ayz2/emmi_wing_processed

Emmi-Wing (processed)

This is a processed, downsampled derivative of Emmi-Wing, not the original dataset. Fields were converted to a common frame and non-dimensionalization, rows were randomly subsampled and some columns were dropped (details below). For the original data, see the upstream source.

A preprocessed, downsampled release of Emmi-Wing (upstream: https://huggingface.co/datasets/EmmiAI/Emmi-Wing). Total size is 0.83 TB. Every sample is a float32 point cloud [n_points, n_cols] stored as .npy, in a common frame and non-dimensionalization shared across the campaigns of this release.

Layout

collated/                 surface
  manifest.json           [{stem, n_points, n_points_source, ...}]
  splits/{train,val}.json
  norm_stats*.npz         per-column normalization statistics
  samples/<stem>.npy   (packed in shards/*.tar, see below)
volume_collated/          volume
  manifest.json
  norm_stats_volume*.npz
  samples/<stem>.npy   (packed in shards/*.tar, see below)
RELEASE.json              downsampling factors, columns and row totals

This dataset has ~29k samples per tree, so the per-sample files are packed into ~5 GB uncompressed tar shards to stay within Hub file-count limits. The member paths are relative to the dataset root, so this reproduces the per-sample layout above:

cd <download dir>
for f in collated/shards/*.tar volume_collated/shards/*.tar; do tar -xf "$f"; done

<tree>/shards/index.json maps each stem to its shard. A single sample can be pulled with tar -xf <shard> <tree>/samples/<stem>.npy.

Columns

Surface (collated, 7 columns):

col	name	meaning
0	x	position
1	y	position
2	z	position
3	cp	pressure coefficient
4	cf_x	skin-friction coefficient
5	cf_y	skin-friction coefficient
6	cf_z	skin-friction coefficient

Dropped from this release: rho_tilde.

Volume (volume_collated, 7 columns):

col	name	meaning
0	x	position
1	y	position
2	z	position
3	ux	velocity
4	uy	velocity
5	uz	velocity
6	cp	pressure coefficient

Dropped from this release: rho_tilde, wx, wy, wz.

Downsampling

tree	samples	row factor	rows (release)	rows (collated source)
collated	29,609	1/1	1.014e+10	1.014e+10
volume_collated	29,606	1/5	1.942e+10	9.711e+10

Rows within each file are in random order: they were permuted when the files were written. The release keeps the first n // factor rows of each collated sample, which is a uniform random subsample of the mesh points. For the same reason, any contiguous window arr[s:s+k] is a uniform random crop, so np.load(..., mmap_mode='r') plus a slice is the intended access pattern. The shipped norm_stats*.npz were computed on the full collated trees and remain valid for the subsample.

Splits

The volume tree uses the surface splits (keyed by stem).

• train: 26,648 samples
• val: 2,961 samples

Source dataset and preprocessing

Emmi-Wing

What it is. ~29.7 k steady compressible RANS solutions over parameterized tapered swept wings — the largest campaign here by sample count, and one of two transonic wing sets. Four geometric parameters (root chord, span, taper ratio, sweep; dihedral is also tabulated) plus two inflow parameters (velocity, angle of attack).

Generation. OpenFOAM v2506 rhoSimpleFoam (steady, compressible, perfect gas) with the Spalart-Allmaras closure, second-order schemes (van Leer limiter on momentum, bounded upwind on energy/pressure). Body-fitted snappyHexMesh grids with prismatic boundary layers at y⁺ 50–200, i.e. wall functions — the only wing campaign here that does not resolve the sublayer. ~3.3 M volume points, 46 k–114 k surface points per case (each case is a distinct geometry, so the point count varies).

Parameters. Measured off the raw solutions rather than assumed: p∞ = 100 000 Pa, ρ∞ = 1.16639 kg/m³, T∞ = 298 K with R = 8314.5/28.9 = 287.7 J/(kg·K) (OpenFOAM's default molecular weight), constant across the campaign; only U∞ varies, 150–300 m/s, so q = ½ρ∞U∞² per run. Mach 0.437–0.875, Re 5–20e6, AoA −10…10°. Angle of attack lives in the freestream, not the geometry: the wing is symmetric about z = 0 and the inlet velocity is U[cos α, 0, sin α].

Cost. Not published. 62 cases are flagged as erroneous upstream and excluded, leaving 29 665 valid, of which 29 609 are converted here.

Processing. The raw release is per-run PyTorch tensors (surface_position.pt, surface_pressure.pt in absolute Pa, surface_rho.pt, surface_wall_shear_stress.pt), so Cp/Cf are precomputed in the builder along with rho_tilde = ρ/ρ∞: [x, y, z, cp, cfx, cfy, cfz, rho_tilde] (8 cols), CF_SIGN = −1. An earlier version of this pipeline assumed sea-level p∞ = 101 325 Pa with R = 287.05 and derived T∞ from a = U∞/Mach; that is wrong by 1325 Pa and 1.57% in ρ∞, biasing Cp by ≈ −0.06 and showing up as a stagnation Cp of 1.016 against the compressible expectation of ~1.08. The 29 609 already-collated samples were corrected in place by a per-run affine rescale rather than re-downloaded (recorded in collated/freestream_fix.json), and every volume conversion now re-asserts both constants on the run's own inlet face. Note that the metadata's Mach column is exactly U∞/346.0295, which corresponds to no consistent (R, T) — it is a usable conditioning label and a wrong physical scale. Frame: x chord/streamwise, y span (half-domain, root at y ≈ 0 on a symmetry plane), z thickness/lift. The volume ships position, pressure, density, velocity and vorticity tensors; output is 11 cols [x, y, z, ux, uy, uz, cp, rho_tilde, ωx, ωy, ωz] with velocity non-dimensionalized as u/U∞ and vorticity as ω·c_root/U∞, since the freestream varies per run. The mesh is so wall-clustered (99% of points within r < 2.6 m of a ±17.4 m cube) that the crop keeps 96.0–99.6%; it exists only to share the surface's frame. Surface Cf is not pruned for this campaign: the tail is continuous and physical — the cf_y outliers are the wing tip and the cf_z outliers the leading edge — so a 3σ cut would be a normalization trade, not a fix. Splits: random 90/10 (26 648/2961).

Citation

TODO: this work, and the upstream dataset's citation.