Numerical modeling of turbulent airflow with Computational Fluid Dynamics (CFD) requires expertise and a substantial amount of computation power. Yet, CFD simulations are considered vital for assessing novel architectural and urban ventilation concepts. As a potential remedy, image upsampling techniques from computer vision can help in significantly reducing necessary computing times while maintaining high prediction accuracy. Hereby, low resolution CFD simulations are provided as inputs to prediction models for upsampling towards high-fidelity flow fields. Several recent studies focusing on the built environment propose various upsampling approaches using Deep Learning (DL). However, they are either constrained to well specified domains, or they use one of the many existing DL architectures without conclusive findings on how model prediction accuracy relates to airflow domain characteristics. This thesis therefore complements the existing literature by comparing several state-of-the-art DL architectures applied to a range of airflow setups, including forced indoor ventilation, as well as urban air flow. Results reveal that sparsity of the flow domain exhibits a limiting factor for upsampling techniques, even if complemented with physics informed loss terms to maintain physical coherence. On the other hand, pronounced and heterogeneous flow fields can be upsampled well even to a magnification factor of 4, and modern DL architectures based on vision transformers and diffusion models clearly outperform canonical models such as U-Net.
La modellazione numerica del flusso d’aria turbolento mediante fluidodinamica computazionale (CFD) richiede competenze altamente specializzate e un’elevata potenza di calcolo. Nonostante ciò, le simulazioni CFD rimangono strumenti fondamentali per valutare nuovi concetti di ventilazione architettonica e urbana. Un potenziale rimedio consiste nell’impiego di tecniche di upsampling delle immagini, utilizzate nell’ambito di computer vision, che permettono di ridurre sensibilmente i tempi di calcolo mantenendo al contempo un’elevata accuratezza predittiva. In questo approccio, le simulazioni CFD a bassa risoluzione vengono fornite come input a modelli di apprendimento profondo (DL) per generare campi di flusso ad alta fedeltà. Numerosi studi recenti applicati alla progettazione edilizia hanno proposto metodi di upsampling basati sul DL; tuttavia, questi risultano spesso limitati a domini ben specifici oppure si basano su singole architetture senza offrire conclusioni definitive riguardo al legame tra accuratezza predittiva e caratteristiche del dominio di flusso. La presente tesi si inserisce in questo contesto confrontando diverse architetture DL all’avanguardia applicate a una gamma più ampia di configurazioni, comprendenti sia la ventilazione forzata in ambienti interni sia il flusso d’aria urbano. I risultati mostrano che la sparsità del dominio di flusso costituisce un fattore limitante per le tecniche di upsampling, anche se supportate da funzioni di perdita basate sulla fisica, mentre campi di flusso eterogenei e complessi possono essere ricostruiti efficacemente anche con fattori di magnificazione pari a 4. Inoltre, le moderne architetture DL basate su Transformer e modelli di diffusione superano nettamente modelli canonici come U-Net.
Physic-Informed Deep Learning for Super-Resolution of Indoor and Outdoor Airflow Analysis
La Ferla, Vittorio
2024/2025
Abstract
Numerical modeling of turbulent airflow with Computational Fluid Dynamics (CFD) requires expertise and a substantial amount of computation power. Yet, CFD simulations are considered vital for assessing novel architectural and urban ventilation concepts. As a potential remedy, image upsampling techniques from computer vision can help in significantly reducing necessary computing times while maintaining high prediction accuracy. Hereby, low resolution CFD simulations are provided as inputs to prediction models for upsampling towards high-fidelity flow fields. Several recent studies focusing on the built environment propose various upsampling approaches using Deep Learning (DL). However, they are either constrained to well specified domains, or they use one of the many existing DL architectures without conclusive findings on how model prediction accuracy relates to airflow domain characteristics. This thesis therefore complements the existing literature by comparing several state-of-the-art DL architectures applied to a range of airflow setups, including forced indoor ventilation, as well as urban air flow. Results reveal that sparsity of the flow domain exhibits a limiting factor for upsampling techniques, even if complemented with physics informed loss terms to maintain physical coherence. On the other hand, pronounced and heterogeneous flow fields can be upsampled well even to a magnification factor of 4, and modern DL architectures based on vision transformers and diffusion models clearly outperform canonical models such as U-Net.| File | Dimensione | Formato | |
|---|---|---|---|
|
Executive_Summary_Final.pdf
solo utenti autorizzati a partire dal 28/09/2026
Dimensione
1.42 MB
Formato
Adobe PDF
|
1.42 MB | Adobe PDF | Visualizza/Apri |
|
Master_Thesis_Polimi_Final.pdf
solo utenti autorizzati a partire dal 28/09/2026
Dimensione
3.88 MB
Formato
Adobe PDF
|
3.88 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/10589/243494