Regime-Conditioned Bayesian Digital Twins for Annular Multiphase Flow: Unified Modeling, Inference, and Control from Sparse Surface Measurements

• Santiago Arévalo

  Universidad de la Amazonia, Avenida Circunvalar 45--21, Departamento de Ingeniería de Sistemas, Florencia, Caquetá, Colombia

  Author

• Julián Restrepo

  Universidad Mariana, Calle 18 No. 34--52, Departamento de Ingeniería de Sistemas, Pasto, Nariño, Colombia

  Author

Multiphase flow in wellbore annuli and near-well conduits is a dominant determinant of pressure integrity, circulation stability, and the detectability of hazardous transients such as gas influx. In operational settings, the downhole state is only partially observed through surface sensors that are delayed, noisy, and confounded by uncertain rheology and evolving geometry. A central difficulty is that the effective constitutive behavior of gas--liquid mixtures is topology dependent: the same superficial velocities can yield distinct momentum losses depending on whether the mixture is dispersed, intermittent, or stratified, and regime transitions can occur abruptly under modest boundary changes. This paper proposes a unified digital-twin formulation that couples mechanistic conservation laws with a probabilistic regime process and a physically constrained closure-learning layer. The key contribution is a regime-conditioned Bayesian state-space model in which a discrete latent regime variable modulates the uncertain closure terms governing friction, slip, and compressibility residuals, thereby aligning state estimation, regime inference, and risk assessment under one probabilistic semantics. A constrained variational filtering method is developed to assimilate sparse surface pressure and flow signals while enforcing admissibility constraints on void fraction and mixture properties. The framework further embeds an influx source model for probabilistic kick detection and connects the posterior to a regime-aware stochastic model predictive control layer for pressure-window regulation. Simulation studies across inclination changes, sensor latency, frictional mismatch, and influx scenarios demonstrate improved calibration and more stable control relative to regime-agnostic probabilistic baselines, while remaining computationally compatible with real-time decision support.

• PDF