Abstract

A number of codes of practice exist that predict the maximum shear capacity of reinforced concrete beams. Since the behavior of reinforced concrete (RC) beams with non-homogeneous, non-isotropic, and nonlinear material under a combined shear and bending state of stress is very difficult to establish, these codes seem to under or over-estimate this shear capacity for many cases. This is attributed to the fact that the factors that affect the shear strength of RC beams are too many, making modeling of its actual behavior a hard task. In this paper, several multilayer perceptrons were constructed as an analytical alternative to existing expressions for predicting the shear capacity of RC beams. A large database of experimental tests of beams (574 samples) was utilized to train and test the networks. Both multilayer perceptrons' predictions and four different codes of practice for the shear capacity of RC beams were examined. It was found that, the predictions of multilayer perceptrons are superior to those of any of the current available code relationships.

Key words: Multilayer perceptron, neural networks, shear capacity, reinforced concrete (RC) beams, shear reinforcement.