Neural network-based surrogate model for a bifurcating structural fracture response

Abstract A finite element model of a tapered tensile specimen with a hardness transition zone in the gauge section and a varying width parameter is used for creating corresponding solution snapshots. Subsequently, a long short-term memory (LSTM) recurrent neural network (RNN) is trained on the selected snapshots, providing a parametrized solution model for a computationally efficient prediction of the structural response, allowing real-time model evaluation. In addition to a parametrized solution of the fracture localization, the model also captures the bifurcating local mesh deformation. The internal solution strategy of the RNN for predicting the bifurcation phenomenon is investigated and visualized.