A Hybrid Neural Network–Based Approach to Predict Crack Propagation Paths

Abstract

A data-driven method based on a hybrid neural network (HNet) model is proposed to predict the crack propagation path. Using images as input enables the HNet model to predict crack propagation paths for different structures and defect types. To validate the effectiveness of this method, crack propagation paths on holed plates are investigated. The HNet model is trained to approximate the nonlinear relationship between the structural geometric parameters and the crack propagation paths. The feasibility of this method is verified by comparing the prediction results of the HNet model with the finite element calculation results. Furthermore, explainable artificial intelligence enhances the transparency of the HNet model, increasing its credibility. The challenge of data acquisition is effectively addressed by active learning, reducing the required training data volume. This method provides a fresh insight into the path prediction of crack growth problems.