Co-authors​ :

 Mahoor MEHDIKHANI (BELGIUM), Rui GUO (BELGIUM), Sina AHMADVASHAGHBASH (BELGIUM), Diego CELENTANO , Yentl SWOLFS (BELGIUM) 

The initiation and propagation of transverse cracking in composites are governed by factors such as defects, arrangement of fibers, and matrix and fiber-matrix interface properties. Microscale computational models have been developed to simulate transverse cracking. The key challenges for such models lie in the 1) development of physics-based material models for the fiber, matrix, and interface, and 2) generation of representative volume elements (RVEs) of fibers embedded in the matrix. Fibers are often simplified to linear (transversely) isotropic materials. For the matrix, a few different plasticity and damage models have been proposed, e.g. [1,2]. A vital challenge for these material models is the correct identification of the input parameters since those measured via macroscale testing do not always represent the behavior of the material at the microscale.
To generate random fiber distributions, a few tools have been proposed, among which the one by Melro et al. [3] is commonly used. Despite their high efficiency, these types of generators are criticized because of not representing features like resin-rich pockets. In the current study, we have exploited deep learning to generate realistic images of transverse microstructures of unidirectional composites, yielding representative fiber distributions. Using these realistic microstructures, finite element (FE) models for the simulation of transverse cracking are created.
A constitutive material model, developed by Morelle et al. [2], accounting for epoxy plasticity and damage, is implemented for the matrix. This model accounts for yielding based on the linear Drucker-Prager criterion and strain hardening that includes pre-peak hardening, softening, and rehardening. Moreover, failure is included based on a triaxiality-dependent critical equivalent plastic strain. Input parameters for the pre-failure phase are taken from compression tests. For failure criterion parameters, tension, compression, and shear tests are considered. Applying periodic boundary conditions, the implemented material model for epoxy can successfully predict the formation and propagation of transverse cracks.
To evaluate the effect of fiber arrangement on transverse cracking, microstructures based on the Melro random generator [3] as well as the real computed tomography images are created and modeled with FE. Analysis of different RVEs demonstrates that the predicted transverse cracking behavior is similar in the deep-learning and real-image based models, and different from that in the random-generated one. The average strength and the standard deviation in the random-generated model differ slightly from those in the other two models. A quantitative approach to compare the crack propagation path between the three microstructures is being developed (Fig. 1). This study can still be enriched with the inclusion of interface debonding as well as other loading scenarios.