MegaFlow2D

MegaFlow2D is a collection of parameterized two-dimensional computational fluid dynamics simulations for machine learning research. It pairs low- and high-resolution simulation results so researchers can study super-resolution and discretization-error correction. The accompanying paper describes more than two million snapshots from 3,000 external- and internal-flow configurations.

What is in this repository

The data are distributed as five parts of one ZIP archive, data.zip.001 through data.zip.005. They are not five independent datasets. The archive contains the raw simulation data used by the MegaFlow2D Python package, which processes paired mesh resolutions into graph data and stores the processed results in HDF5. The Hugging Face dataset viewer cannot preview this split ZIP archive.

The package supports circle and ellipse subsets, a mixed subset, or the full dataset. Its graph samples contain paired low- and high-resolution results. See the source repository for the data loader, processing code, and a model-training example.

Getting started

The easiest route is to let the Python package download and process the archive:

pip install MegaFlow2D

from megaflow.dataset.MegaFlow2D import MegaFlow2D

dataset = MegaFlow2D(
    root="/path/to/megaflow-data",
    download=True,
    transform=None,
    pre_transform=None,
    split_scheme="mixed",
    split_ratio=[0.5, 0.5],
)
low_resolution, high_resolution = dataset.get(0)
print(low_resolution.num_nodes, high_resolution.num_nodes)

Run the processing code in a Python script guarded by if __name__ == "__main__":, as shown in the project's example, because processing uses multiple processes. The initial download and processing require substantial disk space and may take hours. Set download=False on later runs to reuse an existing local copy.

To download the raw files yourself, keep all five parts in the same directory and concatenate them in numeric order before extracting the resulting ZIP archive. The package's download method performs those steps automatically.

Some processing utilities require FEniCS/dolfin and are intended for Linux or Windows Subsystem for Linux. See the installation notes before setting up a minimal environment.

Intended use and limitations

MegaFlow2D is intended for research on machine learning methods for two-dimensional CFD, especially multi-fidelity learning, super-resolution, and error correction. The simulations cover the geometries and boundary conditions described in the paper; performance on other flow regimes or geometries should be evaluated separately. The archive is provided in a split ZIP format, so Hub preview and direct row-based loading are unavailable.

Citation

If you use MegaFlow2D, cite the paper:

@inproceedings{xu2023megaflow2d,
  author = {Xu, Wenzhuo and Grande Gutierrez, Noelia and McComb, Christopher},
  title = {MegaFlow2D: A Parametric Dataset for Machine Learning Super-Resolution in Computational Fluid Dynamics Simulations},
  booktitle = {Proceedings of Cyber-Physical Systems and Internet of Things Week 2023},
  year = {2023},
  pages = {100--104},
  doi = {10.1145/3576914.3587552},
  url = {https://doi.org/10.1145/3576914.3587552}
}

The dataset on this Hub page is marked Apache-2.0. The Python package source is separately licensed under MIT.