ayz2/ahmedml_processed

AhmedML (processed)

This is a processed, downsampled derivative of AhmedML, 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 AhmedML. Total size is 0.28 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	p	static pressure (raw solver units; Cp = (p - p_inf) / q)
4	tau_x	wall shear stress (raw)
5	tau_y	wall shear stress (raw)
6	tau_z	wall shear stress (raw)
7	Cp	pressure coefficient (as shipped upstream)
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	500	1/1	5.345e+08	5.345e+08
volume_collated	500	1/1	9.193e+09	9.193e+09

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: 400 samples
• val: 100 samples

Source dataset and preprocessing

AhmedML

What it is. 500 variants of the Ahmed body — the canonical simplified automotive bluff body — spanning slant angle, body length/width/height and ground clearance. It models base separation and the slant-angle drag-crisis regime.

Generation. Plain SA-DDES with the standard f_d shielding function (no σ-model, no EP shielding), same pimpleFoam / OpenFOAM v2212 lineage as DrivAerML, second-order in time. ~21.3 M prismatic/hex-dominant cells, deliberately high-y⁺: 3 prism layers giving y⁺ ≈ 50 with Spalding wall functions. 40 convective transit times, with averaging started after the first 10. Δt = 6e-4.

Parameters. U∞ = 1 m/s with Re = 7.68e5 set through the viscosity instead (ν = 3.75e-7 m²/s), Re based on body height H = 0.288 m (2.78e6 length-based). Verified in the data rather than taken from the paper: regressing the stored Cp column on the stored kinematic p returns q = 0.500000 to 12 digits.

Cost. 288 cores on 4 AWS hpc6a.48xlarge nodes, ~30 min to mesh and ~48 h to solve (≈ 184 node-hours, ≈ 1.4e4 core-hours per case), so ≈ 6.9e6 core-hours for the campaign.

Processing. Cell-based surface, [x, y, z, p, τx, τy, τz, Cp, yPlus] (9 cols, kinematic p, CF_SIGN = −1). Axes as DrivAerNet++/DrivAerML, full domain, no mirroring. One campaign-specific subtlety: the tunnel floor is at z = 0 but the body sits on stilts, so the surface z-minimum is ≈ 0.05, above the floor. The volume crop therefore clamps its lower vertical bound down to the ground plane z = 0 rather than to the surface bbox, which keeps the under-body gap flow and raises retention from ~62% to ~74%. Crop margins otherwise [0.5, 0.75, 0.0]/[2.0, 0.75, 0.5] L. Volume output [x, y, z, ux, uy, uz, p]. Splits: random 80/20 (400/100).

Citation

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