In this repository, you will find the different python scripts to train the available models on the AirfRANS dataset proposed at the NeurIPS 2022 Datasets and Benchmarks Track conference. You can find the paper here. Disclaimer: An important update correcting an inconsistency in the Machine Learning experiments proposed in the main part of the NeurIPS version of the paper has been done. Please refer to the ArXiv version for the up to date version.
- Python 3.9.12
- PyTorch 1.11.0 with CUDA 11.3
- PyTorch Geometric 2.0.4
- PyVista 0.36.1
- Seaborn 0.11.2
To train a model, run main.py with the desired model architecture:
python main.py GraphSAGE -t full
Note that you must have the dataset in folder Dataset/ at the root of this repository, you can download the dataset here. You can change the parameters of the models and the training in the params.yaml file.
usage: main.py [-h] [-n NMODEL] [-w WEIGHT] [-t TASK] [-s SCORE] model
positional arguments:
model The model you want to train, choose between MLP, GraphSAGE, PointNet, GUNet
optional arguments:
-h, --help show this help message and exit
-n NMODEL, --nmodel NMODEL
Number of trained models for standard deviation estimation (default: 1)
-w WEIGHT, --weight WEIGHT
Weight in front of the surface loss (default: 1)
-t TASK, --task TASK Task to train on. Choose between "full", "scarce", "reynolds" and "aoa"
(default: full)
-s SCORE, --score SCORE
If you want to compute the score of the models on the associated test set.
(default: 0)
For example, if you want to train two MLP models on the scarce data regime task and compute the scores of both models on the associated test set, you will have to run:
python main.py MLP -t scarce -n 2 -s 1
After training a model, the different training plots and training logs will be at the root of the metrics folder. The scores of the trained model on the associated test set are written in the folder scores/TASK/. The visualization of the force coefficients plot and the boundary layers can be generated via the notebook visualization.ipynb. You can find the scores and the plots given in the paper here. The dataset statistics can also be generated via the notebook dataset_stats.ipynb.
Please cite this paper if you use this dataset in your own work.
@inproceedings{
bonnet2022airfrans,
title={Airf{RANS}: High Fidelity Computational Fluid Dynamics Dataset for Approximating Reynolds-Averaged Navier{\textendash}Stokes Solutions},
author={Florent Bonnet and Jocelyn Ahmed Mazari and Paola Cinnella and Patrick Gallinari},
booktitle={Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
year={2022},
url={https://arxiv.org/abs/2212.07564}
}