Accurate reduced-order prediction of stratified gas-liquid flow in horizontal pipes remains strongly limited by the closure of unresolved wall and interfacial momentum-transfer mechanisms. Classical one-dimensional two-fluid models are computationally efficient and physically interpretable, but their accuracy depends on empirical closure relations for wall shear, interfacial shear, and turbulent momentum transport. Experiments provide reliable global quantities such as pressure gradient, void fraction, liquid holdup, and superficial velocities, while CFD can provide local flow information but remains computationally expensive and model-dependent. This work develops a hybrid inverse physics-informed neural-network framework for reduced-order turbulence-closure discovery in smooth, fully developed stratified air-water pipe flow. The closure correlations are derived from eight operating conditions with Re_L = 3450-5310 and Re_g = 2444-10134, and are therefore intended for this smooth-interface reduced-flow envelope rather than as universal turbulence laws. The proposed method combines experimental operating quantities, sparse CFD-derived velocity samples, and reduced phase-wise momentum equations. Separate gas and liquid neural networks infer the nondimensional axial midplane velocity profiles and the corresponding effective turbulent-viscosity fields. In this role, the framework can be interpreted as a physics-informed virtual sensor, estimating hidden closure-relevant quantities from measurable operating conditions, CFD-derived local velocity samples, and governing physical constraints. The inverse problem is constrained by the measured pressure-gradient forcing, wall no-slip conditions, phase mean-flow constraints, and dimensional velocity and shear-stress continuity at the gas-liquid interface. As a consequence of the reduced-order midplane formulation and the data-assisted inverse procedure, the inferred turbulent viscosity should be regarded as an effective momentum-transport model for the present stratified-flow configuration, rather than as a universal three-dimensional eddy-viscosity model. The inferred closure fields reveal distinct phase-dependent structures. In the liquid phase, the effective turbulent viscosity increases monotonically from the wall toward the interface and can be represented by a compact wall-to-interface algebraic form. In the gas phase, the inferred profiles exhibit near-wall damping, an interior peak, and a finite interfacial value, motivating a blended closure based on wall-side and interface-side contributions. The resulting algebraic closures are correlated with Reynolds-number-based variables and then inserted into an independent finite-difference reduced solver. The solver reconstructs coupled gas-liquid velocity profiles while satisfying the reduced momentum equations, mean-flow constraints, and interface-continuity conditions without using neural-network outputs or PINN residual minimization. The results demonstrate that physics-informed inversion can be used not only for sparse-data velocity reconstruction, but also as a closure-discovery tool for reduced-order stratified-flow modeling. Within the present smooth-interface, one-dimensional reduced formulation, the framework provides a physically interpretable bridge between experimental measurements, sparse CFD information, inferred effective closure fields, and solver-compatible algebraic turbulence closures.
La previsione ridotta accurata del flusso stratificato gas-liquido in tubazioni orizzontali resta fortemente limitata dalla chiusura dei meccanismi non risolti di trasferimento della quantità di moto alla parete e all'interfaccia. I modelli monodimensionali classici a due fluidi sono computazionalmente efficienti e fisicamente interpretabili, ma la loro accuratezza dipende da relazioni di chiusura empiriche per lo sforzo di taglio alla parete, lo sforzo di taglio interfaciale e il trasporto turbolento della quantità di moto. Gli esperimenti forniscono grandezze globali affidabili, come gradiente di pressione, frazione di vuoto, holdup liquido e velocità superficiali, mentre la CFD può fornire informazioni locali sul flusso, ma rimane computazionalmente costosa e dipendente dal modello. Questo lavoro sviluppa un framework ibrido inverso basato su reti neurali physics-informed per la scoperta di chiusure di turbolenza di ordine ridotto in flussi aria-acqua stratificati, lisci e completamente sviluppati in tubazioni. Le correlazioni di chiusura sono derivate da otto condizioni operative con Re_L = 3450-5310 e Re_g = 2444-10134, e sono quindi destinate a questo inviluppo di flusso ridotto a interfaccia liscia, piuttosto che a leggi universali di turbolenza. Il metodo proposto combina grandezze operative sperimentali, campioni sparsi di velocità derivati da CFD ed equazioni ridotte della quantità di moto per ciascuna fase. Reti neurali separate per gas e liquido inferiscono i profili adimensionali di velocità assiale lungo il piano medio e i corrispondenti campi di viscosità turbolenta efficace. In questo ruolo, il framework può essere interpretato come un sensore virtuale physics-informed, stimando quantità nascoste rilevanti per la chiusura a partire da condizioni operative misurabili, campioni locali di velocità derivati da CFD e vincoli fisici governanti. Il problema inverso è vincolato dal forzante di gradiente di pressione misurato, dalle condizioni di non scorrimento alla parete, dai vincoli di portata media di fase e dalla continuità dimensionale di velocità e di sforzo di taglio all'interfaccia gas-liquido. Come conseguenza della formulazione ridotta sul piano medio e della procedura inversa assistita dai dati, la viscosità