Gas bubble detection and segmentation using a machine learning approach leveraging semi-supervised training

This work presents a novel training approach for machine learning based instance segmentation, without the need for manual annotated datasets. With applications in experimental investigations in bubbly flows in reactor safety research. This semi-supervised process consists of two different neural networks, a conditional generative adversarial network used for data generation and a U-net style convolutional neural network for instance segmentation. We validated the approach using a fully automated experimental setup creating layered bubble curtains, which enables an evaluation of the bubble size distribution with and without overlapping bubbles. The final neural network is able to measure the average bubble size as well as to recreate the bubble size distribution accurately. • Training machine learning models on segmenting gas bubbles without manual annotated data • Dual stage training using a conditional generative adversarial neural network in combination with u-net style network for segmentation • Evaluation of the segmentation performance with reference experiments using spargers creating layered bubble curtains • Fully automated experimental setup