Leveraging machine learning for advanced flow field design in PEMFCs

The overall performance of proton exchange membrane fuel cells (PEMFCs) strongly depends on the design of the flow field. This study presents a novel, enhanced tapered parallel flow field featuring sub-channels with widths that vary in a precisely engineered converging–diverging pattern. This innovative design significantly improves oxygen transport in both through-plane and in-plane directions, thereby enhancing water management and ensuring highly consistent reactant delivery to reaction sites. In addition, a machine-learning-based optimisation framework is developed for this flow field. Using a rigorously validated three-dimensional, two-phase CFD model, an extensive dataset of 184 cases is generated to train seven distinct data-driven surrogate models: adaptive neuro fuzzy inference system (ANFIS), artificial neural network (ANN), response surface methodology (RSM), random forest (RF), CatBoost, XGBoost, and LightGBM. Notably, CatBoost demonstrates superior predictive accuracy for key oxygen mass-transfer metrics and was consequently employed in a sophisticated multi-objective optimization. This process yields an optimal flow-field geometry with a tapering ratio of 3.8, a cycling amplitude of 0.57 mm, and eight cycles at 0.694 V, achieving a high mean oxygen concentration of 0.020 kmol.m −3 and an excellent uniformity index of 0.93. This integrated machine learning-accelerated optimization framework enables rapid and reliable flow-field optimisation and provides practical, actionable design guidelines for effectively reducing oxygen starvation in next-generation, high-performance fuel-cell stacks.