ML4ProcessSimulation

Multi-scale approaches are used to reduce the computational effort for the simulation of fiber-reinforced polymer composites (FRP). For example, impregnation is simulated separately at the fiber, textile and component levels, and data is exchanged between the scale levels. As with any simulation, various simplifications are made, such as neglecting capillary forces and homogenizing local properties when transferring to the higher level. In the “ML4Process-Simulation” project, the integration of machine learning methods is intended to increase the accuracy and efficiency of the simulation workflows. So far, numerical simulation methods (including finite volume and Lattice-Boltzmann methods) have been investigated for the efficient generation of training data. Permeability prediction could achieve comparable accuracies to experiment at the fiber level using neural networks (NN). In addition to feed forward NNs that require characteristic values as input, 3D convolutional NNs that use geometry models as input were investigated. These NNs can be integrated to efficiently predict local micropermeability at the textile level. Further research focused on the integration of physical laws in model training (physics-informed neural network) and transient simulations of multiphase flows, which allow the integration of further physical effects. This method can be used to estimate air pockets in the fiber structure, which typically represent mechanical weak points in the FRP.

The goal of the project is to integrate machine learning methods into multiscale simulation approaches to efficiently account for relevant physical effects for accurate simulation of resin injection processes.