Single-jump 3D velocity prediction across unseen solid families.
Starting after the numerical solver’s 40-step warm-up, the operator maps the current 3D velocity field and binary occupancy mask directly to step 32, 64, or 128 in a single pass.
- Shipping bundle
- reyn-flow3d-obstacle-32-v1.reynmodel (SHA-256
0e0524e9c665b1b46138…) - Grid & boundaries
- 32³ Cartesian lattice (32,768 cells) in a 3D periodic box
- Warm-up & jump
- 40 solver warm-up steps, then single neural jump to H = 32, 64, or 128
- Occupied vs empty
- Occupied-family mean 0.062 (40 cases); empty-box mean 0.003 (8 cases)
Whole-field RelL2 from 32 to 128 steps.
Error grows gracefully with jump horizon while the persistence baseline diverges rapidly as wakes advect downstream.
- 0.052H128
- 128-step velocity RelL2 across all 48 held-out cases, versus 0.209 for hold-last-state (4.0× margin).
- 0.020 / 0.028
- 32-step and 64-step velocity RelL2 on the same corpus, against hold-last-state scores of 0.084 (4.2×) and 0.136 (4.9×).
- 0.062 / 0.003
- Occupied-obstacle mean vs empty-box mean at 128 steps, reported separately so trivial obstacle-free cases never dilute bluff-wake error.
Error stratified across all six geometry families.
Multi-body wake impingement is substantially harder than isolated bluff flow. We publish all six eight-case families unaggregated so failure modes are transparent.
| Geometry family (n = 8 each) | One-jump RelL2 | Hold last state | Margin | Visual comparison (H = 128) |
|---|---|---|---|---|
| Single bluff body | 0.041 | 0.187 | 4.6× | |
| Irregular non-convex body | 0.052 | 0.207 | 4.0× | |
| Ground-adjacent body | 0.063 | 0.280 | 4.4× | |
| Multiple bodies (wake interaction) | 0.077 | 0.301 | 3.9× | |
| Irregular multiple bodies | 0.076 | 0.271 | 3.6× | |
| Empty periodic box | 0.003 | 0.006 | 2.0× |
How 32³ behaves on real STL meshes, including refusals.
When an engineer imports CAD into Reyn Studio, preflight voxelizes the mesh onto this exact 32³ lattice. Select all six reference geometries to inspect solid-cell counts, velocity divergence RMS, and why thin wings are refused pre-solve.
Ahmed reference body
1.044 m · R100 nose · 25° slant (500 STL triangles)
- Occupied solid voxels
- 60 / 32,768
- Thinnest span (y-axis)
- 2.26 cells
- Axis-ray disagreement
- 6.25%
- Divergence RMS ‖∇·u‖₂
- 0.17713
- Wake deficit peak
- 0.994
- Streamwise Cx
- 0.1891
- Field SHA-256
- 3f6fc76a44a30d40…
Ahmed-like cabin
1.04 m · 25° rear slant (1,692 STL triangles)
- Occupied solid voxels
- 56 / 32,768
- Thinnest span (y-axis)
- 2.26 cells
- Axis-ray disagreement
- 6.67%
- Divergence RMS ‖∇·u‖₂
- 0.18597
- Wake deficit peak
- 0.995
- Streamwise Cx
- 0.2116
- Field SHA-256
- 33e6c8be86653eb4…
Closed fairing
1.30 m · Ø0.30 m (1,248 STL triangles)
- Occupied solid voxels
- 52 / 32,768
- Thinnest span (y-axis)
- 3.06 cells
- Axis-ray disagreement
- 0.00%
- Divergence RMS ‖∇·u‖₂
- 0.15142
- Wake deficit peak
- 0.992
- Streamwise Cx
- 0.1040
- Field SHA-256
- 8202e1dbd035cbb4…
Capsule
Ø80 × 260 mm (1,920 STL triangles)
- Occupied solid voxels
- 36 / 32,768
- Thinnest span (y-axis)
- 3.06 cells
- Axis-ray disagreement
- 0.00%
- Divergence RMS ‖∇·u‖₂
- 0.15905
- Wake deficit peak
- 0.993
- Streamwise Cx
- 0.1052
- Field SHA-256
- efae0206dda2e0b8…
Cylinder in cross-flow
Ø60 × 200 mm (256 STL triangles)
- Occupied solid voxels
- 0 (empty occupancy mask)
- Thinnest span (x-axis)
- 0.92 cells
- Auto-fit scale factor
- 0.0030
- Operator invocation
- Aborted before forward pass
Diagnostic refusal: Voxelization produced an empty solid: after auto-fit the body is 0.92 cells thick across x on a 32³ grid. Resolving thin airfoils or slender struts requires a finer declared grid rather than extrapolating from an empty voxel mask.
NACA 0012 wing
c 120 mm · span 300 mm (632 STL triangles)
- Occupied solid voxels
- 0 (empty occupancy mask)
- Thinnest span (z-axis)
- 0.15 cells
- Auto-fit scale factor
- 0.0020
- Operator invocation
- Aborted before forward pass
Diagnostic refusal: Voxelization produced an empty solid: after auto-fit the body is 0.15 cells thick across z on a 32³ grid. Resolving thin airfoils or slender struts requires a finer declared grid rather than extrapolating from an empty voxel mask.
Uninterpolated 32³ cuts across XY, XZ, and YZ.
We render raw 32³ grid slices without bicubic smoothing so engineers can see the actual spatial resolution of the operator across isolated and multi-body wakes.
Why 32³ is never upsampled to pretend 64³.
This study covers a 32³ periodic box, six held-out obstacle and empty-box families, whole-field velocity RelL2, and a single jump to 128 steps after a 40-step warm-up. It does not qualify high-Reynolds turbulent boundary layers, compressible aerodynamics, internal ducting, or rotating machinery.
CAD Run in Reyn Studio voxelizes imported STL geometry strictly to the bundle’s declared 32³ lattice (32,768 cells). Studio never interpolates a 32³ operator onto a 64³ or 128³ display grid to create an illusion of finer boundary-layer resolution.