Data-driven exact model order reduction for computational multiscale methods to predict high-cycle fatigue-damage in short-fiber reinforced plastics

Motivated by, amongst others, the development of more energy-efficient machines and means of transport, lightweight construction has gained enormous importance in recent years. One important class of lightweight materials comprises fiber-reinforced plastics. The present work focuses on the development of material models for the fatigue behavior of short glass-fiber reinforced thermoplastics. These materials differ from thermoset-based materials in their meltability and, thus, their better recyclability. Additionally, in contrast to long fibers, short glass-fibers allow for a simple and time-efficient production of complex components.
Fatigue is an important failure mechanism in these materials for components subjected to vibration-like loads, e.g., in the transport domain. However, the characterization and prediction of this failure mechanism are experimentally extremely time-consuming. Thus, fatigue assessment represents a significant challenge in the development process and for the broader application of short-fiber reinforced components. Therefore, the development of complementary simulative methods is of great interest.
In the present work, methods to predict fatigue damage of short-fiber reinforced materials are developed within the framework of a multiscale method. Multiscale models offer the possibility to predict complex, anisotropic effects of the composite material based solely on the experimental characterizations of the material parameters of the constituents, i.e., fiber and matrix. The experimental effort can thus be reduced significantly. For this purpose, first, material models for the constituents are developed on the microscale. Then, using FFT-based computational homogenization, the material behavior of the composite is predicted for different microstructures and load cases. The precomputed load cases at the microstructure level are transferred to the macroscale using data-driven methods. This enables efficient computations of engineering components, which would not be predictable by methods resolving the fiber structure on state of the art computers in years of computational time.
Various damage models for the matrix are investigated and advantages as well as disadvantages are analyzed. The microstructure simulations provide insight into the influence of various statistical parameters such as fiber length and fiber volume content on the composite behavior. A new model order reduction procedure is developed and successfully applied to the simulation of fatigue damage. Further, model extensions are developed to account for the stress ratio and viscoelastic effects in the evolution of fatigue damage. Both extensions are validated with experimental results. The resulting simulation framework allows the engineer to perform an efficient macrosimulation of the component after precomputations on a set of microstructures. Effects such as viscoelasticity and stress ratio dependence can be taken into account or excluded depending on the desired modeling depth in order to always use the simplest possible model that captures all relevant effects.