A neural network approach to elevated temperature creep–fatigue life prediction

Abstract

A new approach using a back-propagation neural network for life prediction was developed and demonstrated for predicting the elevated temperature (0.7–0.8 T m) creep–fatigue behavior of Ni-base alloy INCONEL 690. The neural network was trained with five extrinsic parameters, characterized via a 2 5–1 fractional factorial design methodology, and an intrinsic parameter (final grain size). The back-propagation network training error, prediction error and training time were minimized using a second fractional factorial design. Life prediction accuracy using only 11 training sets, few training iterations (<20,000) and a simple network architecture (6-2-2-1) showed significant improvement, for sets not previously used for training, when compared to Coffin–Manson, linear life fraction and hysteresis energy prediction techniques.