Machine Learning-Driven Optimization of Gas Diffusion Layer Microstructure for PEM Fuel Cells

Proton exchange membrane fuel cells (PEMFCs) have long been studied as a clean energy technology, known for their high efficiency, low operating temperatures, and zero carbon emissions.¹ As the electric vehicle market grows, PEMFCs are increasingly viewed as a complementary solution to battery electric vehicles (BEVs), particularly for applications like heavy-duty trucks where BEVs face limitations².<br/>A key factor in PEMFC performance is the gas diffusion layer (GDL), which plays a critical role in mass and heat transport. Optimizing the GDL's microstructure can significantly improve these transport properties, leading to more efficient and durable fuel cells.<br/>In the present study, we developed a novel machine learning approach to optimize GDL microstructure and its associated functional properties. Our approach starts with a design of virtual experiments (DoVE) to systematically investigate the relationship between GDL microstructure and its functional properties. The DoVE considered five manufacturing parameters that are used to stochastically generate a large amount of GDL microstructures: fiber diameter, fiber volume fraction, binder volume fraction, GDL thickness, and compression ratio. Seven functional properties are then evaluated: through-plane and in-plane thermal conductivity, electrical conductivity, diffusivity, and contact resistance between the GDL and the microporous layer. Latin hypercube sampling was employed to achieve an efficient exploration of the hyperparameter space and generate a comprehensive well distributed set of input parameters. A Random Forest + Adaboost regression model was then employed to learn the complex relationships between the five input and seven output parameters. After training the regressor, a multi-objective optimization was performed for the following 4 cases: : 1. minimum contact resistance, maximum TP, IP diffusivity; 2. minimum contact resistance, maximum TP, IP electrical conductivity; 3. minimum contact resistance, maximum TP, IP thermal conductivity; and 4. minimum contact resistance, maximum TP, IP electrical and thermal conductivities and diffusivity. This work has taken inspiration from our previous study on machine learning-assisted optimization of electrode manufacturing of lithium-ion batteries.³<br/>The implemented optimization framework, the first of its kind as far as we know, exhibited high efficacy in identifying optimal manufacturing parameters for GDL microstructures, and their associated functional properties. The validation of our machine learning approach was carried out by comparing the predicted GDL functional properties to the ones calculated through physics-based simulations using the optimal manufacturing parameters predicted by the optimizer. This optimization strategy holds promise for enhancing gas transport, water management, efficient current collection, and thermal regulation within PEMFCs.⁴<br/> <br/><b>REFERENCES</b><br/>(1) Mench, M. M. <i>Fuel Cell Engines</i>, 1st ed.; Wiley, 2008. https://doi.org/10.1002/9780470209769.<br/>(2) Cunanan, C.; Tran, M.-K.; Lee, Y.; Kwok, S.; Leung, V.; Fowler, M. A Review of Heavy-Duty Vehicle Powertrain Technologies: Diesel Engine Vehicles, Battery Electric Vehicles, and Hydrogen Fuel Cell Electric Vehicles. <i>Clean Technologies</i> <b>2021</b>, <i>3</i> (2), 474–489. https://doi.org/10.3390/cleantechnol3020028.<br/>(3) Duquesnoy, M.; Liu, C.; Dominguez, D. Z.; Kumar, V.; Ayerbe, E.; Franco, A. A. Machine Learning-Assisted Multi-Objective Optimization of Battery Manufacturing from Synthetic Data Generated by Physics-Based Simulations. <i>Energy Storage Materials</i> <b>2023</b>, <i>56</i>, 50–61. https://doi.org/10.1016/j.ensm.2022.12.040.<br/>(4) Omongos, RL.; Galvez-Aranda, D. E.; Zanotto, F. M.; Vernes, A.; Franco, A. A. Machine Learning-Driven Optimization of Gas Diffusion Layer Microstructure for Enhanced PEM Fuel Cell Performance. (to be submitted, 2024)