The increasing awareness and urgency of climate change have led to an increase in investments and research into power sources not reliant on fossil fuels. Proton exchange membrane (PEM) fuel cells are a promising technology for automotive, maritime, and auxiliary power applications converting chemical energy into electricity. In order for this upcoming technology to compete with well-established alternatives such as diesel generators and combustion engines, it is of vital importance to improve the power density and durability.

PEM fuel cells produce water and heat during operation. The presence of superfluous liquid water in the fuel cell stack gives rise to flooding of the electrodes, which hampers operation and induces degradation processes. On the other hand, it is of great importance to maintain a high membrane humidification to reduce Ohmic losses over the membranes and avoid the formation of cracks. Therefore, water management plays a vital role both in maintaining the power density and guaranteeing durability.

This thesis is written under the auspices of both TU Delft and PowerCell Group, a manufacturer of PEM fuel cell systems located in Gothenburg, Sweden. Currently, a substantial amount of water condenses in the anode subsystem of PowerCell’s PEM fuel cells in certain operating ranges which subsequently enters the fuel cell stack. This study identifies the influence of certain system operating parameters on responses as the water crossover through the membranes, the relative humidity at the stack inlet, the temperature at the inlet of the stack, the condensation rate in the mixing chamber of the recirculation loop, and the mass flow rates of liquid water and water vapor in and out of the stack. Multiphysics simulation software provided by Gamma technologies is used to simulate 5600 operating points in which the system operating parameters are varied according to the Latin Hypercube sampling method. These simulations give a clear overview of the influence of the operating parameters on the aforementioned responses over the entire operating range of PowerCell’s PS-100 system.

The simulated experiments are subsequently used as a basis to construct metamodels. These metamodels predict the behavior of the system based on the operating parameters varied in the 5600 simulations. The metamodels are constructed both as Krigings and as multilayer perceptrons (MLPs). Kriging is a statistical method which produces an output for a certain response based on known input data where input points which resemble the unknown point are given a greater weight. MLPs are neural networks which recognise patterns in input data and use those to predict an output for a certain response. Finally, the operating parameters are optimized using the metamodels to minimize the liquid water mass flow rate into the stack, to prevent condensation in the mixing chamber of the anode subsystem, and to target a certain inlet relative humidity of the hydrogen feed to the stack. Concluding from these optimizations it would be beneficial to preheat the hydrogen before it reaches the mixing chamber. A number of alternative designs for the anode loop are proposed which require further investigation.