(a) Captured image (a single non-overlapping bubble and three overlapping bubbles), (b) detection results (green lines) using rule-based image recognition, and (c) detection results (colors) using the proposed method.

A bubble detection technique for gas–liquid two-phase flows was developed using the latest deep learning technology, Shifted Window Transformer (Swin Transformer). By achieving highly accurate bubble shape detection, essential data could be acquired in greater detail, such as bubble size, distribution, and void fraction for thermal-hydraulic studies, including nuclear engineering.

Conventional bubble detection is based on rule-based image processing using differences in brightness. However, accuracy of these methods decreases when bubbles are deformed or overlap.

In this study, the deep-learning-based image recognition model, Swin Transformer, was used. Training was conducted using only several dozen bubble images. The proposed method was compared with conventional rule-based techniques in Fig. 1. This approach successfully detected bubble shapes even when they overlapped as shown in Fig. 1 (c), outperforming conventional methods. Furthermore, the detector was applied to a 3×3 rod bundle channel and could successfully identify bubbles in complex geometries (Fig. 2).

This technique demonstrates high accuracy even with limited training data and adapts well to complex fluid structures and experimental conditions. It is expected to serve as a reliable image analysis tool in future thermal-hydraulic research and simulation validation.

Uesawa, S. et al., Deep Learning-Based Bubble Detection With Swin Transformer, Journal of Nuclear Science and Technology, vol.61, issue 11, 2024, p.1438–1452.

Paper URL: https://doi.org/10.1080/00223131.2024.2348023/a>