Abstract
<jats:title>Abstract</jats:title> <jats:p>Flow regime identification in co-current upward gas-liquid flow through annular conduits is a critical challenge in petroleum engineering, with significant safety and operational implications, and it is likewise important across industries involving the transport of multiphase fluids. Misidentifying flow regimes can introduce major operational risk, yet regime boundaries in annular gas-liquid flow are often visually complex and context dependent. This motivates a variety of data-driven classification methods, including image-based, though practical performance remains limited by dataset size/consistency and the reliability of model confidence. In this study, convolutional neural networks (CNN) models are trained on split datasets to explicitly investigate how this dataset (~947-images within an annular conduit), while broadly representative of annular flow imagery yet specific in its composition, and its partitioning influence measured performance. Using dynamic, key-parameterized CNN architectures trained with cross-entropy loss and the Adam optimizer, temperature scaling calibrated via LBFGS, and three progressively refined foundational implementations (E1-E3), a total of 871 models were generated to probe preprocessing/input-handling choices and select hyperparameter effects. Results show that performance improves substantially across the implementation sequence and that moderately deep, more-pooled designs are the most consistently successful, after which three top models are examined more closely for calibration quality and practical efficiency. The discussion then focuses on why validation/calibration often underperform, how split/seed sensitivity and ambiguity may contribute, and how future work (more/better images, improved preprocessing and ROI/pipe detection, and stronger split controls) can further strengthen confidence in these models and approaches.</jats:p>