Residual-based physics-informed neural network modelling for coupled transport phenomena in porous gas diffusion layers

Proton exchange membrane fuel cells (PEMFCs) are a promising low-carbon power solution, with the gas diffusion layer (GDL) critically influencing mass, electrical potential, and charge transport. Accurate prediction of these coupled physical fields is essential for optimising fuel cell performance, yet conventional computational fluid dynamics (CFD) methods are computationally expensive, limiting their application in real-time control and rapid design iteration. In this study, a residual-based physics-informed neural network (RPM) was developed to predict multi-physical fields under steady state in the hydrogen-side GDL. By embedding governing equations into the loss function, RPM ensures physical consistency, achieves significantly lower PDE loss than standard PINNs (PPM), and demonstrates generalisation across both interpolation and extrapolation scenarios, outperforming PPM and black-box models (BBM). Mesh resolution analysis identified a medium mesh of 12,570 elements as optimal, balancing prediction accuracy, physical fidelity, and computational efficiency. RPM achieved single-case prediction times of 4.77 ms, approximately 1470 × faster than conventional CFD simulations. These results indicate that RPM provides a reliable and efficient tool for real-time, physically consistent GDL performance prediction.