Deep Learning to Automate Image Analysis of Oxygen Bubble Dynamics in Proton Exchange Membrane Electrolyzers

Abstract

The massive use of fossil fuels due to strong industrialization and exponential population growth is the cause of human-made climate change and its unprecedented devastating environment benign, necessitating an immediate energy transition to renewable energy technologies. In this context, green hydrogen, seen as a fuel and a means of storing energy, represents an important option for an effective energy shift. With the aim of large scale hydrogen production and storage, significant efforts have been made in the field of material science to enable the development of economically viable proton exchange membrane (PEM) electrolyzers. However, water electrolyzers suffer from performance loss due to the formation and accumulation of oxygen bubbles at the anodic side. Understanding the processes involving bubbles and their impact on transport and reaction properties is crucial for optimizing PEM electrolyzers. Manual analyses of bubble distributions turn out to be inefficient, unreliable and expensive. In this master science work, I have tackled the manual-analysis-related problem by making use of deep learning to develop reliable, less expensive and rapid approach to analyse bubble area distributions. The work encompassed the training and evaluation of a Unet (2D) model on a collection of 35 manually annotated images (at a scale of 25 mm) for semantic segmentation, followed by the analysis of bubble area distributions. The results obtained demonstrate good model performance and robustness.