Physics-Guided Neural Network Approach for Pressure Gradient Prediction in Two and Three-Phase Flow Through Pipe and Annular Geometries
Abstract
Accurate prediction of pressure drop in multiphase flow systems is essential for the design and operation of pipelines in the energy and process industries. While numerous correlations exist for single-phase and gas-liquid two-phase flows, reliable prediction of pressure drop in three-phase flow systems remains a challenge, particularly for annular geometries. This study presents an artificial intelligence (AI)-based model for predicting the pressure gradient dP/dL in two and three-phase flows, developed using experimental and literature data. Experiments were conducted using a multiphase flow facility comprising horizontal pipe and annular flow loops, each with a 2 in. [5.08 cm] hydraulic diameter. Air-water, air-biopolymer (non-Newtonian, apparent viscosity 3.4 to 5.4 cp), and air-water-oil tests were performed across both geometries, covering superficial liquid velocities of 3.28-10.83 ft/s [1.0-3.3 m/s] and gas injection rates of 0.21-1.05 ft/s [0.063-0.32 m/s]. Dynamic and differential pressure sensors recorded high-frequency pressure signals. These experimental datasets serve as the training and validation basis for the physics-guided neural network model. Model performance is benchmarked against established correlations including Lockhart-Martinelli and Beggs-Brill. Experimental results indicate that three-phase (air-water-oil) flow reduces mean differential pressure relative to two-phase (air-water) flow at matched flow conditions, and that annular geometry produces a more damped pressure response than pipe geometry; a companion experimental study on the same facility further distinguishes the two geometries by transient response shape, with the pipe showing a predominantly peaked response to gas injection and the annulus a predominantly drifting one. For non-Newtonian air-biopolymer systems, geometric confinement was found to govern pressure gradient behavior more strongly than fluid rheology. These experimental findings motivate the need for a predictive framework that can simultaneously account for geometry, phase count, and fluid rheology, which existing empirical correlations do not fully address. The framework predicts the dP contribution from gas only with a median relative error of 41.3% under pooled cross-validation, competitive with gradient boosting (42.5%) and a pooled power law (42.7%). This work presents one of the first applications of a physics-guided neural network to pressure gradient prediction in three-phase multiphase flow encompassing both Newtonian and non-Newtonian liquid phases across pipe and annular geometries. The proposed approach enforces physical constraints -- positivity, the single-phase limit, and monotonicity in gas volume fraction -- by construction rather than by learning, and admits an entire geometry as a one-sided constraint when its signal lies below instrument resolution. The companion experimental dataset, acquired on a single matched-hydraulic-diameter facility in both geometries, provides a uniquely controlled basis for cross-geometry model evaluation.