Crack path predictions in heterogeneous media by machine learning

The interaction between stress fields at the crack-tip and nearby microstructural heterogeneities influences the crack path, and in turn, the effective fracture toughness of a material. In this paper, we assess the ability of artificial neural networks (ANNs) for machine learning to predict the crack paths from given initial void defect distributions in the material, and to provide insights into the underlying crack growth mechanics. Our ANN accurately captures the process zone size, crack growth sequence, and resulting crack patterns in the simplistic case where the proximity of voids to the crack-tip forms the criterion for crack advance. In a ductile medium, pre-existing voids grow with deformation and link up with the primary crack either contiguously or through the formation of multiple unconnected damage zones. The complex crack patterns for both these ductile fracture processes are successfully captured by an ANN, trained on the cracking sequences in a micromechanics-based ductile fracture model containing two size-scales of voids. In addition, the ANN architecture is capable of predicting stochastic crack growth, by providing a multiplicity of possible crack paths, along with a quantified likelihood of each path. Results further demon-strate the utility of autonomous crack path predictions in enabling fracture-by-design.

Automated summary

In this paper, the ability of artificial neural networks (ANNs) for machine learning to predict crack paths and provide insights into crack growth mechanics is assessed. The results show that ANN accurately captures the process zone size, crack growth sequence, and resulting crack patterns. Furthermore, ANN is capable of predicting stochastic crack growth and providing a quantified likelihood of each path.