MACHINE LEARNING APPROACHES FOR SHEAR STRENGTH PREDICTION OF REINFORCED CONCRETE BEAMS: A SYSTEMATIC LITERATURE REVIEW

Abstract

Shear failure in reinforced concrete beams is a critical and often brittle mode of failure that poses significant challenges for structural design and assessment. Traditional empirical and semi-empirical equations, while widely used, frequently exhibit limited accuracy across diverse design parameters and material types. This systematic literature review was therefore conducted to comprehensively examine how machine learning approaches have been applied to predict the shear strength of reinforced concrete beams. Our objective was to synthesize the state of the art across four distinct dimensions: standard reinforced concrete beams, beams reinforced or strengthened with fiber-reinforced polymers, fiber-reinforced concrete and advanced cementitious composites, and special structural configurations including degradation effects such as corrosion or fatigue. The methodology involved a structured and replicable process of identifying, screening, and critically appraising relevant studies from the past two decades. We then extracted and analyzed data on model architectures, input features, training strategies, and reported performance metrics. The results reveal that ensemble methods, particularly gradient boosting and random forests, have consistently outperformed single models such as artificial neural networks in terms of predictive accuracy and generalization across all four dimensions. Furthermore, the inclusion of parameters that capture size effect and aggregate interlock was shown to substantially reduce prediction errors for standard beams, while for FRP-strengthened systems, models that integrated delamination-related features proved most effective. For fiber-reinforced composites, the choice of fiber type and volume fraction emerged as critical input variables, and for degraded beams, time-dependent corrosion parameters were essential. We conclude that machine learning offers a powerful and adaptable framework that can capture complex nonlinear interactions often missed by code-based formulas, yet the field still suffers from a lack of standardized benchmark datasets and consistent reporting of model uncertainty. This review thereby provides a roadmap for future research and practical deployment of these predictive tools.