InterPore2026

Residual-based PINN Modeling for Coupled Transport Phenomena in Porous Gas Diffusion Layers

20 May 2026, 11:50

15m

Oral Presentation (MS15) Machine Learning in Porous Media MS15

Speaker

Ms Hui Zhang (University of Bristol)

Description

Abstact:
The gas diffusion layer (GDL) of high-temperature proton exchange membrane fuel cells plays a critical role in regulating the coupled transport of species, heat, and charge. The simulation accuracy of these transport phenomena directly affects the predictive reliability of fuel cell performance. However, conventional computational fluid dynamics (CFD) simulations suffer from prohibitively high computational costs, while standard physics-informed neural networks (Pure PINNs) struggle to capture the complex field gradients within the GDL due to gradient vanishing in deep architectures. To address these challenges, this study proposes a residual physics-informed neural network (Res-PINN) framework for accurately modelling the multiphysics coupling processes within the hydrogen-side GDL. The proposed model embeds the governing equations of momentum, mass, and charge conservation directly into the loss function, thereby ensuring strict adherence to physical laws. To overcome the training limitations of deep Pure PINNs, a residual architecture with skip connections is introduced. By constructing identity-mapping pathways, this design effectively mitigates gradient vanishing during backpropagation and significantly enhances the network's ability to capture strong nonlinear gradient variations in porous media. The results indicate that the Res-PINN consistently outperforms the Pure PINN, achieving substantial improvements in predictive accuracy, with overall error levels reduced by 95% to 212% across different physical fields. In particular, the pressure field predictions exhibit near-perfect agreement with the reference solutions. In terms of computational efficiency, the proposed model achieves a 374-fold speedup compared with conventional CFD methods, reducing the inference time per evaluation from 1.0 s to 2.7 ms, whilst maintaining excellent generalization performance under previously unseen operating conditions. Overall, these findings demonstrate the superior capability of the Res-PINN architecture in handling complex multiphysics coupling problems. By preserving strong physical consistency while alleviating the training bottlenecks of deep PINNs, the proposed approach provides an efficient and reliable digital modeling tool for real-time simulation and engineering optimization of next-generation hydrogen-powered aircraft propulsion systems.

Keyword: Residual Physics-Informed Neural Network (Res-PINN), Gas Diffusion Layer (GDL), Multiphysics Coupling, Porous Media Simulation, Gradient Vanishing Mitigation