# INC Dataset: Kuramoto-Sivashinsky and Backward-Facing Step Flow Trajectories for Neural PDE Solver Correction

## Facts

- Provider: Thuerey group, Technical University of Munich
- License: Apache-2.0 (https://huggingface.co/datasets/thuerey-group/INC_Data)
- Size: 11,449 files
- Format: JSON, NumPy (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/thuerey-group/INC_Data
- Access: public, readable without an account
- Category: Physics
- Source: https://openreview.net/forum?id=s3Uk3lrfjy
- Page: https://scigantic.com/archive/6d821f4d-26f7-574e-bc90-06d789294375
- Keywords: Physics, PDE simulation, JSON, NumPy, Hugging Face

## About

Simulation data released by the Thuerey group (TU Munich) with INC, an indirect neural corrector for autoregressive hybrid PDE solvers (NeurIPS 2025), for training and evaluating learned corrections to coarse numerical solvers. Two systems are included. The Kuramoto-Sivashinsky equation (1D chaotic dynamics, periodic, 64 grid points, dt = 0.01) has 27 training, 3 validation and 6 test trajectories of 10,001 steps with varying domain sizes, stored as PyTorch .pth dictionaries. The backward-facing step (BFS) is 2D incompressible Navier-Stokes flow in a five-block channel geometry at Reynolds numbers 1300 to 1400 and step heights 0.85 to 1.0, downsampled to about 128 x 32 and saved every 0.1 time units as JSON metadata plus NPZ tensors per step (velocity, pressure, vertex coordinates, boundary conditions), with 3 training configurations of about 801 steps, 1 validation and a 3,001-step test run. A Results folder holds uncorrected baseline and INC-corrected rollouts for BFS. Layout is KS/Dataset/{train,valid,test} and BFS/Dataset/{train,valid,test}; about 3 GB. Use with the tum-pbs/INC code and PICT solver. Released under Apache-2.0 by the authors.

## Load the data

```python
from huggingface_hub import hf_hub_download
path = hf_hub_download(repo_id="thuerey-group/INC_Data", filename="BFS/Dataset/test/Re1400_Base20-2_Step1.0_Dt0.1/DownBFSdomain_3000.npz", repo_type="dataset")
print(path)
```

## Data structure

File types in sampled files: JSON, NumPy

Example file: `BFS/Dataset/test/Re1400_Base20-2_Step1.0_Dt0.1/DownBFSdomain_3000.npz`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/6d821f4d-26f7-574e-bc90-06d789294375
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/6d821f4d-26f7-574e-bc90-06d789294375
- Over MCP: `get_schema_card` with id `6d821f4d-26f7-574e-bc90-06d789294375` at https://api.scigantic.com/mcp

## Related topics

- [JSON datasets](https://scigantic.com/topics/json)
- [Physics datasets](https://scigantic.com/topics/physics)
- [NumPy datasets](https://scigantic.com/topics/numpy)
