Hero Default

Methods & Software

JL_VMD_Methoden

Pic: Hereon/N. Huber

Several software tools are being developed at Hereon for the JL VMD. The non-commercial highly efficient and scalable finite element code 4C is developed together with the TU Munich (TUM) and the Universität der Bundeswehr München. To ensure exascale computing abilities, the multiphysics code – based on its predecessor code BACI - is relying on the deal.ii librariy. In this context, Hereon and TUM are regularly providing to the world wide open source community novel features developed for 4C as part of new releases of the deal.II library. With 4C it will be possible in the future to solve problems in solid mechanics, fluid mechanics, corrosion. Likewise, coupled electromechanical and diffusive problems can be solved highly efficiently on different discretizations. Great emphasis is placed on scalability, which has already been successfully tested for nonlinear plasticity on up to 1000 CPUs.
Codes, which are based on the numerical implementation of physical equations, are complemented by the development of machine learning methods. There, the focus lies on self-optimizing architectures, the blending of simulation data with experimental data, and the hybridization of data-driven approaches with physics-informed architectures.

graphical abstract cropped

Pic: Hereon/E. Schiessler

We developed a framework for a genetic neural architecture search algorithm that combines optimization of the network architecture with training the network itself. Based on mathematical and statistical criteria, modified network versions are created and then compete against each other. Winning modifications are selected from the pool of generated candidates and receive further training. This process is repeated until a fully developed fully trained final network emerges.
Source: Elisabeth J. Schiessler, Roland C. Aydin et al., Neural network surgery: Combining training with topology optimization, Neural Networks, 2021 Publication