The application of numerical simulations to optimise manufacturing processes, such as deep drawing, is currently state of the art. However, during such processes the behaviour of materials may change due to the evolution of damage. To create an appropriate estimation of material behaviour, damage evolution has to be coupled with constitutive relationships. At the current state, there are several damage models which are able to describe such damage evolution on a macro level. However, for a more accurate estimation it is crucial to take microstructural information, such as texture and grain size distribution, into account. A homogenisation from micro- to macro-schemes is computationally expensive when conducted with finite element methods. Hence, a new approach using a machine learning based framework is suggested, which can map damage from the micro- to the macro level.
In the scope of this work, numerical data based on finite element simulations is used to train and test the machine learning algorithm. This data is generated using a microstructurally informed synthetic representative volume element (RVE), which undergoes different loading states. The material model consists of phenomenological crystal plasticity as well as damage evolution based on equivalent plastic strain. Local quantities from these RVE simulations, such as stress, strain and damage, are homogenised into global averages. The global damage evolution is then predicted as a function of macroscopic parameters (e.g. equivalent strains, equivalent and hydrostatic stresses) and material properties (e.g Young's modulus and Poisson's ratio) by the trained machine learning algorithm. The results are compared to analytical, well accepted, damage models such as Chaboche or Lemaître for validation. Furthermore, different machine learning algorithms such as artificial neural networks, support vector machines and random forest are used and their results are compared with each other as well.