In the production of technical components made from long-fibre-reinforced plastics, compression moulding is used. In this process, a quasi-isotropic, fibre-reinforced preform is pressed into the target geometry under pressure. The resulting material flow causes fibre reorientation, which significantly influences the anisotropic mechanical properties of the part. To tailor these properties to the specific load case, the positioning and geometry of the inserted preform must be selected such that the desired fibre orientation is achieved. While the process can be represented through numerical simulation, such computations are time-consuming and often impractical for iterative optimisation.
This project explored the use of artificial neural networks (feedforward) as metamodels to provide reliable predictions based on limited simulation data. The focus lay on optimising model structure, selecting suitable input parameters, applying training strategies, and using synthetic data augmentation. A single-stage model — directly correlating insert position with deformation — was compared to a two-stage approach that additionally considered material flow fronts.
© IKVThe methodology developed makes it possible to create a metamodel based on the associated numerical simulation, independent of material and component geometry, and to use this for optimisation within the intervals of the training data set. Increasing the extrapolation capability is the subject of further research.
Project data and funding
We would like to thank the DFG for funding the project (funding reference HO 4776/74-1) and the project partners for their cooperation.
Project duration: 01.10.2022 – 31.03.2025
Promotion:
