File size: 3,262 Bytes
752938e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 | ---
license: cc-by-nc-4.0
tags:
- physics
- computational-fluid-dynamics
- rans
- graph-neural-networks
pretty_name: Steady-RANS cross-family generalization dataset
---
# Steady-RANS cross-family generalization dataset
Data for the paper *"Towards generalized flow field prediction: one model across unseen
object families"* (under double blind review; this account is anonymous for that reason).
Trained checkpoints and evaluation code are in the companion model repo:
[steady-rans-surrogates](https://huggingface.co/BlidReview/steady-rans-surrogates).
Steady incompressible k-omega SST (OpenFOAM simpleFoam) external flow around 855 distinct
shapes (17 scripted parametric families plus 40 ModelNet object categories), three
log-uniform Reynolds numbers in [500, 1e5] per shape, random yaw 0 to 20 degrees. All
solves are nondimensional (U_inf = rho = L = 1).
## Contents
| Folder | Size | What |
|---|--:|---|
| `cache_v3/` | ~13 GB | 2546 processed PyG graphs (.pt), one per converged case: ~10^4 sampled nodes with 12 node features, kNN edges (k=8, symmetrized), and the 7 target fields (u, v, w, p, and transformed k, omega, nu_t) at every node. This is the exact training and evaluation input; the deterministic train / val / OOD split is derived from it in code. |
| `ext_dataset_of/` | ~0.5 GB | The 9 zero shot external geometry families (AhmedML, WindsorML, DrivAerML, DARPA SUBOFF, NASA-CRM, motorbike, swept wings, slotted plates, CAARC buildings), re-simulated through the same pipeline at matched conditions: per case meta.json plus VTK exports (boundary and internal fields), the zero shot ground truth used in the paper. |
| `dataset_stls/` | ~0.4 GB | The 855 watertight training shape STLs, for re-meshing or extending the corpus. |
| `dataset_of/` | ~35 GB | Raw OpenFOAM cases for the training corpus (meta.json + exported fields per case), for full pipeline reproducibility. |
## Splits
The split is deterministic and byte identical across architectures and seeds; it is
computed by `split()` in `code/ezflow_v3/gnn/train_v5.py` of the model repo. Held out:
six whole ModelNet categories (car, airplane, bottle, cone, chair, lamp; 267 cases,
89 shapes) on the geometry axis, and the Reynolds band [20000, 50000] (375 cases) on the
flow axis, checked in that order; a random 10 percent of the remainder is validation
(198 cases); 1706 cases train.
## Use
```bash
huggingface-cli download BlidReview/steady-rans-generalization --local-dir ./data --repo-type dataset
# then, from the model repo:
python easy_eval.py --cache ./data/cache_v3 --weights ./weights
```
Graph field layout is defined in `code/ezflow_v3/gnn/etl.py` (features) and
`code/ezflow_v3/gnn/features.py` (target transforms) of the model repo.
## Provenance and license
All cases were generated by the authors with OpenFOAM. The external zero shot geometries
are derived from public benchmark shapes (AhmedML, WindsorML, DrivAerML, DARPA SUBOFF,
NASA-CRM, the OpenFOAM motorbike tutorial, CAARC); we release only our own simulations of
them at our conditions, not the original benchmark data. License: CC BY-NC 4.0
(non-commercial, attribution). Surrogates trained on this data are approximations; do not
use them as the sole basis for safety critical decisions.
|