04. Boiling and Multi-Phase Flow

In two-phase immersion cooling, identifying bubble characteristics is crucial for heat transfer evaluation, yet remains challenging. Manual detection has been the norm, relying heavily on cognition of detailed features, which often leads to human error. To address this issue, this study develops an algorithm that accurately identifies and delineates bubble boundaries in images, even under challenging conditions such as overlapping, shadowing, and reflections. We employ a machine learning model to determine the probability of bubble presence and use a non-suppression method to extract and segment bubble boundaries. The multi-headed model is based on the U-Net architecture, which learns local and global features through down-sampling and skip connections. With the probability and the boundary heads sharing the same feature space, the model achieves accurate bubble prediction with reduced computational power. Additionally, the model handles blur bubbles by utilizing their global features, compensating for the loss of local details. Once the profile of an individual bubble is acquired, precise tracking becomes possible, facilitating the quantification of key parameters such as the velocity and volume of individual bubbles—essential for evaluating boiling heat transfer performance. Moreover, training data can be easily generated with auto-relabeled bubbles, enabling model calibration and transfer learning for bubble images taken under various conditions. While validation against previous works confirms the accuracy of our approach, it is also proven to accommodate images captured during pool boiling on a large heated surface, which often have unpredictable shadows and out-of-focus objects. We further demonstrate its practical application by establishing the relationship between the input heat flux and the bubble removal rate. This work represents a significant step forward in the automated analysis of bubble characteristics in pool boiling.