Evaluation of Crack Front Geometry by Strain Field Classification Using Neural Networks

Determining a relation between the crack length and the compliance in compact tension C(T) specimens during experiments is a topic of great interest, which has been addressed in several researches with 2D finite element (FE) analyses assuming a straight crack front. Relevant standards for the evaluation of the fatigue crack growth rate and fracture toughness of metallic materials prescribe limitations on the crack front curvature. This is because different curvatures in the crack front result in different stress-strain fields, making the prediction of the crack length a problem depending on the crack front shape too. Additionally, observation of the crack front during tests or operation of mechanical components is challenging. For all these reasons, this work demonstrates a methodology that aims to simultaneously predict the crack front curvature and track the crack growth in numerical simulations of a standard specimen. This objective is addressed by investigating two different strategies. A first approach aimed to leverage local strain measurements; however, analysis of data revealed that this strategy was not feasible because local measurements were not representative of scenarios with different crack front geometries. In a second approach, the framework is setup as a classification problem and it consists of the combined employment of the full field strain measurement on the lateral face of C(T) specimens and of a convolutional neural network that provides the classification. A dataset is generated by means of parametric 3D FE simulations reproducing different crack front curvatures and lengths for a selected geometry. Results indicate that the crack front geometry is correctly predicted with an average accuracy above 90%, and that the normalized crack length is predicted with RMSE of 0.02.