Perception-Aware Modeling and Vision-Based Identification of Two-Phase Flow During Gas Kick Migration

Abstract

Background conditions in drilling systems are increasingly shaped by the interaction of multiphase transport, uncertain downhole sensing, and limited direct observability of annular flow structure. Gas kick migration is therefore not only a hydraulic control problem but also an inference problem, because the quantities that matter operationally, such as gas holdup distribution, slip behavior, and transition to severe flow regimes, are only partially observed through pressure, return flow, and sparse surface measurements. This paper develops a perception-aware framework for gas kick analysis in vertical annuli that unifies mechanistic transport modeling with computer vision based flow regime identification. The central thesis is that regime recognition should be treated as an explicit state variable in well control and not as a descriptive afterthought. A physically structured model is formulated for transient gas migration under shut-in and circulating conditions, with closures that depend on local regime state, dissolution dynamics, and annular geometry. A complementary visual learning pipeline is then designed for regime classification from high-resolution annulus imagery, using temporal representation learning, calibration, and cross-domain adaptation to bridge laboratory observations and operational monitoring. The two components are fused through probabilistic state estimation so that image-derived regime beliefs update hydraulic state trajectories in real time. The resulting architecture supports earlier detection of hazardous regime transitions, more stable uncertainty quantification, and improved interpretation of surface indicators during kick handling. The paper argues that the most practical path toward robust digital well surveillance is not purely mechanistic or purely data-driven, but a coupled estimation architecture in which transport physics and visual evidence constrain one another continuously.