Abstract
`Gas--liquid two-phase flow in annular passages exhibits visually distinct macrostates that correlate with pressure gradient, phase distribution, and intermittency. High-speed imaging makes these macrostates observable, but building reliable recognition models is limited by the cost of dense labeling, by strong temporal correlation in videos, and by domain shift across optical configurations and operating points. This paper develops a self-supervised pretraining framework for two-phase flow vision that leverages unlabeled annulus imagery to learn transferable representations prior to supervised regime fine-tuning. The approach combines instance-level invariance with physically plausible augmentations, masked reconstruction that emphasizes interface geometry, and temporal consistency objectives that exploit frame-to-frame continuity while avoiding information leakage. The learned encoder is then adapted to regime classification under constrained label budgets, with explicit strategies for class imbalance, transition ambiguity, and calibration. To enable credible comparisons, the paper specifies an evaluation protocol that separates experiments at the split level, reports class-balanced discrimination and probability quality, and quantifies robustness under controlled photometric and blur perturbations. The analysis emphasizes when self-supervision improves generalization to unseen runs and when gains vanish due to shortcut learning or over-tuning on validation. The resulting methodology is intended as a practical blueprint for researchers who wish to exploit abundant unlabeled two-phase imagery to reduce labeling demand while improving cross-condition reliability of regime recognition in annular geometries.