Bubble Detection and Analysis with Deep Learning

Deep Learning-based Approach to R-134a Bubble Detection and Analysis

Paper being submitted to International Journal of Heat and Mass Transfer (see publications)


This study used a deep learning method to analyse the bubble images of R134a fluid (ORC) produced on two types of heat exchanger plates: bare and coated stainless steel, where the coating was performed using novel GeoHex material on a stainless-steel substrate. The Mask R-CNN model has been used to train the deep learning algorithm for bubble detection. The deep learning algorithm calculated the number of detected bubbles, their sizes and centroids per frame, tracking each bubble frame by frame and evaluating the number of nucleation sites. Using the calculated data by the algorithm, the number of bubbles with time, the variation of single bubble diameter with time and the bubble size distribution are analysed. After that, active nucleation site density, average bubble departure diameter and bubble departure frequency are calculated using known formulas from the literature. The results showed no significant differences regarding bubble parameters between the coated and uncoated plates. (see Publications)

Fig: Experimental setup for boiling (bubble) video capturing

Fig: Bubble detection by eye (left) and the model for uncoated plate.

Fig: Bubble detection by eye (left) and the model for coated plate.