Active learning assisted optimization unlocks optimal gas diffusion layer properties for anion exchange membrane water electrolyzers

Anion exchange membrane water electrolyzers are widely recognized as nextgeneration electrochemical devices for green hydrogen production, integrating the advantages of alkaline water electrolysis and proton exchange membrane water electrolysis while avoiding their respective limitations. The gas diffusion layer serves as a critical component linking the flow channel and catalyst layer, performing multiple functions including electron conduction, gasliquid twophase mass transport, and heat transfer. However, systematic quantitative frameworks for the synergistic optimization of GDL physical properties remain lacking.

In a study published in ENG. Chem. Eng., researchers at China University of Petroleum (Beijing) and collaborators established a threedimensional steadystate multiphysics numerical model of AEMWEs using COMSOL Multiphysics 6.3. The model was validated against experimental polarization data, with a mean relative error of only 0.7 % and a maximum relative deviation of 1.3 % over the entire operating range. Systematic singlefactor parametric analyses revealed the regulatory mechanisms of permeability, electrical conductivity, and porosity on mass transport and heat transfer.

To overcome the high computational cost of highfidelity numerical simulations, the team constructed a weighted ensemble surrogate model integrating four Gaussian process regression models with heterogeneous kernels, a Universal Kriging model, and a Bootstrap ExtraTrees model. An active learning framework with uncertaintydriven dynamic sampling and batch clustering strategies achieved efficient global optimization in the highdimensional parameter space. After nine iterations, the model achieved a testset coefficient of determination of 0.9994, a root mean square error of 5.0240, and a mean absolute error of 4.0694.

Parametric analysis revealed that permeability exerts a dominant influence on mass transport. When permeability increased to 1 × 10⁻¹⁰ m², a discontinuous transition occurred from viscous resistancedominated flow to inertiadominated twophase separated flow, dramatically improving gas uniformity. However, excessively high permeability beyond this critical value led to loss of driving force and deteriorated mass transport. Electrical conductivity monotonically improved gas uniformity, while porosity exhibited differential effects on hydrogen and oxygen distribution.

Heat transfer analysis showed that the optimal permeability of 1 × 10⁻¹⁰ m² and porosity of 0.60 produced the most uniform thermal field distribution. Electrical conductivity had no significant effect on the temperature field morphology within the investigated range.

The active learning framework identified the optimal GDL parameter combination for electrochemical performance: a permeability of 1.18 × 10⁻¹⁰ m², an electrical conductivity of 2208.86 S·m⁻¹, and a porosity of 0.605. Under this configuration, the normal current density reached 14934.52 A·m⁻², representing a 4.1 % improvement over the baseline. Multiphysics verification confirmed that this optimal combination effectively suppressed local heat accumulation, eliminated mass transport bottlenecks, and maintained uniform species and temperature distributions.

This study provides both qualitative and quantitative theoretical support for the material selection and engineering design of highperformance electrolyzers.