The applicability of neural network model to predict flow stress for carbon steels

A number of semi-empirical models are available in literature to predict flow stress of steel during hot deformation. In recent years, neural networks have also been used. Quantitative assessment of these models shows that the prediction errors range from 2 to 60% of the mean flow stress, when used over a range of strain rates (2-120 s(-1)), temperatures (900-1100 degreesC) and strains until 0.8. A neural network model, which can be used to predict flow stress for carbon steels, ranging from 0.03 to 0.34%C, is proposed. The network is able to simulate the flow stress behavior with an average error of 3.7% of the mean flow stress using strain, strain rate, temperature and carbon equivalent as inputs. The network is able to interpolate not only over the domain of strain rates and temperatures but also over the domain of carbon equivalents in which it is trained. (C) 2002 Elsevier B.V. All rights reserved.