Neural networks for critical high temperature component life prediction

Abstract

The aim of the present research was to subject a well known parametric model and a neural network model to an acid test of extrapolation in order to determine which can produce improved long term creep rupture life predictions for 2·25Cr–1Mo steels. Linear, squared and cubic parametric models were used and the accuracy of the predictions assessed by calculating the mean percentage absolute error. Many different neural network geometries were developed and the accuracy of the predictions was assessed again by calculating the mean percentage absolute error. As the predictions are concerned with the long term rupture life of components, the accuracy below 60 MPa is of the greatest importance. A neural network model with a 2–5–5–1 architecture provided the lowest error for predictions below 60 MPa, compared to other neural networks and the parametric models, and is therefore the optimum model from this study for predictions of long term rupture creep life for a 2·25Cr–1Mo steel.