Investigation of Gas Diffusion Layer Intrusion in PEM Fuel Cell Using Physics-informed Machine Learning
Skip to main content
eScholarship
Open Access Publications from the University of California

UC Irvine

UC Irvine Electronic Theses and Dissertations bannerUC Irvine

Investigation of Gas Diffusion Layer Intrusion in PEM Fuel Cell Using Physics-informed Machine Learning

Abstract

This research investigated the gas diffusion layer (GDL) membrane intrusion into a gas flow channel (GFC) using physics-informed machine learning for PEM fuel cells. This study was done to reduce the time it takes to simulate the model and find the intrusion area. To establish the training data for machine learning, different configurations of the GFC and GDL were created using COMSOL to simulate the GDL intrusion into a gas flow channel. The data from these simulations then get exported into Design-Expert, a statistical software to determine the parameters with the highest impact on the % intrusion. The GDL's Young's Modulus (E_GDL) was found to be the parameter with the most significant impact, with an F-value of 3776.66 compared to the next parameter Bipolar Plate Channel Width (CW_BP) with an F-value of 992.59, or the lowest parameter GDL Height (H_GDL) with an F-value of 159.96. Design-Expert also shows that (E_GDL) and the Bipolar Plate Rib R_BP positively affect the \% intrusion area, where both will yield lower % intrusion the higher the parameters are. Whereas higher the values are (CW_BP), (H_GDL) and Pressure (P) will increase the % intrusion.

With the finding from the Two-Factorial test, the simulation interval for (E_GDL) is higher than the other parameters to yield more accurate training data for machine learning. Then performing, a parametric study will be done to find the x-y coordinates of the intrusion curve to export to MATLAB to find the intrusion area. Four machine-learning algorithms, linear regression, Decision Tree, SVR, and KNN, were deployed to train using 70% of the data set and the remaining 30% for testing. The accuracy of each model were calculated based on how close the prediction is compared to the actual value. It was found that out of the four algorithms, Linear Regression has the lowest model accuracy at 68.5% and the highest RMSE at 0.0751, and Decision Tree has the highest model accuracy at 95.5% and lowest RMSE at 0.0303. Thus, Decision Tree was used to make predictions for various ranges of the five parameters to find the optimization parameters for design. This study was done to help reduce the time it takes to simulate the model and find the intrusion area. It takes an average of 20 minutes to create the model, simulate it, and calculate to find the intrusion. With the trained machine learning models, the intrusion area can be found in less than a minute. The machine learning model also identifies the parameter ranges for less than 10% and 20% intrusion to guide fuel cell material selection and design.

Main Content
For improved accessibility of PDF content, download the file to your device.
Current View