ALIENX · FLOW LAB 3D
a 32³ neural operator vs a Navier–Stokes spectral solver — Taylor–Green turbulence
"It was trained on three Reynolds numbers. This one it has never seen."
Re = 900. Zero-shot. 5.7% rollout RMSE over 25 steps.
Left: pointwise MSE of the autoregressive rollout against the spectral DNS truth,
at Re=900 (a value outside training {300, 628, 1200}). The dashed line is the same run with
QSA's complex phase channel severed (gi=0) — the phase carries the predictive signal,
~5× degradation when cut. Right: kinetic energy — the vortex decay the network reproduces.
Every step is divergence-free by construction (Biot–Savart recovery):
max |∇·u| ≈ 2×10⁻⁶, float32 round-off.
A 421,507-parameter network holding a 32×32×32 turbulence field — no mesh, no grid,
no Reynolds it met before.
"The mid-plane of the vortex."
|vorticity| at the mid-plane (z = 16 of 32) — solver (top) vs
the operator (bottom), late in the rollout. The vortex filaments, their roll-up,
and their decay were never memorized: the network carries them in complex attention
phase, locally reframed SO(3) geometry, and an isotropic 26-neighbor stencil.
Turbulence without a grid. — Grimoire of Elbàlor · Eric Yaka · AGPL-3.0