This thesis focuses on the application of data-driven methods to the prediction of the aerodynamic coefficients of a 3D front wing for automotive applications, a problem traditionally tackled by computational fluid dynamics. Three different neural network architectures were used to predict the pressure coefficient and the wall shear stress on the surface of a parametrized 3D wing running in close proximity to the ground. The machine learning models used were DoMINO and X-MeshGraphNet, both developed by NVIDIA, and AB-UPT, developed at JKU Linz. These 3 models each represent a different approach to neural networks, being based on convolutional neural networks, graph neural networks, and transformers respectively. The training and testing data for the models was generated ad-hoc for this thesis, by performing CFD simulations of variations of the wing, with changes to the planform, the angle of attack, the distance from the ground, and the profile shape. The models achieved a mean Cl prediction error of 6.7% for DoMINO, 10.96% for X-MeshGraphNet, and 1.96% for AB-UPT. In addition, AB-UPT was used to predict the flow velocity and pressure coefficient in a limited size volume in close proximity to the wing, and was also tested, but not trained, on an additional set of wing variations created using profile shapes obtained by interpolating the ones used previously. The mean Cl prediction error achieved was 2.02%. The overall goal of this thesis is creating a foundation for further development of methodologies that could accelerate the design process of the aerodynamic appendages of a racecar. The entire work of the thesis was performed at BMW M Motorsport, with support from the BMW Group IT.
Questa tesi si focalizza sull'applicazione di metodi data-driven per la predizione dei coefficienti aerodinamici di un'ala anteriore tridimensionale per applicazioni automobilistiche, un problema tradizionalmente affrontato mediante la fluidodinamica computazionale. Sono state utilizzate tre diverse architetture di reti neurali per predire il coefficiente di pressione e lo sforzo a parete sulla superficie di un'ala 3D altamente parametrizzata operante in prossimità del suolo. I modelli di machine learning impiegati sono DoMINO e X-MeshGraphNet, entrambi sviluppati da NVIDIA, e AB-UPT, sviluppato presso la JKU Linz. Questi tre modelli rappresentano ciascuno un approccio differente alle reti neurali, essendo basati rispettivamente su reti neurali convoluzionali, reti neurali grafiche, e transformer. I dati utilizzati per il training e il test dei modelli sono stati generati ad-hoc per questa tesi, effettuando simulazioni CFD di varianti dell'ala, con modifiche alla forma in pianta, all'angolo di attacco, alla distanza dal suolo e alla forma del profilo alare. I modelli hanno ottenuto un errore medio nella predizione del Cl del 6.7% per DoMINO, del 10.96% per X-MeshGraphNet e dell'1.96% per AB-UPT. Inoltre, AB-UPT è stato utilizzato per predire la velocità del flusso e il coefficiente di pressione in un volume di dimensioni limitate in prossimità dell'ala, ed è stato testato, ma non allenato, su un set aggiuntivo di varianti dell'ala create utilizzando profili alari ottenuti interpolando quelli utilizzati in precedenza. L'errore medio ottenuto nella predizione del Cl è del 2.02%. L'obiettivo di questa tesi è creare una base per l'ulteriore sviluppo di metodologie che possano accelerare il processo di progettazione delle appendici aerodinamiche di un'auto da corsa. L'intero lavoro di tesi è stato svolto presso BMW M Motorsport, con il supporto di BMW Group IT.
Data-Driven prediction of flow variables on a 3D wing in ground effect
Cenzato, Andrea
2025/2026
Abstract
This thesis focuses on the application of data-driven methods to the prediction of the aerodynamic coefficients of a 3D front wing for automotive applications, a problem traditionally tackled by computational fluid dynamics. Three different neural network architectures were used to predict the pressure coefficient and the wall shear stress on the surface of a parametrized 3D wing running in close proximity to the ground. The machine learning models used were DoMINO and X-MeshGraphNet, both developed by NVIDIA, and AB-UPT, developed at JKU Linz. These 3 models each represent a different approach to neural networks, being based on convolutional neural networks, graph neural networks, and transformers respectively. The training and testing data for the models was generated ad-hoc for this thesis, by performing CFD simulations of variations of the wing, with changes to the planform, the angle of attack, the distance from the ground, and the profile shape. The models achieved a mean Cl prediction error of 6.7% for DoMINO, 10.96% for X-MeshGraphNet, and 1.96% for AB-UPT. In addition, AB-UPT was used to predict the flow velocity and pressure coefficient in a limited size volume in close proximity to the wing, and was also tested, but not trained, on an additional set of wing variations created using profile shapes obtained by interpolating the ones used previously. The mean Cl prediction error achieved was 2.02%. The overall goal of this thesis is creating a foundation for further development of methodologies that could accelerate the design process of the aerodynamic appendages of a racecar. The entire work of the thesis was performed at BMW M Motorsport, with support from the BMW Group IT.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_Cenzato_Andrea.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Tesi Cenzato Andrea
Dimensione
34.7 MB
Formato
Adobe PDF
|
34.7 MB | Adobe PDF | Visualizza/Apri |
|
Executive_Summary_Cenzato_Andrea.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Executive Summary Cenzato Andrea
Dimensione
3.02 MB
Formato
Adobe PDF
|
3.02 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/10589/250741