Update README.md

README.md

@@ -7,9 +7,11 @@ size_categories: [10B<n<100B]
 
 # SuperWing (processed)
 
-> **This is a processed, downsampled derivative of SuperWing, 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.
+This is a processed, downsampled version of SuperWing, not the original dataset. 
+Fields were converted to a common frame and non-dimensionalized, rows were randomly subsampled and some variables were dropped. 
+For the original data, see the original [paper](https://arxiv.org/abs/2512.14397) 
 
-A preprocessed, downsampled release of SuperWing. Total size is 0.70 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.
+The original size was 2.6 TB, therefore the volume field was downsampled by 3x to obtain a size of 0.7 TB. 
 
 ## Layout
 
@@ -65,15 +67,6 @@ Dropped from this release: rho, T.
 
 Dropped from this release: rho_tilde, layer.
 
-## Downsampling
-
-| tree | samples | row factor | rows (release) | rows (collated source) |
-|---|---|---|---|---|
-| `collated` | 28,856 | 1/1 | 1.215e+09 | 1.215e+09 |
-| `volume_collated` | 28,294 | 1/3 | 2.394e+10 | 7.182e+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).
@@ -81,65 +74,4 @@ The volume tree uses the surface splits (keyed by stem).
 - `train`: 23,111 samples
 - `val`: 5,745 samples
 
-## Source dataset and preprocessing
-
-### SuperWing
-
-**What it is.** 28 856 transonic solutions over 4239 kinked (double-swept) wing planforms, each
-at up to eight operating conditions. 56 design variables per wing (7 planform, 18 spanwise,
-20 CST section coefficients, 11 operating). The only wall-resolved *and* structured-mesh
-campaign here.
-
-**Generation.** ADflow (MDOLab), steady compressible RANS with Spalart-Allmaras in a fully
-turbulent setting (no transition model). Structured O-mesh hyperbolically extruded from the
-surface: 249 circumferential × 41/125 spanwise (inner/outer) × 81 wall-normal layers at
-**y⁺ ≈ 1** (the release ships the inner 70 layers). Solution strategy is 3w multigrid, 500
-cycles on coarse levels then up to 3000 on the fine mesh, approximate Newton-Krylov until the
-total residual drops below 1e-8, then Newton-Krylov to **1e-10**.
-
-**Parameters.** Mach 0.75–0.90, Re fixed at 2.0e7, T∞ = 300 K ⇒ U∞ ≈ 260–313 m/s (derived from
-a = 347 m/s), AoA 2–12°. **The release carries no dimensional quantity at all** — no density,
-no p∞, no viscosity — which is why Reynolds number is unavailable as a normalization scale for
-this campaign.
-
-**Cost.** "The simulations are conducted on the 160-core high-performance computing cluster at
-Tsinghua University for over four months", i.e. ≈ 4.6e5 core-hours for the campaign and ≈ 16
-core-hours per case (derived from the wall time, not published per case). Released as 26.7 GB
-of compressed surface data plus 3.2 TB of volume.
-
-**Processing.** The surface release is a single dense array, `(28856, 9, 44096)` float64
-(~92 GB), in `[sample, channel, point]` layout with raw channel order
-`x, y, z, ρ̃, Cp, Cfx, Cfy, Cfz, T̃`; the 44 096 cell centres are fixed per wing geometry. The
-builder transposes to `[44096, 9]`, reorders channels into the shared contract
-`[x, y, z, cp, cfx, cfy, cfz, rho, T]`, and row-shuffles. Cp/Cf are already coefficients, so no
-conversion. The **raw frame** is chord = +x, span = z, vertical = +y with +y the suction side —
-established by an upper/lower surface pairing test binned on chord × span, then checked for 100%
-sign agreement with the released CFD lift coefficient — so SuperWing carries the same `ORIENT`
-permutation as WindsorML (`x ← x, y ← −z, z ← y`), which puts negative Cp on the +z side of a
-lifting wing. The volume arrives as 46 shards of `(n, 8, 3086720)` float64 (5.72 TB) with
-freestream-normalized channels `x, y, z, ρ̃, p̃, Ṽx, Ṽy, Ṽz`. Its topology is undocumented
-upstream and was recovered here: 3 086 720 = 44 096 × **70**, i.e. the surface cells extruded
-through 70 wall-normal layers in 13 structured blocks, layer-major within a block, with the
-blocks in a *different* order in the volume array than in the surface array. The grading is
-exponential (median wall distance 9e-7 at layer 0 to 6.5 at layer 69), so **half of all volume
-points lie within 0.0085 chord of the wall**, and layer 0 *is* the surface field — its density
-matches the surface file bit-for-bit. A `layer` column is therefore carried in the output
-(`[x, y, z, ux, uy, uz, cp, rho_tilde, layer]`, 9 cols) so that wall-distance thinning is a
-filter over the collated array rather than a re-read of the raw shards. Volume Cp is recovered
-as `(p̃ − 1)/(½γM²)` with γ = 1.4, reproducing the surface Cp column to 4e-8; velocity is kept
-non-dimensional (`u/U∞`) since no dimensional scale is shipped. The alignment of each shard to
-its samples is asserted by un-shuffling the paired surface file and matching the 44 096 layer-0
-rows (a wrong offset gives a positional discrepancy of 3.13 against 1.1e-6 for the correct one).
-One upstream defect: shard 32 is byte-identical to shard 31 (same published checksum), so 562
-samples have no volume data and the volume tree holds 28 294 of 28 856. Crop margins are
-`x: 0.25/0.75 L_x`, `z: 0.15/0.02 L_z`, with the thin-axis (y) half-width at
-`0.20 max(L_x, L_z)` — the streamwise reference, not L_y, or the slab clips the shock off the
-suction side — keeping 81.4–83.8%. **Surface Cf is pruned** at 3σ per sample on all three
-components (1 272 434 176 → 1 214 523 900 rows, −4.55%); the removed rows concentrate at the
-wing tip and aft chord and the pruning is a normalization trade, recorded in
-`collated/pruned_cf.json` with a keep-mask so the volume's alignment gate still works. Splits
-are **geometry-grouped** 80/20 (23 111/5745).
-
-## Citation
-
-TODO: this work, and the upstream dataset's citation.
+Either use norm_stats_centered.npz, or norm_stats_thinned.npz. The centered stats are computed on the simulation mesh (i.e., native discretization), while the thinned stats are computed after (roughly) uniform sampling.