Predicting through-Plane Porosity Profiles of Fibrous Porous Media from 2D Images with a Single-Input Multi-Output Convolutional Neural Network

While significant advances have been made in polymer electrolyte membrane fuel cells (PEMFC), water management in the gas diffusion layer (GDL) continues to be an important avenue for increasing overall cell efficiency (1). Efforts have been made to better understand how the structure of the GDL impacts water management through advanced microstructure characterization techniques such as synchrotron x-ray and laboratory-based computed tomography imaging, as well as neutron imaging (2-4). Deep learning tools like convolutional neural networks (CNNs) are openly available resources and have been employed for image classification and object detection since the late 1990s (5). While CNNs are less often used for real number regression tasks, they present a unique opportunity to gain meaningful insight into the GDL microstructure from lower quality data sources. Recently deep learning has been implemented in geological porous media applications to predict morphological, hydraulic, and mechanical properties with good success, and should be further investigated for use in the fuel cell community (6). In this study, a database containing over 2200 3D fibrous porous materials was created. The materials exhibited porosities ranging from 40% to 95% and were designed to represent GDLs in a PEMFC. The materials were used to create 2D images for training CNNs to predict average porosities and through plane porosity profiles. The CNNs, based on the popular ResNet50 and Xception network architectures, were adapted for real number regression rather than for traditional classification tasks. Both architectures accurately predicted average porosities with an R 2 of 0.98 (ResNet50) and 0.99 (Xception). Xception was then further adapted into a single-input multi-output convolutional neural network (SiMo CNN) and trained to predict through-plane porosity profiles using only 2D images. We achieved good results with the SiMo CNN for predicting porosity profiles with an R 2 of 0.91 and a mean absolute error of 1.7%. This study illustrates the usefulness of CNNs in image analysis of fibrous porous materials like the GDL and highlights the potential for CNNs to be further applied in the design and characterization of materials for electrochemical energy conversion. References Ijaodola OS, El-Hassan Z, Ogungbemi E, Khatib FN, Wilberforce T, Thompson J, et al. Energy efficiency improvements by investigating the water flooding management on proton exchange membrane fuel cell (PEMFC). Energy. 2019;179:246-67. Ince UU, Markötter H, George MG, Liu H, Ge N, Lee J, et al. Effects of compression on water distribution in gas diffusion layer materials of PEMFC in a point injection device by means of synchrotron X-ray imaging. Int J Hydrogen Energy. 2018;43(1):391-406. Battrell L, Patel V, Zhu N, Zhang L, Anderson R. Imaging of the desaturation of gas diffusion layers by synchrotron computed tomography. J Power Sources. 2019;416:155-62. Siegwart M, Harti RP, Manzi-Orezzoli V, Valsecchi J, Strobl M, Grünzweig C, et al. Selective visualization of water in fuel cell gas diffusion layers with neutron dark-field imaging. J Electrochem Soc. 2019;166(2):F149. LeCun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning applied to document recognition. Proc IEEE. 1998;86(11):2278-324. Rabbani A, Babaei M, Shams R, Da Wang Y, Chung T. DeePore: a deep learning workflow for rapid and comprehensive characterization of porous materials. Adv Water Resour. 2020;146:103787.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.