hiliftaeroml_processed

This is a processed, downsampled derivative of HiLiftAeroML, 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 HiLiftAeroML (upstream: https://huggingface.co/datasets/nvidia/HiLiftAeroML). Total size is 0.88 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, 7 columns):

Dropped from this release: rho_tilde, yPlus.

Volume (volume_collated, 7 columns):

Downsampling

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.

The collated HiLiftAeroML surface had already been halved before this release, so its surface factor relative to the raw mesh is 1/24.

Splits

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

• train: 1,608 samples
• val: 179 samples

Source dataset and preprocessing

HiLiftAeroML

What it is. 1787 wall-modelled LES solutions over the NASA CRM-HL high-lift configuration: 180 geometry variants (slat and flap deflections and gap multipliers, inboard and outboard — 8 design variables) at 10 angles of attack, 4–22° in 2° steps. It models high-lift aerodynamics through separation onset, and it is by far the largest campaign here in points per sample.

Generation. Fidelity Charles (charLES), explicit unstructured finite-volume, second-order in space and third-order in time, with a dynamic Smagorinsky SGS model (coefficient computed dynamically in space and time, which is what makes it usable through separation) and an equilibrium wall model solving simplified boundary-layer equations on an embedded near-wall grid. Voronoi meshes from Fidelity Stitch, 300–500 M cells (150–300 M after conversion for ML use). Statistics are gathered over 15 CTU following a 10 CTU transient below 12°, and 30–200 CTU at the separated high-AoA conditions. Meshes differ per AoA (adaptive refinement) — it is not one mesh reused.

Parameters. M = 0.2, Re_MAC = 1.6e6, MAC = 275.80 in (7.005 m), half-span 1156.75 in. Solver units are Imperial with Rankine temperatures, confirmed from the volume far field: avg(T) = 518.67 °R (= 288.15 K), avg(P) = 176.352 in solver units against the derived p∞ = 176.352 (far-field Cp = −3e-6), density ratio 1.000000. Reference values are q_ref = 4.937856, u_ref = 2679.505 (68.06 m/s at M 0.2), γ = 1.4. Because the geometry is full-scale in inches while Re_MAC is 1.6e6, the solver viscosity is scaled rather than sea-level.

Cost. Typically 8 NVIDIA GH200 or GB200 nodes per case; no wall-clock or aggregate cost is published, so no campaign total is quoted here.

Processing. Surface fields are on mesh points and are raw dimensional solver quantities (no Cp/Cf in the file, no freestream either), so they are non-dimensionalized in the builder against the single global freestream: Cp = (PROJ(AVG(P)) − p∞)/q_ref, Cf = AVG(TAU_WALL)/q_ref, rho_tilde = PROJ(AVG(RHO))/ρ∞, giving [x, y, z, cp, cfx, cfy, cfz, rho_tilde, yPlus] (9 cols). Only 6 of the 21 point arrays are read, which roughly halves I/O and memory. Sanity checks: Cp_max ≈ 1.1 (the compressible stagnation value, which validates the single global freestream), Cp_min ≈ −12 at 22°, rho_tilde ∈ [0.7, 1.02], y ≥ 0 (symmetry plane). Frame: x streamwise, y spanwise (half-domain, symmetry at y = 0, tip at +1160), z lift; positions are in inches, so pos_scale ≈ 0.001205. The surface tree was halved in place to 50% of its points (9.08 → 4.54 TB, ~63–77 M points per sample) — which is free precisely because rows were permuted at write time: the leading n/2 rows are already a uniform subsample, so the operation is an .npy header rewrite plus an ftruncate rather than a 13.5 TB read-and-rewrite. The volume is the hardest case in the set: one .tgz per sample at ~23.7 GB (~42 TB for the 1787 cases), each holding one ~90 GB VTU of ~235 M points / ~200 M polyhedral cells, at a measured gzip ratio of 3.775 — so extracting even a third of the campaign would exceed the available allocation, and extracting all of it (~160 TB) is impossible here. It is instead streamed: the appended-data offsets are monotonic in declaration order, so one forward-only inflate pass reaches everything needed, with the faces and faceoffsets blocks (51.5 GB, 57% of each file) inflated and discarded. Measured at 4 min 24 s and 6.4 GB peak RSS per sample, 235.6 M points in to 117.7 M rows out at --keep-frac 0.5, ≈ 130 core-hours to convert the campaign. The per-sample gate is again the freestream: far-field mean velocity reproduces u_ref[cos α, 0, sin α] to 6e-8 and recovers α to 10.000° on geo_LHC001_AoA_10, which is sharp enough to catch a stem↔archive mismatch (one AoA step is 35× the tolerance). The crop keeps 99.894% — the domain is a ±24 533 box but the LES mesh is glued to the wing — and its vertical margins scale with L_z here (unlike Double-Delta) because this wing's vertical extent is a third of its streamwise extent, not a sixteenth. Volume output [x, y, z, ux, uy, uz, cp] (7 cols, 208.0e9 rows, 5.82 TB); rho_tilde/t_tilde are omitted because at M 0.2 they are near-redundant with Cp. Conditioning is (aoa, mach) with Mach a zero-width range (0.2 everywhere), normalized to exactly 0 rather than min-maxed. Splits are geometry-grouped 90/10 (1608 samples / 162 geometries vs. 179 / 18).

Citation

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