Artificial Intelligence accelerated multi-scale simulations of Composite Material Failure

Multi-scale simulations of composite materials are often unpractical for complex non-linear material responses, in particular because of their inherent computational cost. Data-driven methods allow to train surrogate models offline from a synthetic dataset built using offline simulations. Once trained, the surrogate can substitute the micro-scale BVP resolution of a multi-scale simulation reducing the online computation time by several orders of magnitude.
In particular, neural-networks (NNWs) offer the potential of reducing the computation time by more than 5 orders of magnitude [1]. However the extension to failure analyses requires accounting for the strain softening which poses challenges, on the one hand, in the definition of the meso-scale response that is RVE-size dependent and, on the other hand, on the handling of the loss of ellipticity at the macro-scale. These limitations are circumvented in this work by redefining a homogenized meso-scale non-local damage material law fed by NNW predictions of the underlying RVE response.
[1] L. Wu and L. Noels. "Self-consistency Reinforced minimal Gated Recurrent Unit for surrogate modeling of history-dependent non-linear problems: Application to history-dependent homogenized response of heterogeneous materials." Comput. Meth. in Appl. Mech. and Engng., Vol. 424 116881 (2024)