Autonomous Design of Ordered Gas Diffusion Layers for High-performing Fuel Cells via Bayesian Machine Learning

Abstract Rational design of gas diffusion layers (GDL) is an example of a long-standing pursuit to increase the power density and reduce the cost of proton exchange membrane fuel cells (PEMFC). However, current state-of-the-art GDLs are designed by trial and error, which is a time-consuming endeavor. Here, we propose an autonomous Bayesian machine learning approach to optimize the design of GDL structures. With the artificial neural network accelerating the calculation of anisotropic transport properties of reconstructed 7621 fibrous GDLs, Bayesian optimization algorithm identifies optimal structures in only 40 steps, maximizing the PEMFC’s limiting current density. Results suggest that the optimal GDL structure consists of highly orientated fibers with moderate diameters (~10 µm), which is successfully fabricated with a controlled electrospinning technique. Impressively, the PEMFC demonstrates a record high power density of 2.17 W cm-2 and a limiting current density of ~7200 mA cm-2, far exceeding that with commercial GDL which only achieves 1.33 W cm-2 and ~2700 mA cm-2. The approach reported here represents how the advanced algorithms can aid the innovative material design to address the critical water flooding issues in PEMFCs, leading to significant performance improvements, which paves the way for further application across various scientific fields.