Machine-learning-aided Shear-capacity Solution of RC Girders with Web Stirrups Based on the Modified Compression Field Theory

It is widely recognized that critical crack angle θ is a pre-requisite to calculate the shear capacity of RC elements in traditional modified compression field theory (MCFT), and it is often determined by an iterative calculation or by presupposing an empirical value. This study proposes a straightforward solution of critical crack angle aided by machine learning. Firstly, 215 reinforced concrete T-girder test samples are collected from published literatures, and are analyzed by traditional MCFT to determine their shear capacity and their corresponding critical compressive strain of the upper edge concrete ε_tom and the critical crack angle θ. Subsequently, integrated BP (back propagation) neural network models are established to seek for a quantitative regression between the iteratively obtained θ, ε_tom and other MCFT input parameters. Finally, the obtained regression equations are incorporated into traditional MCFT framework to determine the shear capacity straightforwardly. The results indicate that critical crack angle θ of reinforced concrete T-girder exponentially grows by increasing the strength eigenvalue ρ_v·f_yv/f_c or decreasing the longitudinal reinforcement ratio ρ_l. While the compressive strain of concrete in the compression region ε_tom exhibits a logarithmic function with the strength eigenvalue ρ_v·f_yv/f_c and the shear span ratio λ. The proposed straightforward calculation approach is superior to other methods both in efficiency and accuracy. Specifically, the goodness-of-fit of the proposed approach is 1.7-fold higher than that of the American ACI318-14, and the coefficient of variation is reduced by 43% compared to the European EN 1992-1-1;2004.