Method for improving parametric creep rupture life of 2·25Cr–1Mo steel using artificial neural networks

Abstract

A number of well known parametric models and a multilayer neural network are subjected to an ‘acid test’ of extrapolation to determine whether the latter can produce improved long term rupture life predictions for 2·25Cr–1Mo steel. Linear and non- linear least squares analysis was used to estimate the parametric models and genetic algorithms were used to identify and train the network. All the parametric models produced lifetime predictions for stresses below 60 MPa that were in error by some 20–40% on average. This reflected their tendency to overfit the data sets used for their estimation. Standard statistical measures of model inadequacy were of little use in overcoming this problem and the more non-linear models (e.g. Manson–Haferd) produced implausible extrapolations. In contrast, the optimised neural network was able to identify general patterns in the training data that were useful for extrapolation purposes and this was reflected in an average error of some 4–5%.