Graph-Based Neural Models for Transonic Aerodynamics: AeroFormer & MeshGAT Architectures

Fuente: Zenodo
Salvato in:
Dettagli Bibliografici
Autore principale: Hussain, Saad
Natura: Recurso digital
Lingua:inglese
Pubblicazione: Zenodo 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866901613083361280
author Hussain, Saad
author_facet Hussain, Saad
contents <p>This repository contains the source code, model architectures, training/inference scripts, and environment files used in the manuscript <em>"</em><em>Transonic Aerodynamic </em><em>Predictions</em><em> with Sparse Edge-Augmented Transformers and Graph Attention Networks: A Comparative Study</em><em>"</em>, currently under peer review.</p> <p>It includes implementations of two models:</p> <ul> <li> <p>MeshGAT – a GAT-enhanced MeshGraphNet</p> </li> <li> <p>AeroFormer – a flow-adaptive sparse attention Edge-augmented Graph Transformer</p> </li> </ul> <p>The dataset structure, training logs, inference outputs, and environment specifications (YAML and requirements.txt) are provided to ensure reproducibility.</p> <p>This repository extends v1.0 with two key enhancements while preserving all original functionality: (1) K-fold cross-validation support via <code>Cross_Valid.py</code> for model's statistical robustness, generating per-fold performance metrics and checkpoints; (2) A new <code>Data_Utils.py</code> module that centralizes dataset handling (Mach/AoA parsing, DGL graph loading, and batching logic) to reduce code duplication. The update maintains backward compatibility with v1.0's dataset structure, model architectures (<code>MeshGAT</code> and <code>AeroFormer</code>), and environment specifications with improved reproducibility.</p> <p>*(v1.0 remains available at DOI: 10.5281/zenodo.15583112)*</p> <p>This upload is shared strictly for review purposes. Please contact the author for further clarifications.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_16791813
institution Zenodo
language eng
publishDate 2025
publisher Zenodo
record_format zenodo
spellingShingle Graph-Based Neural Models for Transonic Aerodynamics: AeroFormer & MeshGAT Architectures
Hussain, Saad
Machine Learning
Computational fluid dynamics
<p>This repository contains the source code, model architectures, training/inference scripts, and environment files used in the manuscript <em>"</em><em>Transonic Aerodynamic </em><em>Predictions</em><em> with Sparse Edge-Augmented Transformers and Graph Attention Networks: A Comparative Study</em><em>"</em>, currently under peer review.</p> <p>It includes implementations of two models:</p> <ul> <li> <p>MeshGAT – a GAT-enhanced MeshGraphNet</p> </li> <li> <p>AeroFormer – a flow-adaptive sparse attention Edge-augmented Graph Transformer</p> </li> </ul> <p>The dataset structure, training logs, inference outputs, and environment specifications (YAML and requirements.txt) are provided to ensure reproducibility.</p> <p>This repository extends v1.0 with two key enhancements while preserving all original functionality: (1) K-fold cross-validation support via <code>Cross_Valid.py</code> for model's statistical robustness, generating per-fold performance metrics and checkpoints; (2) A new <code>Data_Utils.py</code> module that centralizes dataset handling (Mach/AoA parsing, DGL graph loading, and batching logic) to reduce code duplication. The update maintains backward compatibility with v1.0's dataset structure, model architectures (<code>MeshGAT</code> and <code>AeroFormer</code>), and environment specifications with improved reproducibility.</p> <p>*(v1.0 remains available at DOI: 10.5281/zenodo.15583112)*</p> <p>This upload is shared strictly for review purposes. Please contact the author for further clarifications.</p>
title Graph-Based Neural Models for Transonic Aerodynamics: AeroFormer & MeshGAT Architectures
topic Machine Learning
Computational fluid dynamics
url https://doi.org/10.5281/zenodo.16791813