Leveraging the capabilities of physics-informed neural networks for channel optimization in PEM fuel cells

Physics-informed neural networks (PINNs) offer a promising alternative to computationally expensive CFD-based design studies of proton exchange membrane fuel cells (PEMFCs). However, their suitability for modeling realistic fuel cell channel flows and their ability to generalize across different geometric designs have not yet been systematically assessed. This work addresses this gap by developing a mesh-free PINN framework that embeds the governing transport equations of multicomponent channel flow, including oxygen consumption, directly into the learning process. Unlike conventional artificial neural networks (ANNs), the proposed PINN does not require labelled CFD training data. In this work, CFD data are used for benchmarking and validation, while the PINN training itself does not rely on CFD field solutions in the interior of the computational domain, and the PINN training itself is driven solely by the embedded governing equations without labelled CFD target fields. The model is evaluated with respect to two core research questions: (i) Can a PINN accurately reproduce the coupled flow and species transport fields in a PEMFC gas channel? and (ii) Can the learned physics be leveraged to predict performance for previously unseen channel widths? Results show that the PINN reproduces CFD reference fields with high accuracy and substantially outperforms ANNs in generalizing to channel widths not included in training. Moreover, the PINN-based design study achieves a computational time reduction of approximately 83.9% compared to full CFD simulations. These findings demonstrate that PINNs provide a data-efficient and computationally lightweight surrogate model suitable for accelerating iterative PEMFC channel design, here exemplarily demonstrated for variations in channel width.