Gas diffusion layer descriptors identified by machine learning to predict electrical contact resistance

Abstract

Proton exchange membrane fuel cells face performance limitations owing to their high electrical contact resistance (ECR) at the gas diffusion layer (GDL)-bipolar plate interface. In this study, a stochastic random Poisson process is employed to numerically generate GDL realizations coupled with various machine learning (ML) techniques, such as deep neural networks (DNNs) and random forests (RFs), which permit the identification of those descriptors characterizing a nonwoven carbon paper GDL that predict the pressure dependence of ECR when compressing the GDL. The DNN trained on a representative number of GDL realizations is used to accurately predict ECR for various sets of descriptors and RF is used to identify the importance of these descriptors within the sets. Using the so-classified descriptors, the ultimate goal is to reduce ECR occurring at the clamping pressure by manufacturing the GDL accordingly. Obviously, compression is shown to be the most important for predicting the pressure dependence of ECR however, some of the descriptors considered here, like fiber length, fiber density, bond density, and fiber-to-bond ratio, seem to be more crucial than traditional and widely used ones, such as porosity and the orientation of fibers, for example. These unexpected findings are just the first results of a large-scale simulation campaign, supported by ML techniques to reduce computational cost, which will presumably identify the GDL manufacturing parameters that unambiguously determine the ECR.