Regime-Switching Neural-Operator Digital Twins for Gas-Liquid Two-Phase Flow with Bayesian Assimilation and Robust Control Interfaces

• Andrés Felipe Montoya

  Universidad de Nariño, Calle 18 50–02, Ciudadela Torobajo, Pasto, Nariño, Colombia

  Author

• Camilo Andrés Pineda

  Universidad del Magdalena, Carrera 32 22–08, Sector San Pedro Alejandrino, Santa Marta, Magdalena, Colombia

  Author

• Julián David Restrepo

  Universidad del Quindío, Carrera 15 12N–20, Barrio La Castellana, Armenia, Quindío, Colombia

  Author

Gas--liquid two-phase flow arises across drilling, well control, and pipeline transport, where rapid changes in holdup, pressure gradient, and interfacial structure can trigger operational risk. Although mechanistic models encode conservation laws, their closure relations and regime transitions are often uncertain under field variability, sensor sparsity, and geometry-dependent flow pattern shifts. This paper develops a real-time digital-twin framework that couples a regime-switching state-space model with physics-constrained neural operators to support inference and control under uncertainty. The core contribution is a hybrid architecture in which latent flow regimes act as discrete modes that gate continuous dynamics and closure surrogates, while a neural operator learns nonlocal constitutive mappings consistent with mass and momentum balances. A Bayesian assimilation layer performs joint filtering of continuous states and regime probabilities using likelihood models tailored to typical downhole and topside measurements. The resulting twin yields calibrated predictive distributions rather than point forecasts, enabling robust decision interfaces for choke and pump commands. We provide identifiability conditions for separating regime-induced dynamics from closure uncertainty, derive stability constraints for mode transitions, and present a computational study showing improved predictive reliability under sensor dropout and distribution shift relative to purely data-driven or purely mechanistic baselines. The framework is designed to integrate heterogeneous datasets, including flow-loop experiments and operational telemetry, and it supports online adaptation while maintaining physical consistency through constrained residual minimization.

• PDF