Electrospinning Machine| Artificial-intelligence-guided design of ordered gas diffusion layers for high-performing fuel cells via Bayesian machine learning

To achieve the carbon neutrality goal, using renewable energy to produce green hydrogen and efficiently convert it into electricity through proton exchange membrane fuel cells (PEMFC) has become an important path for energy transition. However, the large-scale application of PEMFC urgently needs to increase its power density, reduce costs, and enhance operational stability, among which the "flooding" problem seriously limits performance and life at high current density. Although existing research attempts to optimize the structure and wettability of the gas diffusion layer (GDL) to improve water management, the random, porous structure makes it difficult to effectively solve water accumulation, especially the prominent problem of water accumulation under the rib area. In addition, the traditional trial-and-error method is difficult to systematically and efficiently optimize the GDL structure. Therefore, an optimization strategy combining artificial intelligence with multi-scale, multi-physics models has become a topic urgently needing solving in current research.

Recently, Academician Tianshou Zhao from the Hong Kong University of Science and Technology published a research achievement titled "Artificial-intelligence-guided design of ordered gas diffusion layers for high-performing fuel cells via Bayesian machine learning" in the internationally renowned journal Nature Communications. This research addresses the challenging problem of optimizing the design of the gas diffusion layer structure in proton exchange membrane fuel cells by proposing a closed-loop machine learning framework based on Bayesian optimization and artificial neural networks. This method accelerates the calculation of anisotropic transport properties of the GDL, obtaining the optimal structural parameters in only 40 steps, significantly enhancing the fuel cell's limiting current density. Based on this guidance, controlled electrospinning was successfully used to prepare an optimal GDL structure with high fiber orientation and moderate diameter, achieving a battery power density of 2.17 W cm⁻² and a limiting current density as high as about 7200 mA cm⁻², far exceeding the level of commercial GDLs.
This study proposes a closed-loop multi-scale modeling method based on Bayesian optimization and artificial neural networks to achieve rapid structural optimization of the gas diffusion layer in proton exchange membrane fuel cells. First, an ANN model was used to significantly accelerate the calculation of pore-scale anisotropic transport properties of the GDL. Then, using the limiting current density as the objective function, the BO algorithm quickly searched and obtained the optimal structural parameters within 40 iterations. The study found that the optimal GDL structure consists of oriented fibers with moderate diameter.

Figure 1. Workflow of closed-loop Bayesian machine learning for GDL design.

This study prepared a controllably fiber-diameter ordered and oriented gas diffusion layer by adjusting the PAN precursor solution concentration and electrospinning conditions. Adding cobalt salt promoted fiber alignment and graphitization. After heat treatment, the fiber diameter increased with the PAN concentration. After hydrophobic treatment, the contact angle increased to 144°, close to that of commercial GDL (147°). Raman spectroscopy indicated that the prepared GDL had a slightly lower degree of graphitization than commercial products, but CT scans showed it had better fiber alignment and a larger pore size distribution range, which is expected to enhance fuel cell performance.

 Figure 2. Preparation and characterization of ordered aGDL.

This study systematically analyzed the influence of gas diffusion layer structure on the performance of proton exchange membrane fuel cells. Using the orthogonal experimental method, the effects of fiber diameter, the presence or absence of a microporous layer (MPL), fiber alignment direction, and the degree of randomization were studied. The results showed that as the GDL fiber diameter increases, the pore size increases, which is beneficial for improving water removal performance at high current densities. Among them, aGDL_d3 had the largest pore size and exhibited the best mass transport performance. The MPL prepared using electrospun fibers effectively reduced the contact resistance. When the fibers were placed perpendicular to the flow channels, the performance was superior to the structure where they were placed parallel, because the latter increases the ohmic resistance and mass transport losses of the cell. In addition, the randomly arranged fiber layers prepared by staggered stacking had reduced pore sizes compared to the aligned fiber layers, leading to reduced water management performance. Therefore, using an ordered structure with larger diameter fibers and placing them perpendicular to the flow channel direction effectively enhanced PEMFC performance.

Figure 3. Performance comparison of ordered aGDL with different structures and arrangements.

The study evaluated the practical application potential of the self-made oriented aGDL by comparing its performance with the commercial random cGDL in proton exchange membrane fuel cells. Under standard conditions, the aGDL cell achieved a limiting current density and peak power density of 7200 mA cm⁻² and 2.17 W cm⁻², respectively, which are 2.7 times and 1.6 times those of the cGDL cell. Furthermore, under harsh operating conditions such as low gas flow rate, low temperature, high humidity, and high back pressure, the aGDL cell still exhibited lower concentration polarization losses and higher stability. Stability tests showed that the aGDL cell could operate stably at 0.4 V, with the current density stable above 5000 mA cm⁻², while the cGDL cell showed significant fluctuations. Meanwhile, accelerated stress tests (AST) indicated that aGDL has excellent corrosion resistance, is reusable, and recovers performance well. Overall, aGDL outperformed cGDL under various operating conditions, demonstrating strong potential for practical application replacement.

Figure 4. Performance verification of ordered aGDL and commercial cGDL under different experimental conditions.

This study used electrochemical impedance spectroscopy (EIS), limiting current experiments, pore-scale simulation, and finite element simulation methods to analyze the water and gas transport performance of proton exchange membrane fuel cells containing oriented gas diffusion layers (aGDL) and commercial gas diffusion layers (cGDL). Fitting the EIS test data with an equivalent circuit model found that cGDL suffered from severe water flooding at current densities exceeding 2 A cm⁻², leading to significant concentration polarization losses, while aGDL remained stable up to 5 A cm⁻², only slightly increasing at 6 A cm⁻². Oxygen transport resistance (OTR) measured by the limiting current density method showed that aGDL exhibited significantly lower OTR under different oxygen concentrations and back pressure conditions, and it showed a linear relationship with pressure, indicating excellent anti-flooding performance. Pore-scale simulations revealed that liquid water saturation in aGDL was significantly lower than in cGDL, showing better drainage performance in both transverse and longitudinal directions. Cell-scale finite element simulations further confirmed that aGDL significantly reduced liquid water saturation and improved oxygen pressure distribution, thereby enhancing cell performance.

Figure 5. Experimental and numerical water-gas transport analysis and mechanism verification.

This study optimized the structure of a GDL with oriented fibers through an AI design strategy combining multi-scale modeling and Bayesian machine learning, and successfully prepared an ordered GDL using a highly controllable electrospinning method. After systematically studying the effects of fiber orientation, diameter, MPL presence, and stacking methods on PEMFC performance, it was found that the optimized oriented GDL, when placed perpendicular to the flow channels, achieved a limiting current density of up to 7200 mA cm⁻² and a peak power density of 2.17 W cm⁻², which are 2.7 times and 1.6 times those of commercial GDL, respectively. This improvement remained stable under different temperatures, humidity levels, flow rates, and back pressures, and the oriented GDL possessed outstanding anti-flooding and anti-corrosion performance. Electrochemical impedance and limiting current tests, as well as pore-scale and cell-scale simulations, all proved its superior water management capability. The AI-driven design and synthesis process brings a breakthrough for the next generation of PEMFC GDL technology and also provides a generalizable paradigm for design optimization of complex problems in other fields.

Paper link: Sun, J., Lin, P., Zeng, L. et al. Artificial-intelligence-guided design of ordered gas diffusion layers for high-performing fuel cells via Bayesian machine learning. Nat Commun 16, 6528 (2025).https://doi.org/10.1038/s41467-025-61794-y