Prediction of the Water Distribution in Randomly Packed Hollow Fibre Humidifiers Using Artificial Neural Networks as Surrogate Models

Hollow fibre membrane humidifiers are typically applied in proton exchange membrane (PEM) fuel cell systems to guarantee a sufficient level of humidification of the fuel cell membrane. This study investigates the effect of the fibre placement on the water distribution in hollow fibre membrane humidifiers using surrogate models derived from detailed CFD simulations. In the literature, it is demonstrated that CFD models and simulations can accurately predict the water transfer in membrane humidifiers. However, the extensive use of detailed CFD simulations is often restricted by the high amount of computational resources required for such studies. To overcome this limitation, we propose to use a U-Net, a special type of artificial neural network (ANN), that acts as a surrogate model to predict the water distribution in randomly packed hollow fibre bundles. Furthermore, a grid search is done to identify suitable hyperparameters for the U-Net. Our results indicate that a U-Net can accurately predict the water distribution in different hollow fibre humidifiers.