Modeling the load carrying capacity of corroded reinforced concrete compression bending members using explainable machine learning

Corrosion of reinforcement can lead to a decrease in the load carrying capacity of reinforced concrete structures and affect their safety. Therefore, accurate evaluation of the ultimate load carrying capacity is crucial for the maintenance and reinforcement of corroded reinforced concrete structures. In this paper, based on experimental research data of 192 corroded reinforced concrete compression bending members, data-driven analysis was conducted using ANN, SVM, RF, and AdaBoost algorithms to establish the relationship between the influencing factors of the load carrying capacity and their ultimate load carrying capacity. The input variables include the section width of the member, section height of the member, length of member, yield strength of reinforcement, diameter of longitudinal reinforcement, compressive strength of concrete, thickness of concrete cover, hoop diameter, original eccentricity, corrosion rate and the ultimate load carrying capacity is the output variable. Additionally, this study innovatively utilizes the Shapley additive explanations (SHAP) method to enhance the interpretability of the ML models, overcoming the "black box" issue associated with ML methods. Furthermore, the performance of the ML models is compared with theoretical formulas. The results indicate that the ML models exhibit good predictive performance, with higher accuracy than thetheoretical calculation formulas. And the predictive performance of ensemble learning models (RF, AdaBoost) is better than that of single learning models (ANN, SVM). The newly developed hybrid ML model is likely to become a new choice for dealing with the load carrying capacity problem of corroded reinforced concrete compression bending members.