Computer vision-enabled multi-level analysis and prediction of physical field distribution for PEMFCs

Most of existing artificial intelligence (AI)-aided optimizations of flow fields for proton exchange membrane fuel cells (PEMFCs) are value-oriented, focusing on scalar performance indicators, while the spatial distribution patterns of internal physical fields are largely overlooked. To address this limitation, this work establishes a multi-level data-driven framework that transforms PEMFC analysis from value-based performance prediction to vision-oriented recognition and generation of full-field physical distributions by leveraging computer vision. Value-level sensitivity analysis identified the dominant cathode hydrothermal parameters, while vision-level statistical analysis of distribution characteristics quantified the spatial signatures. Unsupervised clustering further revealed five typical vision patterns of field distribution (Silhouette = 0.627), bridging the gap between value parameters and visualized distributions, which established a correlation between value and vision. Then, two new variants of conditional generative adversarial network (cGAN) are proposed: one is Class-GAN inspired by clustering ideas, which introduces pixel-level classification constraints to enhance boundary fidelity, and the other is Lite-GAN that adopts a lightweight architecture to improve computing efficiency. The results showed that the highest structural similarity index (SSIM) of Class-GAN reached 0.9297, and Lite-GAN reached 0.8036, both significantly better than the baseline cGAN (0.4672). The proposed AI-aided model reduces the single case prediction time from approximately 2.5 h by manual simulation to within 2 s. This framework enables interpretable quasi-real-time analysis, providing a useful tool for rapid evaluation and optimization of bipolar plate design.