An improved machine learning strategy for shear analysis in structural concrete based on the Newton’s method and the solvability region of the steel constitutive model

The Compression Field Theories (CFTs) can predict the full non-linear response of reinforced and prestressed concrete beams subjected to shear, considering the equilibrium and compatibility conditions in the cracked web of the beam, and the corresponding stress-strain relationships of the involved materials. These theories are supported by a set of non-linear algebraic governing equations whose numerical solution (which contains, among others, the inclination of the diagonal concrete struts, or crack angle) has been calculated through several strategies in the last years. Between them, machine learning methods arise as the best and most effective procedure for this aim, since, contrary to the traditional Newton-type methods, they do not require taking initial approximations to the numerical solution. Thus, in this work we present a new variant of the application of machine learning to the CFTs, based on Newton’s method as optimizer. Moreover, the training of this new strategy considered the existence of a solvability region for the steel constitutive model, as well as the location of the steel apparent yield strain in such region. In this sense, a classification sub-procedure about the location of such strain is implemented. As result, this new hybrid regression model, based on the Newton’s method and trained with previous classification, not only does not depend on initial approximations, but it predicts significantly better the experimental shear response of cracked reinforced and prestressed concrete beams than those machine learning strategies developed in previous works. In this work we propose an improved machine learning strategy for shear analysis in structural concrete based on the Newton’s method and the solvability region of the steel constitutive model