The prediction of flow fields around bluff bodies remains a significant computational challenge, particularly during the preliminary design phase. RANS simulations continue to represent the industrial standard due to their computational efficiency. However, because of the excessive numerical diffusion, they are unable to accurately capture the complex unsteady turbulent phenomena that characterise bluff-body flows. The present work aims to overcome these limitations through an innovative multifidelity data-driven framework based on state-of-the-art Geometric Deep Learning architectures. The objective is to develop a model capable of mapping low-fidelity RANS results to flow fields with an accuracy approaching IDDES level of precision. Since no suitable datasets are currently available in the literature, a dedicated one was generated for the present work, involving a surface-mounted cube with variable aspect ratio and side-slip flow angles. The proposed framework is distinguished by its high degree of modularity and flexibility. Two distinct modelling architectures were implemented and compared: one based on Graph Convolutional and the other on global attention mechanisms (Transformers). The results demonstrate that the developed models provide a substantial improvement in predictive accuracy over the original RANS data in nearly all test cases considered. Furthermore, several opportunities for future development identified during this work suggest significant potential for further gains in both efficiency and accuracy. In conclusion, this work outlines a fundamental paradigm shift in CFD. Rather than relying exclusively on the numerical solver and mesh resolution to achieve high accuracy, a fast, low-fidelity CFD solver is now only required to generate a coarse solution. IDDES-equivalent predictive accuracy is then achieved through a neural network acting as a corrector from the inaccurate solution. By leveraging the computational efficiency of deep learning, this approach drastically reduces the need for high-fidelity CFD simulations, enabling an unseen combination of computational efficiency and predictive accuracy.
La predizione del campo di moto attorno a corpi tozzi rappresenta tuttora una sfida computazionale complessa, specialmente nelle fasi di progettazione preliminare. Le simulazioni RANS rimangono lo standard industriale per la loro efficienza; tuttavia, non riescono a catturare con la necessaria accuratezza i fenomeni turbolenti instazionari e complessi caratteristici dei corpi tozzi. Il presente lavoro mira a superare tali limiti attraverso un innovativo framework multifidelity data-driven basato sulle più recenti architetture di Geometric Deep Learning. L'obiettivo è lo sviluppo di un modello in grado di mappare risultati RANS a bassa fedeltà verso campi di moto aventi precisione prossima a risultati IDDES. Data l'assenza in letteratura di dataset idonei, è stato generato un dataset ad-hoc focalizzato sul caso studio di un surface mounted cube con Aspect Ratio variabile, analizzato sotto molteplici angoli di sideslip. Il framework sviluppato si distingue per un'elevata modularità e flessibilità. In particolare, vengono implementate e messe a confronto due architetture: una basata su Graph Convolution e una basata su meccanismi di attenzione globale (Transformers). I risultati ottenuti dimostrano che i modelli sviluppati garantiscono un sensibile incremento della precisione rispetto ai dati RANS di partenza nella quasi totalità dei casi testati. Inoltre, i margini di sviluppo futuro individuati durante la ricerca promettono importanti miglioramenti in termini di efficienza e precisione. In conclusione, questo lavoro delinea un radicale cambio di paradigma nella CFD. Si passa da un approccio tradizionale, in cui l'accuratezza dipende esclusivamente dal solutore e dalla griglia, a una visione in cui il solutore CFD ha il solo compito di produrre un dato di partenza che catturi la fisica macroscopica. La precisione predittiva viene delegata alla rete neurale, che agisce come correttore intelligente. Sfruttando l'efficienza del Deep Learning, questo approccio riduce drasticamente la necessità di eseguire simulazioni CFD ad alta fedeltà garantendo un connubio senza precedenti tra efficienza computazionale e accuratezza predittiva.
Multifidelity framework for flow prediction around bluff bodies
Baio, Matteo
2025/2026
Abstract
The prediction of flow fields around bluff bodies remains a significant computational challenge, particularly during the preliminary design phase. RANS simulations continue to represent the industrial standard due to their computational efficiency. However, because of the excessive numerical diffusion, they are unable to accurately capture the complex unsteady turbulent phenomena that characterise bluff-body flows. The present work aims to overcome these limitations through an innovative multifidelity data-driven framework based on state-of-the-art Geometric Deep Learning architectures. The objective is to develop a model capable of mapping low-fidelity RANS results to flow fields with an accuracy approaching IDDES level of precision. Since no suitable datasets are currently available in the literature, a dedicated one was generated for the present work, involving a surface-mounted cube with variable aspect ratio and side-slip flow angles. The proposed framework is distinguished by its high degree of modularity and flexibility. Two distinct modelling architectures were implemented and compared: one based on Graph Convolutional and the other on global attention mechanisms (Transformers). The results demonstrate that the developed models provide a substantial improvement in predictive accuracy over the original RANS data in nearly all test cases considered. Furthermore, several opportunities for future development identified during this work suggest significant potential for further gains in both efficiency and accuracy. In conclusion, this work outlines a fundamental paradigm shift in CFD. Rather than relying exclusively on the numerical solver and mesh resolution to achieve high accuracy, a fast, low-fidelity CFD solver is now only required to generate a coarse solution. IDDES-equivalent predictive accuracy is then achieved through a neural network acting as a corrector from the inaccurate solution. By leveraging the computational efficiency of deep learning, this approach drastically reduces the need for high-fidelity CFD simulations, enabling an unseen combination of computational efficiency and predictive accuracy.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_Baio_Tesi.pdf
accessibile in internet per tutti
Descrizione: Testo della tesi
Dimensione
26.81 MB
Formato
Adobe PDF
|
26.81 MB | Adobe PDF | Visualizza/Apri |
|
2026_07_Baio_Executive Summary.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
686.9 kB
Formato
Adobe PDF
|
686.9 kB | 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/10589/260632