Neural network approach to estimate stable crack growth in welded specimens

This paper describes an application of the hierarchical neural network to the generation phase stable crack growth analysis of two kinds of welded CT specimens using the GE/EPRI simplified method. One of the specimens was machined from a submerged-arc-welded plate of nuclear pressure vessel A533B Class 1 steel, the other from an electron-beam-welded plate of A533B Class 1 steel and high-strength HT80 steel. A ratio of mixture of material constants was introduced to apply the GE/EPRI method to the analysis of crack growth in the welded specimens. The best ratio of mixture was identified using the neural-network-based inverse analysis approach as follows. At first, a number of generation phase crack growth analyses based on the GE/EPRI method were tested by parametrically varying the ratio of mixture. The relationship between the ratio of mixture and the calculated crack growth behavior is called here ‘learning data sets’. The neural network was then ‘trained’ using the learning data sets. In the training process, the calculated crack growth behavior is applied to the input units of the network, while the ratio of mixture is applied to its output units in the form of teaching data. Finally, the best ratio of mixture was estimated by applying measured crack growth behavior to the input units of the ‘trained network’. The effects of material inhomogeneity on crack growth behavior in the welded specimens are discussed with respect to the best ratio of mixture obtained.