The continuous pursuit of Aerodynamic Shape Optimization (ASO) in modern railway design requires thousands of evaluations, a process severely hindered by the prohibitive computational costs and the simulation time of traditional CFD. To overcome this bottleneck, this thesis proposes an innovative, geometric-feature-free Deep Learning surrogate model designed to instantaneously predict the drag coefficient Cd and surface pressure fields. Leveraging a Graph Attention Network Autoencoder (GAT-AE), the model natively processes the complex topologies of 3D unstructured meshes, entirely bypassing the need for rigid parameterizations or manual feature engineering.The framework was trained on an extensive offline dataset comprising 500 high-speed and 300 regional train geometries, rigorously simulated using OpenFOAM. To ensure scalability across industrial-grade meshes, an auxiliary "Scout" network performs a physics-guided hierarchical subsampling, dynamically retaining critical aerodynamic regions characterized by high pressure gradients. The GAT-AE subsequently compresses the physical domain into a low-dimensional latent space, which a Multi-Layer Perceptron (MLP) exploits to accurately map the total drag. A rigorous generalization test demonstrated the pipeline's exceptional robustness across both the streamlined attached flows of high-speed trains and the chaotic bluff-body wakes of regional configurations, effectively bounding the predictive error within the CFD solver's inherent aleatoric uncertainty. Capitalizing on this reliability, the surrogate was coupled with an optimization algorithm to drive an automated geometric shape optimization, successfully generating highly efficient frontal profiles. By slashing inference times from 3 hours to merely 6 seconds and drastically reducing the energy footprint of online evaluations, this methodology enables an extensive, rapid, and radically sustainable design space exploration.
La ricerca verso l'Ottimizzazione Aerodinamica della geometria (ASO) nel design ferroviario moderno richiede migliaia di valutazioni, un processo ostacolato dai costi computazionali e dalle tempistiche proibitive dettate dalle tradizionali CFD. Per superare questo limite, questa tesi propone un innovativo modello surrogato geometric-feature-free basato sul Deep Learning, progettato per prevedere istantaneamente il coefficiente di resistenza Cd e i campi di pressione superficiale. Sfruttando un'architettura Autoencoder basata su Graph Attention Networks (GAT-AE), il modello elabora nativamente le complesse topologie delle mesh 3D non strutturate, eliminando del tutto la necessità di ricorrere all'estrazione manuale di feature ingegnerizzate. Il framework è stato addestrato su un dataset offline comprendente 500 geometrie di treni ad alta velocità e 300 regionali, simulate tramite OpenFOAM. Per garantire la scalabilità su mesh di livello industriale, una rete ausiliaria "Scout" esegue un sottocampionamento guidato dalla fisica, preservando dinamicamente le regioni aerodinamiche critiche. Successivamente, il GAT-AE comprime la fisica in uno spazio latente a bassa dimensionalità, mappato sul Cd globale tramite un Multi-Layer Perceptron (MLP). Un rigoroso test di generalizzazione ha dimostrato l'eccezionale robustezza della pipeline sia sui flussi adesi dell'alta velocità, sia sulle complesse scie da bluff-body dei regionali, contenendo l'errore strettamente entro l'incertezza aleatoria intrinseca del solutore. Capitalizzando tale affidabilità, il surrogato è stato accoppiato a un algoritmo per l'ottimizzazione automatica della forma, generando profili altamente efficienti. Riducendo il tempo di inferenza da 3 ore a soli 6 secondi e abbattendo l'impronta energetica computazionale, questa metodologia abilita un'esplorazione dello spazio di progettazione rapida, estensiva e radicalmente sostenibile.
Deep learning geometric-feature-free surrogate model for train aerodynamics
Di Guida, Alessio
2025/2026
Abstract
The continuous pursuit of Aerodynamic Shape Optimization (ASO) in modern railway design requires thousands of evaluations, a process severely hindered by the prohibitive computational costs and the simulation time of traditional CFD. To overcome this bottleneck, this thesis proposes an innovative, geometric-feature-free Deep Learning surrogate model designed to instantaneously predict the drag coefficient Cd and surface pressure fields. Leveraging a Graph Attention Network Autoencoder (GAT-AE), the model natively processes the complex topologies of 3D unstructured meshes, entirely bypassing the need for rigid parameterizations or manual feature engineering.The framework was trained on an extensive offline dataset comprising 500 high-speed and 300 regional train geometries, rigorously simulated using OpenFOAM. To ensure scalability across industrial-grade meshes, an auxiliary "Scout" network performs a physics-guided hierarchical subsampling, dynamically retaining critical aerodynamic regions characterized by high pressure gradients. The GAT-AE subsequently compresses the physical domain into a low-dimensional latent space, which a Multi-Layer Perceptron (MLP) exploits to accurately map the total drag. A rigorous generalization test demonstrated the pipeline's exceptional robustness across both the streamlined attached flows of high-speed trains and the chaotic bluff-body wakes of regional configurations, effectively bounding the predictive error within the CFD solver's inherent aleatoric uncertainty. Capitalizing on this reliability, the surrogate was coupled with an optimization algorithm to drive an automated geometric shape optimization, successfully generating highly efficient frontal profiles. By slashing inference times from 3 hours to merely 6 seconds and drastically reducing the energy footprint of online evaluations, this methodology enables an extensive, rapid, and radically sustainable design space exploration.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_DiGuida.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
9.53 MB
Formato
Adobe PDF
|
9.53 MB | Adobe PDF | Visualizza/Apri |
|
2026_07_DiGuida_Executive_Summary.pdf
non accessibile
Descrizione: Executive Summary
Dimensione
2.75 MB
Formato
Adobe PDF
|
2.75 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/10589/259678