Measurement of bubble characteristics in rod bundle flow channel using deep learning, 2; Development of bubble detection technique using deep learning

Hiramatsu, Natsuki*; Ono, Koji*; Uesawa, Shinichiro  ; Yoshida, Hiroyuki  

For the early practical deployment of innovative reactors, the use of detailed two-phase flow analysis is being considered as an alternative or complementary approach to large-scale mock-up experiments. In this study, we are developing a technique to calculate bubble centroid positions, diameters, and velocities and to reconstruct the three-dimensional bubble distribution by applying image segmentation, bounding-box-based tracking, and bubble matching between images acquired from two different directions, in order to validate the detailed two-phase flow analysis. This presentation reports the current status of the development of this technique and evaluates the impact of using not only experimental data but also numerical simulation data as training data on the detection accuracy.