Accurate modelling of fiber reinforced polymers (FRP) plays an important role not only in part design but also for the development of robust and reliable manufacturing processes. However, the inherent structural complexity of engineering textiles and the presence of different physics at multiple scales lead to major challenges in numerical modelling of these processes. This is particularly evident in Liquid Composite Molding processes, widely used to manufacture e.g. wind turbine blades, where a dry fiber structure is impregnated with a liquid polymer resin system. During the infiltration, resin flow occurs within fiber bundles (flow channels in the range of μm) and between fiber bundles (flow channels in the range of mm). Resolving both spatial scales simultaneously requires very fine computational grids, which with current computational constraints becomes infeasible. Hence, numerical modelling of these phenomena is conventionally accomplished by using a finite element or finite volume scale separated approach.

Since their introduction in 2017, Physics Informed Neural Networks (PINNs) have been used for a wide range of problems, including flow in porous media, heat transfer, and thermochemical curing. PINNs incorporate physical laws in the form of partial differential equations (PDEs) into the training of neural networks (NNs). The meshless nature of PINNs offers potential advantages in dealing with complex geometries, such as fibrous microstructures. Moreover, PINNs can be trained rapidly on modern computational clusters, thanks to the excellent parallelization capabilities of neural networks. We investigate the use of PINNs for predicting the flow in fibrous microstructures using simplified models of porous media. Once the correct NN approximation scheme is found, the results are in good agreement with those provided by conventional solvers, with the advantage of significant computational leverage based on various deep learning techniques.