Real-Time Air-Water Volume Fraction Prediction Using Deep Learning and High-Speed Imaging

Abstract Submission: Accurate real-time measurement of gas volume fractions is essential for analyzing multiphase flows, especially in pipe systems where understanding gas-liquid dynamics is critical for optimizing fluid transport and identifying potential issues like pressure surges or blockages. This study used non-intrusive high-speed imaging to capture transient water-air interactions within a transparent pipe, providing high-resolution visual data of multiphase flow regimes. However, limited visual access in turbulent flows with gas-liquid interfaces complicates the interpretation of key quantities (e.g., void fraction, interfacial area). To address this, we developed a method to process video frames using a deep learning framework, combining the Segment Anything Model (SAM) with K-means clustering for image segmentation. A U-Net-based neural network was trained to predict water-air volume fractions across flow regimes. This method enables real-time multiphase flow analysis with applications in fluid transport optimization and pipeline monitoring. Validation was performed by comparing results with low-intrusive wire-mesh sensor (WMS) measurements taken in parallel on the same experimental loop, ensuring accurate predictions and assessing uncertainties in the image processing methodology.

Learning Objectives/Expected Outcome (Optional) :