turbolenta inferita deve essere considerata come un modello efficace di trasporto della quantità di moto per la presente configurazione di flusso stratificato, piuttosto che come un modello universale tridimensionale di viscosità turbolenta. I campi di chiusura inferiti rivelano strutture distinte dipendenti dalla fase. Nella fase liquida, la viscosità turbolenta efficace aumenta monotonicamente dalla parete verso l'interfaccia e può essere rappresentata da una forma algebrica compatta parete-interfaccia. Nella fase gassosa, i profili inferiti mostrano smorzamento vicino alla parete, un picco interno e un valore interfaciale finito, motivando una chiusura blended basata su contributi lato parete e lato interfaccia. Le chiusure algebriche risultanti sono correlate con variabili basate sui numeri di Reynolds e quindi inserite in un solutore ridotto indipendente a differenze finite. Il solutore ricostruisce profili di velocità gas--liquido accoppiati soddisfacendo le equazioni ridotte della quantità di moto, i vincoli di portata media e le condizioni di continuità all'interfaccia, senza utilizzare output delle reti neurali o minimizzazione dei residui PINN. I risultati dimostrano che l'inversione physics-informed può essere utilizzata non solo per la ricostruzione della velocità da dati sparsi, ma anche come strumento di scoperta di chiusure per la modellazione ridotta di flussi stratificati. Nell'ambito della presente formulazione ridotta monodimensionale a interfaccia liscia, il framework fornisce un ponte fisicamente interpretabile tra misure sperimentali, informazioni CFD sparse, campi di chiusura efficace inferiti e chiusure algebriche di turbolenza compatibili con un solutore.
Reduced-order turbulence-closure discovery in smooth-interface stratified gas-liquid flow using a hybrid inverse physics-informed neural-network framework
Maldar Mohammadabadi, Ali
2025/2026
Abstract
Accurate reduced-order prediction of stratified gas-liquid flow in horizontal pipes remains strongly limited by the closure of unresolved wall and interfacial momentum-transfer mechanisms. Classical one-dimensional two-fluid models are computationally efficient and physically interpretable, but their accuracy depends on empirical closure relations for wall shear, interfacial shear, and turbulent momentum transport. Experiments provide reliable global quantities such as pressure gradient, void fraction, liquid holdup, and superficial velocities, while CFD can provide local flow information but remains computationally expensive and model-dependent. This work develops a hybrid inverse physics-informed neural-network framework for reduced-order turbulence-closure discovery in smooth, fully developed stratified air-water pipe flow. The closure correlations are derived from eight operating conditions with Re_L = 3450-5310 and Re_g = 2444-10134, and are therefore intended for this smooth-interface reduced-flow envelope rather than as universal turbulence laws. The proposed method combines experimental operating quantities, sparse CFD-derived velocity samples, and reduced phase-wise momentum equations. Separate gas and liquid neural networks infer the nondimensional axial midplane velocity profiles and the corresponding effective turbulent-viscosity fields. In this role, the framework can be interpreted as a physics-informed virtual sensor, estimating hidden closure-relevant quantities from measurable operating conditions, CFD-derived local velocity samples, and governing physical constraints. The inverse problem is constrained by the measured pressure-gradient forcing, wall no-slip conditions, phase mean-flow constraints, and dimensional velocity and shear-stress continuity at the gas-liquid interface. As a consequence of the reduced-order midplane formulation and the data-assisted inverse procedure, the inferred turbulent viscosity should be regarded as an effective momentum-transport model for the present stratified-flow configuration, rather than as a universal three-dimensional eddy-viscosity model. The inferred closure fields reveal distinct phase-dependent structures. In the liquid phase, the effective turbulent viscosity increases monotonically from the wall toward the interface and can be represented by a compact wall-to-interface algebraic form. In the gas phase, the inferred profiles exhibit near-wall damping, an interior peak, and a finite interfacial value, motivating a blended closure based on wall-side and interface-side contributions. The resulting algebraic closures are correlated with Reynolds-number-based variables and then inserted into an independent finite-difference reduced solver. The solver reconstructs coupled gas-liquid velocity profiles while satisfying the reduced momentum equations, mean-flow constraints, and interface-continuity conditions without using neural-network outputs or PINN residual minimization. The results demonstrate that physics-informed inversion can be used not only for sparse-data velocity reconstruction, but also as a closure-discovery tool for reduced-order stratified-flow modeling. Within the present smooth-interface, one-dimensional reduced formulation, the framework provides a physically interpretable bridge between experimental measurements, sparse CFD information, inferred effective closure fields, and solver-compatible algebraic turbulence closures.| File | Dimensione | Formato | |
|---|---|---|---|
|
Thesis.pdf
non accessibile
Descrizione: Thesis
Dimensione
6.52 MB
Formato
Adobe PDF
|
6.52 MB | Adobe PDF | Visualizza/Apri |
|
Executive_Summary.pdf
non accessibile
Descrizione: Executive Summary
Dimensione
1.09 MB
Formato
Adobe PDF
|
1.09 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/261507