Revealing anodic multi-class bubble dynamics in PEMWE systems using deep learning and post-processing detection

Oxygen bubbles that emerge in the anodic side of a Proton Exchange Membrane Water Electrolyzer (PEMWE) can significantly decrease the efficiency of the system. Therefore, a deeper understanding of the bubble's behavior is crucial. However, this two-phase flow analysis is a challenging problem due to its complexity and remains a major scientific issue. In this paper, a fine-class deep learning detection tool is developed to tackle this issue. The proposed strategy is designed for the detection of three classes of bubbles: bubbly, slug, and stagnated. Based on these detections, several indicators are computed such as the number of bubbles or the covering rate. A high-density acquisition system coupled with a transparent anodic side PEMWE are used to capture anodic high-resolution bubble pictures. The proposed deep learning tool in combination with an image post-processing method carries out the detection of multiple bubble labels. Curve trends for the three different classes are obtained and are in concordance with the literature. For the first time, stagnated bubble dynamics are extracted from data. It is found that the water flow rate has no influence on stagnated bubbles covering rate, amount, and mean stagnated bubble size. However, increasing the current density decreases the covering rate and amount of stagnated bubbles which frees active areas. When the water flow rate increases, the global bubble covering rate decreases, nevertheless the amount of bubbles counter-intuitively increases. Thanks to the multi-class bubble detection, this phenomenon can be explained by the fact that slug amount decreases due to the non-coalescence phenomenon, and the bubbly amount increases. The developed tool is efficient and could be used to analyze bubble characteristics after modifying the PEMWE such as the porous transport layer, catalyst layer, or even the membrane.