Improving accuracy of plural crack growth prediction by machine learning considering physical quantities limited small dataset

Abstract

This paper presents the improvement of the accuracy of plural crack propagation prediction by machine learning considering physical quantities limited small dataset. A dataset is obtained from the results of crack propagation analyses using s-version FEM combined with an automatic mesh generation technique. The input parameters are coordinates of the four crack tips. The output values to be predicted are crack propagation vectors and a number of crack propagation cycles of 0.25mm. Crack propagation paths and rates were predicted within 0.07 percent accuracy. This is a reason why independent multilayer perceptrons are separately configured in order to avoid an influence between different physical phenomena. Moreover, we tried to reduce the error by applying the data augmentation technique as a regularization. We observed the distribution of activations in the hidden layer to validate the generalization performance. We show that it is possible to predict with high accuracy even on small dataset with appropriate input and output parameters and appropriate configuration of training dataset.