Blame for README.md

---
license: cc-by-4.0
task_categories:
- time-series-forecasting
size_categories:
- n<1K
tags:
- computational-fluid-dynamics
- graph-neural-networks
- unstructured-mesh
- openlb
- hdf5
pretty_name: OpenLB MultiGeometry Flow 456
---

# OpenLB MultiGeometry Flow 456

OpenLB MultiGeometry Flow 456 is an openly available collection of 456
independent two-dimensional unsteady-flow trajectories around eight obstacle
families on variable triangular meshes. It is created and maintained by
[`DingDong1921`](https://huggingface.co/DingDong1921).

## Scale and splits

- 408 train, 24 validation, and 24 test trajectories.
- 57 trajectories for each of eight families: ellipse, triangle, rectangle,
  pentagon, smooth star, irregular, bunny, and elephant.
- 601 frames per trajectory over dimensionless time `0:0.1:60`.
- 2,831--4,365
  mesh nodes per trajectory (mean 3580.2).
- 28.774 GiB of HDF5 payload.
- Reynolds numbers 500--1000 in increments of 50.
- Channels `(u, v, p)`: two velocity components and recovered
  density-normalized specific pressure.

The three splits are predefined for reproducible comparisons. Test files are
publicly accessible, so this is a conventional held-out split rather than a
secret benchmark. When reporting a clean test result, choose methods and
hyperparameters using only train and validation data.

All eight geometry families occur in training. The supplied split supports
within-family geometry/parameter generalization studies; it does not establish
unseen-family generalization.

## Download

Download the complete dataset:

```bash
hf download DingDong1921/openlb-multigeometry-flow \
  --repo-type dataset \
  --local-dir openlb-multigeometry-flow
```

Download one split:

```bash
hf download DingDong1921/openlb-multigeometry-flow \
  --repo-type dataset \
  --include "data/test/*" \
  --local-dir openlb-multigeometry-flow
```

Inspect one trajectory after download:

```bash
python examples/load_one_case.py data/train/ellipse_00.h5 --frame 0
```

See `metadata/schema.md`, `metadata/provenance.md`, and
`metadata/cases.jsonl` before constructing a loader. Each manifest row records
the split, geometry, physical/numerical parameters, source-code identities,
file size, and SHA-256 identity.

## Representative previews

| Family | Shared-scale frame at `t*=30` | Full trajectory |
| --- | --- | --- |
| Ellipse | ![](previews/ellipse_00_t30.png) | [GIF](previews/ellipse_00_speed_pressure.gif) |
| Triangle | ![](previews/triangle_00_t30.png) | [GIF](previews/triangle_00_speed_pressure.gif) |
| Rectangle | ![](previews/rectangle_00_t30.png) | [GIF](previews/rectangle_00_speed_pressure.gif) |
| Pentagon | ![](previews/pentagon_00_t30.png) | [GIF](previews/pentagon_00_speed_pressure.gif) |
| Smooth star | ![](previews/smooth_star_00_t30.png) | [GIF](previews/smooth_star_00_speed_pressure.gif) |
| Irregular | ![](previews/irregular_00_t30.png) | [GIF](previews/irregular_00_speed_pressure.gif) |
| Bunny | ![](previews/bunny_00_t30.png) | [GIF](previews/bunny_00_speed_pressure.gif) |
| Elephant | ![](previews/elephant_00_t30.png) | [GIF](previews/elephant_00_speed_pressure.gif) |

## Numerical semantics and limitations

- Solver: OpenLB 1.9.0, D2Q9/TRT, cold start, fifth-order inlet/farfield ramp,
  convection outlet, and no-slip obstacle wall.
- Pressure is reconstructed from the unsteady Navier--Stokes relation and fixed
  to zero at the right outlet. It is not a direct raw LBM pressure field.
- The inlet/farfield ramp ends at dimensionless time `t*=15.025`; early windows
  should not be treated as stationary-boundary data.
- The data are generated by OpenLB and are not interchangeable with Nektar++
  trajectories without a matched-semantics comparison.
- Representative images are navigation aids. Visual plausibility alone does
  not establish physical correctness.

## License and attribution

The generated dataset payload, dataset-specific metadata, documentation, and
previews are released under
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Cite or attribute
the dataset to `DingDong1921` and link this repository. OpenLB, DGN4CFD, and
source geometry assets retain their own licenses and are acknowledged in
`metadata/provenance.md`.