Bubble shape classification in a bubble column based on multi-input ConvNeXt model

Abstract

A comprehensive understanding of the continuous variation and deformation of rising bubbles is essential for precise reactor scale-up and process optimization. This work combines telecentric vision probe and bubble boundary R-CNN with a newly developed multi-input ConvNeXt to pioneer bubble shape classification under realistic flow conditions. The classification results demonstrate that ellipsoidal bubbles constitute the predominant shape (49.5%), while oblate ellipsoidal bubbles represent the smallest proportions (2.7%). Statistical analysis of E and d_m across all classified bubbles reveals significant distinctions: spherical and oblate ellipsoidal bubbles exhibit pronounced differences in E compared to other classes, whereas the remaining shapes show relative consistency. d_m varies substantially across classes, progressively increasing in the order: spherical, ellipsoidal, spherical cap, ellipsoidal cap, oblate ellipsoidal, and wobbling ellipsoidal. While our model achieves bubble classification, unresolved issues span image quality constraints, 3D shape inference from 2D data, turbulent flow coupling, and high-velocity applicability—necessitating integrated imaging-algorithm advances.