Co-authors​ :

 Lu KE (NETHERLANDS), Iuri ROCHA , Marina MAIA , Frans VAN DER MEER (NETHERLANDS) 

Fiber reinforced composite laminates are inherently multiscale materials. There is a microstructure of fibers in a matrix and a mesostructure of plies that form the laminate. Models for nonlinear physical processes, such as plastic deformations and environmental aging, are more easily described on the microscale than on the mesoscale. The microscale offers the possibility to combine relatively simple models for the constituents with a geometric representation of the microstructure describe relatively complex behavior for the composite material. In this work, we are interested in analyzing the fracture process zone around a growing crack, employing a multiscale model to shed light on how different microscopic nonlinear processes such as diffuse microcracking and matrix plasticity contribute to the fracture energy for delamination in different modes of fracture.

One way to include micromechanics in mesoscale analysis is through computational homogenization, or FE2. However, the computational cost of FE2 is enormous, limiting its practical applicability. For that reason, we pursue an approach where the micromodel is replaced with data-driven surrogates. The vision is that data-driven surrogates can be richer than a priori formulated mesoscale constitutive models, but more efficient than the micromodel. Given sufficient micromechanical training data these surrogates can be as expressive as the original micromodel, but they can be much faster to evaluate. In this contribution, we present data-driven multiscale analysis of delamination crack growth in mixed-mode bending (MMB) tests. A micromodel with random fiber distribution is surrogated with machine-learning techniques and linked to a mesomodel of the MMB setup.

Specifically, mesoscale crack growth is modeled with cohesive XFEM, while the micromodel includes matrix plasticity, matrix cracking and fiber/matrix debonding. Two different surrogates are used in a single framework. For the cohesive tractions and the elements ahead of the crack, which all undergo relatively similar local histories, an active learning framework with Gaussian Processes is used, where a micromodel is evaluated on the fly to generate additional data if the surrogate is probed too far from the range it has been trained for. For the surrounding bulk material where a larger range of different strain histories is seen during the analysis, but without softening in the homogenized stress-strain response, a pre-trained data-driven surrogate is used. This surrogate is a neural network with classical constitutive models included in its hidden layer for improved generalizability. The proposed approach enables simulation of delamination crack growth without making constitutive assumptions on the macroscale. The influence of different physical processes on energy dissipation is investigated by varying which ingredients are included in the micromodel.