Degradation mitigation control of fuel cells with physics-informed machine learning – DEFINING

WP1 will develop a control-oriented multiscale fuel cell model using physics-informed neural networks.
WP2 will observe nano/microscale material degradation in real-time using Bayesian approach.
WP3 will optimize fuel cell operating parameters using model predictive control.

In this Tremplin project, we will take actions to overcome the weakpoints raised by reviewers for both project and for PI.
To improve the project, I will make efforts to 1) deepen Objective #3 (related to control) and explain this part more clearly and specifically; 2) put more evidence to demonstrate the advantages of the PINN approach in modelling the real situations.
To improve the PI, I will 1) publish more recent independent works, as the leading author and without involving PhD supervisors, to demonstrate the independent thinking capability; 2) complement my experiences and expertise in fuel cell physical modelling by conducting and publishing relevant works.

Complete the proof-of-concept study by integrating the control aspect.
With the support of a PhD student, I have conducted a case study demonstrating how the control strategy can be designed based on a Physics-Informed Neural Network (PINN) model. This result will be included as part of the proof-of-concept study in the next proposal.
In the same study, we also showed that PINN models can outperform purely physical models when the underlying physical phenomena are not fully understood.

In this ERC Tremplin project and my ongoing ANR JCJC project, I carried out a physical modelling study of a fuel cell, focusing on two-phase water flow, to demonstrate my expertise in this area. The software package developed has been published here:
doi.org/10.1016/j.softx.2024.102002

Comparative degradation experiments intended to demonstrate the potential of the control strategy are still ongoing. We encountered technical issues during this task: the pressure regulation system of the dedicated test bench broke down.
To address this, we ordered alternative fuel cell systems to carry out the tests without relying on the original bench. However, we needed to adapt these new systems to make them compatible with long-duration testing. We expect to begin these experiments in June 2025.

4. In addition to fuel cells, I have set up a test bench for electrolyzers, allowing for similar experimental investigations.
This action was not included in my previous ERC proposal, but it will be integrated into the upcoming one.

1. Comparative degradation tests to show the potential of the proposed methodology in both fuel cell and electrolyzer applications.

2. Proposal writing.

1. About fuel cell multiphysics modeling: doi.org/10.1016/j.softx.2024.102002.
2. About fuel cell degradation modeling: doi.org/10.1016/j.jpowsour.2024.235628
3. About PINN application in electrolyzer modeling: doi.org/10.1016/j.egyai.2025.100474

Hydrogen fuel cells are playing a more and more important role in different energy sectors, whereas fuel cell durability still needs to improve. Controlling operating parameters of fuel cells to mitigate the degradation rate is recognized as a crucial solution. Implementing this control needs to understand how nano/microscale material degradations impact the macroscopic performance, and how global operating parameters cause the material degradations. Multiscale modelling has been identified as an effective tool to achieve the linkage. But adopting multiscale degradation models for degradation mitigation control has been difficult so far. Because physical multiscale models are computationally inefficient, and hard to connect with data. These drawbacks further hinder accessing the nano/microscale material degradation situation and optimizing macroscopic operating parameters in real-time, which are the key elements of degradation mitigation control. DEFINING will overcome the drawbacks by using physics-informed machine learning approach and achieve the key elements of degradation mitigation control by realizing three objectives:
1. DEVELOP We will develop a control-oriented multiscale model using physics-informed neural networks.
2. OBSERVE We will observe nano/microscale material degradation in real-time using Bayesian approach.
3. OPTIMIZE We will optimize macroscopic fuel cell operating parameters using model predictive control.
DEFINING will empower hydrogen fuel cell technology with a degradation mitigation control strategy which will significantly prolong fuel cells’ lifetime. The other research domains of fuel cells will benefit from the unified degradation analysis tools offered by the project. DEFINING will also contribute by providing common methodologies for the analysis, observation, and control of multiscale processes to electrochemical, product quality control, and biomedical fields.