---
pretty_name: DiffPhyCon 2D Jellyfish Dataset
license: other
task_categories:
- reinforcement-learning
- unconditional-image-generation
---

# DiffPhyCon 2D Jellyfish Dataset

This repository is an archival mirror of the official 2D Jellyfish dataset
released by **AI4Science-WestlakeU** for the NeurIPS 2024 paper
**“DiffPhyCon: A Generative Approach to Control Complex Physical Systems.”**

The dataset was not generated or authored by `bcTiann`. It is mirrored here to
support reproducible DiffPhyCon baseline and Flow Matching experiments. The
data values have not been modified; only lossless tar sharding was applied to
make storage and transfer through the Hugging Face Hub reliable.

## Original source

- Official repository: https://github.com/AI4Science-WestlakeU/diffphycon
- Original dataset and checkpoint download linked by the authors:
  https://drive.google.com/drive/folders/1_ECmMZ77Lm02znhQ72MvwKe1MqGQquzU
- Paper: https://openreview.net/forum?id=MbZuh8L0Xg

Please cite the original DiffPhyCon paper and credit its authors when using
this dataset.

## Contents

- `train_data`: 30,000 simulations.
- `test_data`: 200 simulations.
- Per-simulation fluid states, forces, boundaries, opening-angle controls, and merged boundary representations.
- Training and test normalization metadata.

To avoid Hub commit-rate limits from roughly 150,000 source files, the files are
stored in lossless tar shards:

- `shards/train_states_*.tar`: the 30,000 training-state files, split into 2,000 simulations per shard.
- `shards/train_auxiliary.tar`: training forces, boundaries, controls, merged masks, and normalization metadata.
- `shards/test_data.tar`: the complete 200-simulation test split.
- `SHA256SUMS`: archive integrity checksums.

Extracting the archives recreates the original `train_data` and `test_data`
directory layout. The main source subdirectories are `states`, `forces`,
`bdry128`, `bdry64`, `bdry_head_thetas`, and
`bdry_merged_mask_offsets`.

## License and redistribution status

The official DiffPhyCon **source-code repository** is distributed under the
MIT License. However, the Jellyfish dataset and checkpoints are distributed
externally through Google Drive, and no separate dataset license was found in
the official repository README or in the downloaded dataset package.

For that reason, this mirror does **not** assert that the dataset itself is
MIT-licensed, and the Hugging Face metadata uses `license: other`. This public
archival mirror preserves explicit attribution to the original DiffPhyCon
authors; users should consult the original authors before further
redistribution or commercial use.

## Citation

```bibtex
@inproceedings{wei2024diffphycon,
  title     = {DiffPhyCon: A Generative Approach to Control Complex Physical Systems},
  author    = {Wei, Long and Hu, Peiyan and Feng, Ruiqi and Feng, Haodong and
               Du, Yixuan and Zhang, Tao and Wang, Rui and Wang, Yue and
               Ma, Zhi-Ming and Wu, Tailin},
  booktitle = {The Thirty-eighth Annual Conference on Neural Information Processing Systems},
  year      = {2024},
  url       = {https://openreview.net/forum?id=MbZuh8L0Xg}
}
```