ayz2/windsorml_processed

WindsorML (processed)

This is a processed, downsampled derivative of WindsorML, 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 WindsorML. Total size is 0.72 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
volume_collated/          volume
  manifest.json
  norm_stats_volume*.npz
  samples/<stem>.npy
RELEASE.json              downsampling factors, columns and row totals

Columns

Surface (collated, 9 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
7	cp_var	pressure-coefficient variance
8	yPlus	wall y+

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	p	static pressure (raw solver units; Cp = (p - p_inf) / q)

Downsampling

tree	samples	row factor	rows (release)	rows (collated source)
collated	350	1/1	7.935e+08	7.935e+08
volume_collated	350	1/3	2.472e+10	7.417e+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: 280 samples
• val: 70 samples

Source dataset and preprocessing

WindsorML

What it is. 355 variants of the Windsor body (a squareback automotive reference geometry), released as time-averaged wall-modelled LES. It is the only campaign here with no RANS branch anywhere in the closure.

Generation. Genuine WMLES with the constant-coefficient Vreman SGS model and an equilibrium wall model imposed as a shear-stress constraint with the outer state probed at a fixed 1.5Δ off the wall. Compressible and explicit: nominally fourth-order finite differences with SSP-RK3 time stepping on Cartesian octree grids (unstructured tree-of-cubes, leaf cubes of 4³ or 8³ cells) with an immersed-boundary geometry representation, in Volcano ScaLES. Baseline grid 275 M cells at 0.75 mm minimum spacing. Fields are time-averaged.

Parameters. U∞ = 40 m/s, Re = 2.9e6 on body length (1.044 m); pitching-moment reference length 0.6375 m (wheelbase), reference frontal area 0.112 m² for the baseline geometry.

Cost. One AWS g5.48xlarge node per case (8 × NVIDIA A10G), ~2 min to mesh and ~28 h to solve = 224 GPU-hours per case, ≈ 7.8e4 GPU-hours for the campaign.

Processing. 350 of the 355 runs are present. The surface file stores fields on points and already ships dimensionless coefficients, so no dynamic-pressure conversion is applied: [x, y, z, cp, cfx, cfy, cfz, cpvar, yPlus] (9 cols). The native frame is rotated relative to every other campaign — x is streamwise, y is vertical (floor at y ≈ 0) and z is the lateral/width axis — so WindsorML carries an ORIENT signed permutation (x ← x, y ← −z, z ← y) applied to positions, to the surface Cf vector and to the volume velocity vector, mapping it into the shared loader frame. The volume is a 21 GB base64/zlib inline VTU per run (~325 M points / 291 M hexahedra) whose fields live on cells with no stored centroids, so cell centres are computed explicitly; one read costs ~227 s and ~36 GB RSS. Crop margins in native axes are [0.5, 0.0, 0.75]/[2.0, 0.5, 0.75] L (stream/vertical/width, never extending below the floor), keeping ~79% of cells. Volume output [x, y, z, ux, uy, uz, p], 74.2e9 cells total. One caveat worth recording: the volume tree stores pressure in Pa against a solver reference that matches no documented ambient; regressing wall-adjacent volume pressure on the paired surface Cp gives q = 1016.6 and p∞ = 37 340.5 at R² = 0.88, and these fitted values (not measured ones) are what the unified physical normalization uses for this campaign. Splits: random 80/20 (280/70).

Citation

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