Project
H2-Sim
Electrolyzers and fuel cells will be an essential part of the future technology. The development is supported in the one hand by the simulation of the complex behavior of the electrochemical processes using OpenFOAM, on the other hand by material characterization of components by means of machine learning (ML). The PI is teaching in the subject area synthetic fuels at the Bochum University, and he is leading the IET-4: Energy Process Engineering at the Forschungszentrum Jülich. His research field is in synthesis of fuels, supported by CFD simulations.
Based on previous work on fuel cells, machine learning methods are investigated on the prediction of macroscopic flow characteristics of porous materials, used in fuel cells and electrolyzers. It was necessary to apply for a PREP project (p022420) in order to re-calculate the required resources. As a result, p0020317 was finished in 2024 and is extended in the frame of p0022420.
Project Details
Project term
November 3, 2023–November 2, 2024
Affiliations
Ruhr University Bochum
Forschungszentrum Jülich
Institute
IET-4
Principal Investigator
Methods
OpenFOAM: the two-phase simulation of fuel cells was employed using the volume-of-fluid (VOF) method. The resolution of the interface between the two fluids requires a fines resolution along the interface between the two phases. This was implemented using adaptive mesh refinement.
Machine learning: the underlying data was generated by Lattice Boltzmann (LB) simulations of gas flow in porous micro-structures. Based on the simulated data, a convolutional neural network (CNN) was developed in order to predict characteristics of gas diffusion layers (GDLs) of fuel cells or porous transport layers (PTLs) of electrolyzers.
Results
OpenFOAM simulations: The VOF simulations were run on different scenarios. Channel simulations were investigated to study the detachment of droplets emerging from random positions of the GDL at the cathode side of a fuel cell. On lower scales the cracks in the microporous layer (MPL), attached to the gas diffusion layer (GDL) were systematically studied.
Machine learning: While the LB simulations require HPC resources, as well as the training of the CNN do, the prediction of the permeability of a porous micro-structures can be applied on a standard computer – given a trained CNN. An application was former shown on a laptop at the ‘Tag der Neugier’ in 2022. A CNN for porous transport layers (PTL) was trained with artificial data from a shere model, then the trained CNN was applied to real micro-structures that were formerly analyzed experimentally. The prediction of the permeability as a material property was successful on some of the real data. The data as well as the code for the geometry model is available at JULICH-DATA, the results were published, see below.
Discussion
The cracks in the MPL need to be adapted to real ones taken from CT images of a real GDL. In future simulations the systematic investigations will be applied to real micro-structures.
The CNN for the prediction of material properties was shown to be sensitive to hidden features of the manufacturing process of the material [2].
Additional Project Information
DFG classification: 404-01 Energy Process Engineering
Software: TensorFlow, LAMMPS, OpenFOAM, Palabos
Cluster: CLAIX
Publications
Dieter Froning, Eugen Hoppe, Ralf Peters. Flow characteristics of sintered titanium-based porous transport layers using machine learning, Discover Mechanical Engineering (2025) 4:2, https://doi.org/10.1007/s44245-025-00087-6