1999 Volume 39 Issue 10 Pages 999-1005
The hot torsion stress-strain curves of steels have been modelled using a neural network, within a Bayesian framework. The analysis is based on an extensive database consisting of detailed chemical composition, temperature and strain rate from new hot torsion experiments. Non-linear functions are obtained, describing the variation of stress-strain curves with temperature and chemical composition. Predictions are associated with error bars, whose magnitude depends on their position in the input space. From the population of possible models, a "committee of models" is found to give the most reliable estimate. The results from the neural network model where found to be consistent with known models, and reasonable estimates are obtained beyond the scope of the experimental data.
