Metamodels of FE Simulations

Deep learning to increase efficiency in finite element simulation

The finite element method (FEM) is an effective means of investigating the behaviour of assemblies over time. However, the high complexity of specific problems often leads to long calculation times for simulations. This disadvantage is particularly noticeable in transient calculations. Even the most advanced methods of model reduction do not sufficiently counteract this.

In addition to FEM models, a metamodel based on artificial intelligence (AI) can be set up using sub-modelling techniques. Using neural networks, sub-areas of the components to be analysed are mapped. Long-short-term memory (LSTM) networks are particularly suitable architectures for AI-based sub-modelling.

Particularly in the field of optimisation - e.g. topology optimisation - many calculations have to be carried out. If there are areas in the structures to be investigated that do not belong to the optimisation area but have an influence on the behaviour of the optimisation area, these areas must always be calculated numerically at great expense. Metamodels therefore offer advantages in terms of efficiency.