Abstract : Capturing sheet-metal localization physics with a trained surrogate inside a finite element (FEM) solver requires thesurrogate to deliver consistent predictions irrespective of the solver’s strain-increment count, which is difficult tocontrol in non-linear explicit codes. We address this by re-implementing a trained recurrent neural network (RNN)damage criterion in Fortran and embed it as a live, increment-by-increment fracture criterion inside an Abaqus/Ex-plicit user material subroutine (VUMAT), advancing the network state alongside the solver. Two architectures arecompared: a SimpleRNN and the proposed Consistent RNN (ConsRNN), whose transition function renders predic-tions invariant to the number of strain increments along a fixed deformation path. Both are trained on bilinear strainpaths and evaluated under varying temporal discretizations and nonlinear histories. SimpleRNN predictions driftas the increment count increases, which disqualifies it for embedding; ConsRNN maintains a stable response, at amodest accuracy cost relative to SimpleRNN on multilinear paths. Deployed at structural scale, the embedded Con-sRNN surrogate reproduces the global force–displacement response of a clamped steel plate to within 0.82% of peakforce relative to an established two-parameter fracture criterion. The results establish increment-count consistencyas the decisive property for embedding recurrent surrogates in explicit FEM solvers and demonstrate the first suchdeployment at structural scale.