Embedding Consistent Recurrent Neural Networks for Fracture Modelling in Large Scale Shell Structures

Data-driven surrogates could replace phenomenological fracture criteria in structural-scale shell element simulations, but a recurrent surrogate can operate inside an explicit solver only if its predictions are independent of the increment count. This condition was tested by embedding a trained recurrent neural network (RNN) localisation surrogate as the erosion criterion in an Abaqus finite element material subroutine. A SimpleRNN was compared against the proposed consistent RNN (ConsRNN), whose transition function makes predictions converge with increasing increment count. Both were trained on 19,098 bilinear strain paths from solid-element simulations of one structural steel. At the 400-increment training resolution, the SimpleRNN was more accurate but drifted with the increment count, whereas the embedded ConsRNN’s prediction varied by only 1.5% between 400 and 200,000 increments. In a clamped steel plate indented to rupture, the force–displacement curves coincided to the peak for network sampling intervals differing by over an order of magnitude. The ConsRNN and an experimentally validated two-factor-scaling (2FS) criterion for the same steel predicted nearly identical fracture onset and peak force at indistinguishable computational cost. No benefit from the surrogate’s history dependence was demonstrated, because the crack-path elements were loaded nearly proportionally, and ranking the criteria requires a fracture experiment. The paper’s main contribution is the embedding of a customised recurrent cell with a non-standard transition function as a live fracture criterion in a structural-scale explicit FE analysis.