Sovereign 8s.is mirror of standardgalactic/Gen4Turbulence
  • Jupyter Notebook 99.7%
  • Python 0.3%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-02-06 15:53:32 -05:00
1_superresolution First Release 2026-02-06 15:53:32 -05:00
2_forecasting First Release 2026-02-06 15:53:32 -05:00
3_flow_reconstruction First Release 2026-02-06 15:53:32 -05:00
.DS_Store First Release 2026-02-06 15:53:32 -05:00
.gitignore First Release 2026-02-06 15:53:32 -05:00
LICENSE Initial commit 2025-08-19 22:02:39 -04:00
README.md First Release 2026-02-06 15:53:32 -05:00
requirements.txt First Release 2026-02-06 15:53:32 -05:00

Learning Turbulent Flows with Generative Models

This code is part of the article: "Learning Turbulent Flows with Generative Models for Super Resolution and Sparse Flow Reconstruction", accepted in Nature Communications.

Neural operators are promising surrogates for dynamical systems but when trained with standard L² losses they tend to oversmooth fine-scale turbulent structures. Here, we show that combining operator learning with generative modeling overcomes this limitation. We consider three practical turbulent-flow challenges where conventional neural operators fail: 1) spatio-temporal super-resolution, 2) forecasting, and 3) sparse flow reconstruction. Check out our Project page for more details.

Requirements

  • Python 3.9.16
  • Install dependencies: pip install -r requirements.txt

Repository Structure

The three main tasks are organized as three main directories in the root. Each directory has a data folder. Before training, please copy the correct data files for each task into the appropriate data folder from Zenodo.

Task 1: Super-resolution

The different models compared in this task are located in:

  • NO: 1_superresolution/no
  • adv-NO: 1_superresolution/no/adv_no
  • NO+VAE: 1_superresolution/no/vae
  • NO+GAN: 1_superresolution/no/gan
  • NO+DM: 1_superresolution/no/dm

Task 2: Forecasting

The different models compared in this task are located in:

  • NO: 2_forecasting/no
  • adv-NO: 2_forecasting/no/adv_no
  • NO+div: 2_forecasting/no_div
  • GenCFD: 2_forecasting/GenCFD (follow the README provided in GenCFD for installation)

Task 3: Flow Reconstruction

The different models compared in this task are located in:

  • GAN: 3_flow_reconstruction/no/adv_training
  • Diffusion Model: 3_flow_reconstruction/dm

Training

To train a model, navigate to the corresponding model directory and run:

python -u train_*.py

Special case for GenCFD: Run the following script:

2_forecasting/GenCFD/scripts/train_hit3d.sh

Post-processing

The postprocess.ipynb notebook in each model's directory loads the trained model (also provided) and saves the prediction.

Special case for GenCFD: Run 2_forecasting/GenCFD/colab/postprocess.ipynb

Model Comparison

To compare the different models for each task, run the comparison.ipynb notebook in each of the three task directories.

Citing This Work

If you find this work useful, please cite:

@article{oommen2025learning,
      author={Oommen, Vivek and Khodakarami, Siavash and Bora, Aniruddha and Wang, Zhicheng and Karniadakis, George Em},
      title={Learning Turbulent Flows with Generative Models: Super-resolution, Forecasting, and Sparse Flow Reconstruction},
      journal={arXiv preprint arXiv:2509.08752},
      year={2025}
      }

License

This project is licensed under the MIT License. See the LICENSE file for details.

Acknowledgments

We acknowledge the following repositories:

  1. Real-Esrgan - https://github.com/lizhuoq/Real-Esrgan
  2. MedicalNet - https://github.com/Tencent/MedicalNet
  3. denoising-diffusion-pytorch - https://github.com/lucidrains/denoising-diffusion-pytorch/tree/main