Data-driven fluid-dynamics of bubbles

Detection, counting and characterization of bubbles in a liquid is an important task in many industrial applications. These applications include monitoring bubbly flows in multi-phase processes used in energy and chemical industries. Typically, the aim is to measure the bubble size distribution (BSD). In this thesis, two approaches are presented for bubble recognition: one based on classical image analysis (segmentation and extraction), and one based on deep learning (especially with CNNs). Both approaches have the aim of extracting the main geometrical parameters of bubbles within an experimental video of flowing bubbles. The global parameters extracted by the two algorithms for each frame are: mean, maximum and minimum radii, mean eccentricity, dispersity, bubble size distribution (BSD) and bubble eccentricity distribution (BED). The analysis shows that the classical image analysis method is robust, effective and computationally fast; the CNNs based method has an interesting potential for the application, even though recursive tuning of the algorithm is necessary to obtain accurate results.