Turbulent flow remains one of the most challenging problems in classical physics due to its nonlinear and multi-scale nature. Large-eddy simulation (LES) offers a cost-effective alternative to direct numerical simulation (DNS) by resolving the dominant energy-containing motions while modeling the smaller scales. However, the development of accurate closure models for the unresolved terms - particularly in reactive flows - remains a major challenge. In recent years, data-driven approaches and deep learning have gained traction for LES closure, enabled by the availability of large datasets and advances in GPU-accelerated computing. Nevertheless, deep learning methods often struggle to provide accurate predictions across diverse physico-chemical regimes. This work demonstrates the robust generalization capabilities of a Physics-Informed Enhanced Super-Resolution Generative Adversarial Network (PIESRGAN) across a wide spectrum of turbulent combustion regimes. PIESRGAN significantly outperforms traditional closure models (e.g., Dynamic Smagorinsky) in predicting subgrid-scale stresses, reducing the reconstruction error by an order of magnitude in in-sample tests. First, generalization between premixed flames in high- and low-Karlovitz regimes is shown to be feasible when a relevant scale factor is held constant. Second, generalization between reactive and nonreactive jet flames is assessed, with cross-prediction of velocity fields proving successful. The best results are obtained with the high-Karlovitz flame, owing to the similarities between its velocity field and that of the nonreactive jet. Finally, we explore cross-prediction between premixed and nonpremixed jet flames. Notably, PIESRGAN successfully generalizes across these fundamentally different mixing regimes, accurately reconstructing both velocity and scalar fields.
La turbolenza è uno dei problemi più complessi della fisica classica a causa della sua natura non lineare e multi-scala. La Large-Eddy Simulation (LES) rappresenta un’alternativa computazionalmente più accessibile rispetto alla Direct Numerical Simulation (DNS), in quanto risolve le strutture dominanti contenenti energia, mentre modella le scale più piccole. Tuttavia, lo sviluppo di modelli di chiusura accurati per i termini non risolti (in particolare nei flussi reattivi) rimane una sfida significativa. Negli ultimi anni, approcci data-driven e metodi di deep learning, resi possibili dalla disponibilità di ampi dataset e dai progressi nel calcolo accelerato tramite GPU, hanno guadagnato crescente attenzione nell’ambito della chiusura LES. Ciononostante, i metodi di deep learning spesso faticano a fornire previsioni accurate su regimi fisico-chimici diversi da quelli su cui sono stati allenati. Questo lavoro dimostra la capacità di generalizzazione di una Physics-Informed Enhanced Super-Resolution Generative Adversarial Network (PIESRGAN) su un ampio spettro di regimi di combustione turbolenta. PIESRGAN supera significativamente in accuratezza i modelli tradizionali (ad esempio, Dynamic Smagorinsky) nella predizione degli sforzi di sottogriglia, riducendo l’errore di ricostruzione di un ordine di grandezza nei test in-sample. In primo luogo, si mostra come la generalizzazione tra fiamme premiscelate in regimi ad alto e basso numero di Karlovitz sia realizzabile quando si mantiene costante un opportuno fattore di scala. In secondo luogo, si valuta la generalizzazione tra getti reattivi e non reattivi, evidenziando come la previsione incrociata dei campi di velocità risulti efficace. I migliori risultati si ottengono con la fiamma ad alto numero di Karlovitz, grazie alle somiglianze tra il suo campo di velocità e quello del getto non reattivo. Infine, viene esplorata la previsione incrociata tra fiamme a getto premiscelate e non premiscelate. È degno di nota che il modello PIESRGAN riesca a generalizzare con successo attraverso questi regimi di miscelazione fondamentalmente differenti, ricostruendo accuratamente sia i campi di velocità sia i campi scalari.
Data-driven subgrid-scale modeling for turbulent hydrogen flames using physics-informed super-resolution
BRUNACCIONI, LUCA
2024/2025
Abstract
Turbulent flow remains one of the most challenging problems in classical physics due to its nonlinear and multi-scale nature. Large-eddy simulation (LES) offers a cost-effective alternative to direct numerical simulation (DNS) by resolving the dominant energy-containing motions while modeling the smaller scales. However, the development of accurate closure models for the unresolved terms - particularly in reactive flows - remains a major challenge. In recent years, data-driven approaches and deep learning have gained traction for LES closure, enabled by the availability of large datasets and advances in GPU-accelerated computing. Nevertheless, deep learning methods often struggle to provide accurate predictions across diverse physico-chemical regimes. This work demonstrates the robust generalization capabilities of a Physics-Informed Enhanced Super-Resolution Generative Adversarial Network (PIESRGAN) across a wide spectrum of turbulent combustion regimes. PIESRGAN significantly outperforms traditional closure models (e.g., Dynamic Smagorinsky) in predicting subgrid-scale stresses, reducing the reconstruction error by an order of magnitude in in-sample tests. First, generalization between premixed flames in high- and low-Karlovitz regimes is shown to be feasible when a relevant scale factor is held constant. Second, generalization between reactive and nonreactive jet flames is assessed, with cross-prediction of velocity fields proving successful. The best results are obtained with the high-Karlovitz flame, owing to the similarities between its velocity field and that of the nonreactive jet. Finally, we explore cross-prediction between premixed and nonpremixed jet flames. Notably, PIESRGAN successfully generalizes across these fundamentally different mixing regimes, accurately reconstructing both velocity and scalar fields.| File | Dimensione | Formato | |
|---|---|---|---|
|
ExecutiveSummary_LucaBrunaccioni.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
2.04 MB
Formato
Adobe PDF
|
2.04 MB | Adobe PDF | Visualizza/Apri |
|
MasterThesis_LucaBrunaccioni.pdf
accessibile in internet per tutti
Descrizione: Master Thesis
Dimensione
42.22 MB
Formato
Adobe PDF
|
42.22 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/10589/243106