Robust machine learning framework for predicting chloride ion diffusion in concrete under load and freeze-thaw conditions

The penetration of environmental chloride ions into concrete and their diffusion to reinforcing steel surface, inducing corrosion, is a common mode of durability failure, which can lead to structural failure in severe cases. Existing calculation methods struggle to account for multi-factor coupling and non-steady-state diffusion of chloride ions. Accurately ascertaining the chloride diffusion coefficient in concrete is therefore crucial for evaluating structural durability and predicting residual service life. This study first identifies key evaluation indicators of chloride ion diffusion through a comprehensive literature review. A dataset comprising 275 sets of experimental results was compiled from published studies, covering various concrete mix proportions, stress ratios, and freeze-thaw cycles. Six machine learning algorithms—decision tree (DT), random forest (RF), gradient boosting decision tree (GBDT), extreme gradient boosting (XGBoost), neural network (NN), and support vector regression (SVR)—served for constructing predictive models for chloride ion diffusion. Different parameters were subjected to feature importance analysis. Model evaluation results indicate that when concrete is subjected to loading, the GBDT model exhibits the highest prediction accuracy, achieving a coefficient of determination R² = 0.9556. Under freeze-thaw conditions, the XGBoost model demonstrates the strongest generalization performance, with R² = 0.9514, identifying the number of freeze-thaw cycles as the most significant factor influencing the chloride diffusion coefficient (D₀). Finally, the predictive models were validated through experimental testing, yielding errors of 9.16 % and 11.83 %, respectively.