Answers are grounded in this paper and cite their sources.
We investigate a physics-constrained neural network for reconstructing transient gas–liquid pipe flow under prescribed initial and boundary conditions. The method combines phase mass and mixture momentum equations with a horizontal drift-flux closure, bounded trainable corrections, and differentiable interpolation of SRK property tables. It uses temporal domain decomposition, a physics-loss schedule based on the training step, and a liquid-phase integral mass constraint. Training uses only the initial spatial profile, inlet gas and liquid mass-flow histories, and outlet pressure. Interior reference fields at positive times are reserved for post-training evaluation and are excluded from the loss and checkpoint selection. The baseline slug-flow case has overall methane/water feed mole fractions of 0.04/0.96. Its relative RMSEs are 0.64% for pressure, 3.52% for liquid velocity, 0.83% for gas void fraction, and 1.41% for gas velocity. The gas void fraction MAE is 0.0026. Gas compressibility and interphase slip influence the spatiotemporal evolution of the gas-phase variables. Additional slug, bubbly, and stratified-related cases are considered to assess reconstruction across flow states. All six cases are trained independently for 20,000 optimization steps.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: