Physics-Informed DeepONet for Field Reconstruction in Incompressible Two-Phase Slug Flow

This work investigates field reconstruction in incompressible two-phase slug flow using a DeepONet-based surrogate model. Time series measured at two sensor locations are used to reconstruct phase fractions and phase velocities along a 75 m horizontal pipeline. Physics-based regularization is incorporated through mixture mass conservation, phase-fraction closure, and sensor-consistency constraints, while a decomposed training strategy reduces computational cost by reusing branch embeddings across trunk queries. Results show modest but consistent improvements over the non-physics-informed model, with the largest gain observed for gas velocity prediction.