Machine-Learning Prediction of Shear Capacity of RC Corbels Under Combined Vertical and Horizontal Loading

Abstract

The shear capacity of reinforced concrete corbel (RCC) remains difficult to predict because these members are disturbed (D-) regions governed by strut-and-tie action rather than by conventional theory, and because their strength is further modified when a horizontal tension force acts together with the vertical load. This study develops and compares data-driven models for the shear strength of RCCs using a compiled database of 53 experimental specimens under combined vertical shear and horizontal force, described by twelve geometric, material, and reinforcement parameters. Five algorithms: multiple linear regression (MLR), random forest (RF), gradient-boosted regression trees (GBRT), support vector regression (SVR), and artificial neural network (ANN) were trained and evaluated under a cross-validation protocol to obtain out of sample predictions. Support vector regression achieved the highest accuracy (R2 = 0.926, RMSE = 59.5 kN, mean absolute percentage error = 15.7%), then followed by gradient boosting (R2 = 0.913) and the neural network (R2 = 0.904), whereas the linear model and the random forest were less accurate. Taylor and radar diagrams confirmed the same ranking across multiple criteria. Feature-importance analysis identified the corbel width, the concrete strength, the shear-span-to-depth ratio, and the applied horizontal force as the dominant predictors, in agreement with established corbel mechanics. The results indicate that the shear strength capacity of RCCs can be predicted with lower scatter than a semi-empirical formulation adopted by standard codes, and therefore provide an interpretable basis for design capacity estimation.