Uncertainty-Aware Surface Pressure Inversion for Real-Time Kick Diagnosis in Managed-Pressure Drilling

Abstract

Reliable kick diagnosis remains challenging in modern drilling because the primary signals available at the surface are indirect, delayed, and entangled with operational transients. In managed-pressure drilling, the same actuation that improves safety by regulating annular pressure also complicates interpretation of pressure histories, since choke control, compressibility, and multiphase slip jointly shape the measured response. This paper proposes a technical framework that treats kick diagnosis as a constrained inverse problem in which surface pressure time series are used to infer a latent, depth-distributed multiphase state with explicit thermodynamic consistency and quantified uncertainty. The core contribution is a dual-timescale reduced-order model that couples a one-dimensional annular multiphase flow description with an equilibrium-constrained dissolved-gas component, while enforcing physically admissible mixture densities, gas fractions, and pressure gradients through projection operators. The reduced model is embedded in a sequential Bayesian inference loop that performs joint state estimation and online calibration of nuisance parameters representing frictional drift, sensor bias, and uncertain mixture rheology. A likelihood design is introduced that separates actuation-induced pressure changes from influx-induced signatures via a disturbance-aware innovation process, enabling probabilistic alarms that are robust to routine choke and pump adjustments. The framework also yields posterior distributions over influx size, type proxies, and migration speed, allowing risk-ranked control recommendations. Numerical experiments illustrate conditions under which the inversion is identifiable from surface pressure alone and quantify the computational budget required for real-time deployment