Optimized regression-based machine learning models for predicting chloride diffusion in concrete

Accurately predicting chloride diffusion in concrete is critical for assessing the durability and service life of reinforced concrete structures exposed to aggressive environments. Traditional models, particularly those relying on Fick’s second law of diffusion, struggle to capture the intricate, time-dependent characteristics of chloride transport. To address these limitations, this study investigates the predictive capabilities of five regression-based machine learning models—K-Nearest Neighbors Regressor (KNNR), Decision Tree Regressor (DTR), Random Forest Regressor (RFR), AdaBoost Regressor (AdaBR), and LightGBM Regressor (LGBMR)—for estimating the chloride diffusion coefficient (CDC) in concrete. A comprehensive dataset, compiled from multiple experimental studies, was used to train and evaluate the models. Bayesian optimization via the Optuna framework was employed to systematically tune hyperparameters and enhance model performance. The findings demonstrate that ensemble learning techniques, especially boosting-based models, provide superior predictive performance compared to conventional regression approaches. LightGBM achieved the highest predictive accuracy, with an R² of 0.95 and the lowest RMSE of 0.83 × 10⁻¹² m²/s in the test phase. Feature importance analysis revealed that the water-to-binder ratio and binder content were the most influential factors governing chloride diffusion, while exposure time, fly ash percentage, and curing time exhibited relatively lower impact. These findings highlight the potential of machine learning as a powerful tool for chloride transport modeling, providing a data-driven approach to improve the durability assessment of concrete structures.