Hierarchical ensemble machine learning for corrosion monitoring in reinforced concrete subjected to carbonation and cast-in-chlorides

Corrosion resulting from the synergistic effects of carbonation and cast-in chlorides remains poorly understood. This study proposed a Quantitative Non-destructive Evaluation (QNDE) framework that integrates multiple non-destructive electrochemical measurements with physicochemical parameters such as pH, electrical resistivity, chloride concentration, half-cell potential, carbonation depth, and exposure time to predict corrosion current density under the synergistic action of carbonation and cast-in chlorides. Accelerated corrosion experiments were conducted to create a comprehensive dataset covering. Corrosion current density was treated as a corrosion indicator and predicted using parameters in the dataset. Three basic ML models, Ridge Regression, Random Forest, and Multi-Layer Perceptron were developed. Hierarchical ensemble techniques, such as the stacking ensemble with a Random Forest meta-learner achieved the highest predictive accuracy, with a R2 of 0.904 and negligible overfitting. The findings by SHAP analysis indicate that electrical resistivity and half-cell potential are the primary NDE parameters in corrosion prediction, together constituting about 85–87% of the model’s predictive power. Bootstrap-based uncertainty quantification further confirms that the Stacking RF model produces well-calibrated prediction intervals, with empirical 90% coverage consistent with the nominal level, supporting its use as a decision-support tool for early damage diagnosis and maintenance planning in reinforced concrete infrastructure.