Abstract
The back-propagation neural networks are becoming very popular for approximating functional formulas. However, there are a lot of trials and errors to develop the networks. In this paper, some effective techniques for approximating a material constitutive equation with neural networks, for example, (a)setting of the initial value of network, (b)determination of training parameters, (c)determination of the network architecture, and so on, are discussed. Since the outputs of the trained network are used directly in numerical simulations, it is important to show the development process clearly and to improve the reliability.
