How deep learning is accelerating multiscale design of porous electrodes for flow cells

To meet urgent net-zero goals, the global energy system is shifting from fossil fuels to renewable sources such as solar and wind. Because these sources are intermittent, efficient storage and conversion are essential. Electrochemical technologies including fuel cells, water electrolyzers, and redox-flow batteries are promising because they decouple energy from power and can operate flexibly. A central bottleneck, however, is the porous electrode. Its complex micro and nano scale features create anisotropic mass transport that is hard to predict and optimize, slowing progress toward higher energy and power density.

In this study, a research team developed a deep learning approach called Electrode Net to accelerate porous-electrode design without sacrificing accuracy. The method represents three-dimensional electrode geometry using signed distance fields and then applies a three-dimensional convolutional neural network to learn the link between structure and transport performance. This combination captures geometry cleanly and enables fast, reliable predictions.

To train and test the model, the researchers built a validated pore-network framework and assembled a comprehensive dataset of 15,433 porous samples paired with their anisotropic transport properties. Across benchmarks, Electrode Net achieved a coefficient of determination (R²) greater than 0.95, outperforming other advanced models on the same tasks. Speed is a key advantage. Guided by the signed distance field representation, Electrode Net cuts computation time by as much as 96% than the conventional numerical simulation models, while maintaining high fidelity. In practice, this turns many hours of simulation into minutes or seconds of model inference, enabling rapid screening of large design spaces.

The team further validated the approach on real electrodes from three technology classes: fuel cells, water electrolyzers, and redox-flow batteries. In each case, the model delivered excellent predictive accuracy, demonstrating strong cross-system generalization and suggesting that the framework can be adopted across diverse subjects.

Beyond fast predictions, the researchers introduced a practical multiscale design workflow. Electrode Net first estimates pore-scale, anisotropic transport parameters. These parameters are then embedded into cell-scale simulations to guide device-level optimization under realistic operating constraints. Using the gas diffusion layer of a proton-exchange-membrane fuel cell as an example, the workflow produced electrode designs with significantly higher limiting power density and limiting current density.

Together, these results show that learning directly from three-dimensional structure can remove a long-standing bottleneck in electrochemical device development. Electrode Net offers a general and scalable path to design porous electrodes faster and more accurately, helping advance next-generation clean-energy technologies.

---

---