Transolver surface-field surrogate for SHIFT-Wing (area sampling)
The model behind the Physics Surrogate Playground: from a wing-body surface (any tessellation, sampled by area) and the angle of attack at Mach 0.5, it predicts the surface pressure coefficient Cp and the skin-friction vector Cf everywhere on the surface; lift, drag and moment follow by integrating them.
Files
| File | Content |
|---|---|
best.pt |
checkpoint (weights, config, normaliser); load with psp_aero.inference.Predictor.from_checkpoint |
ood.npz |
out-of-distribution detector fitted on the training geometries |
demo/ |
4 held-out test cases (CAD mesh plus CFD fields) for the playground |
Test-split results
| Metric | Value |
|---|---|
| Cp relative L2 | 0.041 |
| Cf_x relative L2 | 0.178 |
| CL MAE (integrated) | 0.0045 |
| CD MAE (integrated, counts) | 4.3 |
| Cm MAE (integrated) | 0.0017 |
The training set has about 100 cases (the auto-approved Wing-sample), a tenth of what the SMART paper used, and covers α from about 0° to 4°. Predictions outside that range, or on geometries far from the CRM family (the OOD score flags them), are not reliable.
Licence and attribution
Geometry: NASA Common Research Model. CFD solutions: Luminary Cloud SHIFT-Wing (Wing-sample), CC-BY-NC-4.0. The weights and the demo cases are derived from that data and are CC-BY-NC-4.0 too: non-commercial use only. This is a clean-room implementation, not affiliated with Luminary Cloud or NASA.
SHIFT-SUV model (suv/)
A second surrogate of the same architecture, trained on the 99 quarter-scale
AeroSUV variants of SHIFT-SUV (50 m/s, zero yaw; one flow condition, so it
has no condition inputs). suv/ holds its best.pt, ood.npz and 4
held-out demo cases on 200k-triangle decimations of the solver surface.
Geometry: FKFS AeroSUV. CFD solutions: Luminary Cloud SHIFT-SUV (SUV-sample),
CC-BY-NC-4.0.
RF PCB board model (em/)
A board-image U-Net with rational (pole) frequency heads that predicts the
S-parameters from 1 to 20 GHz and the near E and H fields around 8 x 8 mm
RF boards (line, stub, ladder, loop and patch families), trained on boards
simulated with an FDTD solver (openEMS models, run with gfdtd). em/ holds
its best.pt, its evaluation summaries and 40 held-out test boards
(design.json, meta.json, sparams.h5) that make up the playground's
board catalogue.