Differentiable Thermo-Solutal Drift-Flux Inference for Real-Time Gas-Kick Detection Under Uncertainty

• Apichai Srisuwan

  Chiang Rai Rajabhat University, 80 Moo 9, Phahonyothin Road, Ban Du, Chiang Rai 57100, Thailand

  Author

• Nattapon Kanchana

  Roi Et Rajabhat University, 59 Moo 11, Roi Et–Selaphum Road, Roi Et 45120, Thailand

  Author

• Worawit Chaiyasit

  Nakhon Phanom University, 103 Moo 3, Nakhon Phanom–That Phanom Road, Nakhon Phanom 48000, Thailand

  Author

Transient gas influx into a drilling annulus remains a principal source of operational risk because small early-stage perturbations can evolve into rapid gas unloading at the surface. Conventional monitoring relies on sparse, noisy surface signals whose relationship to downhole multiphase dynamics is strongly nonlinear and history dependent through compressibility, slip, and thermodynamic exchange between phases. This paper develops a real-time, physics-constrained inference framework that couples a thermo-hydrodynamic drift-flux model with a solubility-consistent thermodynamic closure, yielding a differentiable simulator suitable for online inversion. The central contribution is a state-space formulation in which gas influx is treated as an unknown, time-varying boundary flux and estimated jointly with uncertain hydraulic and mass-transfer parameters from pit-volume, pump-rate, and wellhead-pressure observations. Differentiability is enforced by implicit solvers and adjoint sensitivity methods, enabling fast gradient-based posterior updates and likelihood-ratio detection statistics. The model incorporates a unified representation of free-gas holdup, dissolved-gas inventory, and energy transport so that delays induced by dissolution and thermal expansion can be separated from hydraulic signatures. Numerical experiments demonstrate stable online estimation across regimes spanning non-circulating to high-rate circulation, with detection latency governed by observability limits rather than solver cost. The framework is intended as a computational backbone for kick-aware automation, emphasizing uncertainty quantification and identifiability over scenario-specific tuning.

• PDF