Deep learning algorithms identify and track bubbles through gas-liquid interfaces in real time
Deep learning algorithms identify and track bubbles through gas-liquid interfaces in real time lead image
The flow characteristics of two-phase gas-liquid systems is vitally important to a wide range of applications, from chemical engineering to nuclear energy and environmental engineering. Predicting and detecting bubbles in such systems remains particularly challenging.
Fang et al. introduced a method for dynamically detecting these bubble interfaces in real time. Using the Deep Simple Online and Real-Time (DeepSORT) and You Only Look Once (YOLO) deep-learning algorithms, they achieved high-precision bubble detection and real-time tracking even in complex gas-liquid two-phase flow environments, which can be adapted to multiple different use cases.
“This system’s real-time capability represents a significant advancement over traditional methods, which often struggle with both speed and accuracy,” said author Yue Feng. “Combining cutting-edge object detection and tracking algorithms allows for immediate analysis and intervention, which is crucial for optimizing processes and ensuring safety.”
To tackle the issue, the group configured the algorithms to train in tandem on videos of bubbles evolving in liquids. Such a process required the team to iteratively tweak parameters and filters of the algorithms to achieve the best performance.
The algorithms automatically identify bubbles at the gas-liquid interface and can accurately delineate their boundaries with precise contours for each bubble. This foundational data about the bubbles helps fuel further bubble dynamics analysis, including tracking from frame to frame.
“Improvements in hardware, such as enhanced GPUs and specialized processors, significantly accelerated our processing capabilities and allowed for more complex computations to be performed in real-time,” Feng said.
The group hopes their work stimulates more efforts in monitoring gas-liquid interfaces. They next look to extend their analysis from two dimensions to three for improved detection.
Source: “A deep learning-based algorithm for rapid tracking and monitoring of gas-liquid two-phase bubbly flow bubbles,” by Lide Fang, Yiming Lei, Jianan Ning, Jingchi Zhang, and Yue Feng, Physics of Fluids (2024). The article can be accessed at https://doi.org/10.1063/5.0222856 .
