Multiscale modeling with artificial neural networks

The aim to adapt materials specifically to the requirements of load-bearing structures plays an essential role for an efficient and sustainable use of resources. The oftentimes resulting anisotropic material behavior can be captured e.g. with multiscale methods. Hereby, the micro- and macroscopic features are consistently coupled on two scales, which results in a homogenized constitutive relationship. In order to surpass the increased computational cost of the multiscale model, appropriate surrogate models, e.g. artificial neural networks, can be applied. In the first instance the constitutive behavior is learned on the basis of numerically generated data sets before the surrogate model can then be used within finite element simulations. The focus of research is hereby set on investigating the general feasibility of neural networks primarily for shell elements, especially considering numerical problems, as well as developing ways of modelling non-linear or time-dependent material behavior.

Contact Person
Jeremy Geiger