Dissolution-Aware Hybrid Reduced-Order Modeling and Bayesian Inference for Real-Time Gas-Kick Detection in Managed Pressure Drilling

• Hamza Qureshi

  University of Swabi, Anbar Road, Swabi 23561, Pakistan

  Author

• Bilal Siddiqui

  Bacha Khan University, Main Umarzai Road, Charsadda 24420, Pakistan

  Author

Gas kicks remain a dominant source of unplanned events in drilling operations, particularly when narrow pressure margins and complex fluid systems compress the time available for diagnosis and response. Modern managed pressure drilling improves controllability, yet it also raises the bar for accurate, real-time inference because annular hydraulics, gas compressibility, and fluid property uncertainty interact nonlinearly with sensor signals such as pit gain, standpipe pressure, and return flow. A recurring difficulty is that dissolved gas can materially alter apparent pit-volume trends and pressure transients, while multiphase migration redistributes mass and momentum along the annulus. This paper develops a dissolution-aware state-space formulation for kick detection that couples one-dimensional transient annular transport with a thermodynamic closure for gas dissolution and mud swelling, while explicitly representing parametric and structural uncertainty in property models. The technical contribution is a structure-preserving hybrid reduced-order model that remains stable under aggressive time discretization and that is differentiable end-to-end, enabling fast Bayesian inversion of influx rate, depth, and composition proxies from streaming surface measurements. The reduced model is trained on physics-consistent trajectories rather than on raw data alone, and it enforces positivity, compressibility consistency, and mass conservation through constrained operator inference. A sequential Bayesian estimator with adaptive covariance inflation is then derived for online assimilation. Numerical studies demonstrate that dissolution-induced swelling can be disentangled from free-gas expansion in regimes where naive pit-gain heuristics fail, and that posterior uncertainty can be propagated into actionable decision thresholds without sacrificing real-time performance.

• PDF