What question did this study set out to answer?

April 1, 2026Open Access

Artificial Intelligence-Based Prediction of Chloride Migration Coefficient in Concrete with Enhanced Data Imputation, Optimization, and Explainability

Key Points

This research aims to enhance the prediction of chloride migration coefficient in concrete using machine learning techniques.
Developed an explainable machine learning framework for predicting chloride migration coefficients.
Evaluated seven machine learning algorithms with a focus on Gradient Boosting.
Utilized Bayesian optimization for efficient hyperparameter tuning.
Applied advanced data imputation techniques to improve data quality.
Conducted explainability analyses using SHAP and ICE methods.
Gradient Boosting model achieved a predictive performance with R² = 0.896.
Bayesian optimization reduced tuning time by up to 95% while maintaining accuracy.
Advanced data imputation techniques improved model stability and insights.
Key predictors identified include water content, concrete age, cement content, and superplasticizer dosage.

Abstract

Accurate quantification of the non-steady-state chloride migration coefficient (D nssm ) is vital for evaluating the durability and service life of concrete structures exposed to chloride environments. This study introduces an explainable machine learning (ML) framework for predicting D nssm , integrating robust data imputation, systematic hyperparameter optimization, and model explainability. Seven ML algorithms were evaluated, with a Gradient Boosting model optimized via grid search achieving the highest predictive performance (MAE = 3.302 × 10 -12 m 2 /s, RMSE = 8.519 × 10 -12 m 2 /s, R 2 = 0.896). Bayesian optimization delivered comparable performance with a 65–95% reduction in computation time, demonstrating its efficiency for scalable model calibration. Advanced imputation techniques preserved incomplete yet informative observations, enhancing data completeness and model generalizability without compromising predictive performance. Model explainability analysis using SHapley Additive exPlanations (SHAP) and Individual Conditional Expectation (ICE) revealed water content, concrete age, superplasticizer dosage, cement content, and cement type as the most influential features, collectively explaining ∼65% of the model’s variance. In addition, a user-friendly graphical interface was developed to support both single and batch predictions, enhancing accessibility for both research and engineering use. The proposed framework provides reasonably accurate and explainable data-informed tool for supporting chloride transport assessment and durability-oriented decision-making in concrete engineering. • Fine-tuned Gradient Boosting achieved R 2 = 0.896 in predicting concrete D nssm values. • Bayesian optimization reduced tuning time by up to 95% with comparable accuracy. • Advanced imputation strategies enhanced data utility, model stability, and insight. • SHAP analysis identified Water, Age, Cement, and Admixtures as key predictors. • SHAP–ICE methods exposed feature interactions critical to chloride resistance design.

Demander à l'IA

Bookmark

View Full Paper