Synthetic Data and Generative Augmentation for Rare Two-Phase Flow Regimes in Vertical Annuli

Abstract

Rare flow patterns are often the most consequential yet the least represented parts of annulus image archives. In gas--liquid two-phase experiments, abundant video is usually collected under operating conditions that generate stable and frequently observed structures, whereas unusual, short-lived, or difficult-to-reproduce regimes occupy only a small fraction of the visual record. This imbalance weakens machine-vision systems because the model is trained mostly on common morphology while the decision regions associated with rare patterns remain poorly constrained. The resulting classifier may appear accurate in aggregate yet fail precisely where extrapolation, uncertainty, and event sensitivity matter most. This paper develops a synthetic-data and generative-augmentation framework for rare annulus flow-regime recognition. The proposed treatment does not view synthetic imagery as a wholesale substitute for real images. Instead, it treats synthetic generation as a controlled mechanism for expanding low-density regions of morphology space, broadening transition-adjacent visual support, and correcting imbalance in operating, optical, and geometric conditions. The framework combines physics-aware procedural generation, simulation-conditioned rendering, class-targeted latent-space synthesis, and uncertainty-weighted selection of accepted synthetic samples. It also formalizes the difference between useful rarity expansion and harmful synthetic drift by introducing acceptance criteria based on morphology consistency, feature-space proximity, calibration behavior, and downstream performance on experiment-disjoint real test sets. The result is a methodology for using generated annulus imagery to strengthen recognition of rare regimes while keeping the training distribution anchored to physically plausible and empirically validated structure.