SHAP） identifies the water-to-binder ratio as the most critical determinant of chloride diffusivity， followed by cement content， exposure time， and silica fume content， whereas fly ash content shows the least significance. Furthermore， the proposed CatBoost model offers significantly higher prediction accuracy than that of the traditional standard calculation methods based on Fick’s law.