Machine Learning for Permeability Determination of Fibrous Structures

The lightweight potential of fiber-reinforced plastics (FRP) plays a key role in reducing climate damaging CO₂ emissions, which is why their use in the field of renewable energies, in automotive and in aerospace industries is becoming increasingly important. The development of robust and efficient production processes is the main obstacle for industrial application of FRP. To overcome this, process simulations are typically used to support process development. Here, physical properties are attributed to finite elements that represent the processing behavior of fibers and polymers. However, the inherent complexity of fiber structures leads to major challenges in determining these properties, since physical effects occur at different temporal and spatial scales.

The project "ML4ProcessSimulation" addresses the key challenge of developing effective multiscale simulation methods that take into account all relevant physical effects from the fiber to the component level, but still remaining manageable in terms of computational effort. For this purpose, machine learning (ML) methods are applied synergistically with conventional simulation methods at all scale levels.

At the microscale, the lowest considered level in the multiscale workflow, data driven and geometry based ML methods are evaluated to determine permeability accurately with minimal effort. To train and evaluate the neural networks, GeoDict was used to generate over 2500 statistically representative FRP microstructures and to predict their permeability.

„ML4ProcessSimulation - Machine Learning for Simulation Intelligence in Composite Process Design” is funded by the Leibniz Association within the Leibniz Collaborative Excellence funding program (funding reference: K377/2021).