Image-Based Classification of Churn-to-Annular Flow Transitions in Upward Gas-Liquid Two-Phase Flow

Multiphase liquid-gas flow in vertical tubulars is central to processes such as oil and gas production, chemical processing, and geothermal systems. Accurate flow-pattern identification is critical for optimizing flow dynamics. This study develops and validates a deep-learning image classification approach using convolutional neural networks (CNNs) to characterize vertical flow patterns. Churn-to-annular flow transitions are classified, specifically churn, churn-annular, and annular flow regimes. In addition, key flow parameters, including superficial liquid and gas velocities (vSL and vSg), are predicted in upward gas-liquid two-phase flow. A dataset of 5000 labeled images from oil-air and water-air experiments in a 25-ft (7.6-m), 2-in. (0.0508-m) ID vertical flow loop is used for training and validation. Data augmentation improves model robustness, and multiple convolutional and pooling layers enable automated feature extraction. The model achieves 88 classification accuracy with RMSEs of 0.4 and 4 m/s for vSL and vSg, respectively, and maintains strong generalization on unseen flow conditions. Results demonstrate that image-based deep learning provides a reliable, scalable tool for flow-pattern prediction, highlighting its potential to enhance the design and optimization of multiphase transport systems.