A strategy to formulate data-driven constitutive models from random multiaxial experiments.

We present a test technique and an accompanying computational framework to obtain data-driven, surrogate constitutive models that capture the response of isotropic, elastic-plastic materials loaded in-plane stress by combined normal and shear stresses. The surrogate models are based on feed-forward neural networks (NNs) predicting the evolution of state variables over arbitrary increments of strain. The feasibility of the approach is assessed by conducting virtual experiments, i.e. Finite Element (FE) simulations of the response of a hollow, cylindrical, thin-walled test specimen to random histories of imposed axial displacement and rotation. In these simulations, the specimen's material is modelled as an isotropic, rate-independent elastic-plastic solid obeying J2 plasticity with isotropic hardening. The virtual experiments allow assembling a training dataset for the surrogate models. The accuracy of two different surrogate models is evaluated by performing predictions of the response of the material to the application of random multiaxial strain histories. Both models are found to be effective and to have comparable accuracy.