Flow-Regime Classification Benchmarks for Gas-Liquid Two-Phase Imaging in Vertical Annuli

Abstract

Gas--liquid two-phase flow in annular geometries appears in drilling hydraulics, wellbore transients, and multiphase transport where regime identification is coupled to pressure gradient, holdup, and slip. With modern high-speed cameras and inexpensive storage, vision-based regime recognition has become technically feasible, yet published results remain difficult to compare because datasets, sampling policies, and split definitions vary, and because correlated frames can inadvertently leak between training and test. This paper specifies a benchmarking methodology for flow-regime classification in vertical annulus imagery that is designed to separate true model generalization from artifacts of data partitioning and preprocessing. The benchmark definition formalizes the unit of independence at the experiment level, prescribes metadata needed to characterize optical and operating contexts, and provides protocols for constructing frame-level and sequence-level tasks without violating temporal or experimental separation. Evaluation is defined through a metric suite that reports class-balanced discrimination, probability quality, calibration, and robustness under controlled photometric and blur perturbations, together with uncertainty estimates computed by experiment-level resampling rather than frame-level assumptions. A baseline model suite is described to anchor comparisons across classical descriptors, convolutional encoders, and temporal aggregation architectures under standardized optimization and augmentation recipes. Finally, a reproducibility and governance layer is proposed that version-controls labels, documents preprocessing, and enforces evaluator-side metric computation so that reported numbers can be audited and re-scored as the benchmark evolves