Machine-learning-based framework for quantifying and formulating oxygen bubble dynamics in PEM water electrolysis

Oxygen bubbles generated near the catalyst layer in proton exchange membrane water electrolysis (PEMWE) obstruct deionized-water supply, hinder local mass transport, and increase mass-transport overvoltage. However, quantitative characterization is challenging because the large number of bubbles makes manual analysis impractical. Here, we propose a rapid and accurate framework to quantify bubble count and volume on the porous transport layer (PTL) surface by combining high-speed imaging with machine-learning–based detection and tracking (Faster R-CNN and Deep SORT). To reduce the effort required for training data preparation, we developed a pseudo-bubble image generator to automatically create annotated datasets. Systematic experiments were performed by varying the current density (0.5–4.0 A cm −2 ) and the anode-side water-supply pressure (1–5 atm). The total detected bubble volume was well described by a scaling model based on Faraday's and Boyle's laws. Bubble count and mean bubble volume increased with current density and decreased with pressure. Power-law fits revealed that operating conditions affected the count (exponent ∼0.6) more strongly than mean bubble volume (exponent ∼0.3), suggesting that changes in bubble count contribute more significantly to the oxygen removal than those in mean volume. The proposed framework provides quantitative data for numerical modeling, facilitating structural optimization of PTLs.