A physics-based synthetic image generator for bubble plumes

Quantitative imaging is a widely used method for characterizing bubbles in bubble plumes. However, overlapping bubbles in two-dimensional images pose significant challenges for image processing. Real-time data analysis is particularly difficult in field applications, such as natural marine seeps and lake aeration or oxygenation systems. While recent advances in machine learning offer new opportunities for automated analysis, the lack of high-quality, physically representative bubble image datasets limits model performance. In this study, we introduce PlumeDEBuG (Plume-Data Empowered Bubble Generator), an open-source synthetic image generator that produces realistic bubble plume images using a database of isolated bubbles extracted from laboratory experiments. The system assembles synthetic plumes using 78,323 individual bubble images (1–10 mm diameter) while preserving key physical characteristics. Specifically, it reproduces experimentally observed bubble size distributions by allowing users to select from Gaussian, Weibull, log-normal, or uniform distributions, depending on the desired flow condition. In addition, PlumeDEBuG provides user-defined controls for parameters including total bubble count, plume center location, plume width, bubble overlap rate, and void fraction, which enables generation of synthetic images that closely resemble real bubbly flow conditions. We tested PlumeDEBuG and trained several machine learning models for both object detection and segmentation methods. We demonstrate real-time bubble detection and size distribution estimation using an NVIDIA Jetson Nano (2GB), showing sufficient processing speed for field applications. Potential applications of PlumeDEBuG include studies of bubble ebullition, lake oxygenation, wastewater treatment, enhanced oil recovery, and other bubble-driven industrial processes.