MACHINE LEARNING APPROACHES FOR PREDICTING STEEL REINFORCEMENT-CONCRETE BOND STRENGTH: A SYSTEMATIC LITERATURE REVIEW

Abstract

The bond between steel reinforcement and concrete is fundamental to the structural performance of reinforced concrete elements, and accurate prediction of this bond strength remains a critical challenge in civil engineering. Traditional empirical models often fail to capture the complex, nonlinear interactions arising from material degradation, environmental exposure, and diverse reinforcement types. This systematic literature review aims to synthesize and critically evaluate the existing body of research on machine learning approaches applied to bond strength prediction. We conducted a structured search across major academic databases, followed by a rigorous screening and data extraction process to identify relevant studies. Our methodological framework focused on categorizing the literature according to key thematic dimensions, including algorithm types, corrosion effects, specialized concrete mixes, fiber-reinforced polymer (FRP) interfaces, and the integration of explainable artificial intelligence. The review reveals that ensemble methods and hybrid models generally achieve superior predictive accuracy compared to standalone algorithms, particularly when addressing datasets with high variability. Corrosion-induced degradation is increasingly modeled using neural networks that explicitly incorporate electrochemical parameters, while physics-guided approaches successfully embed mechanical priors to maintain physical consistency. Furthermore, we observe a growing trend toward interpreting model predictions through SHAP and partial dependence plots, which enhances the trustworthiness of these tools in structural engineering practice. The findings also highlight significant gaps in the literature, such as the limited validation of models under fire exposure and the underrepresentation of slip behavior predictions. We conclude that machine learning offers a powerful complement to conventional design equations, yet future work must prioritize standardized benchmark datasets, robust cross-validation protocols, and the development of models that generalize across diverse concrete compositions and environmental conditions.