3D occupancy CFD evaluation

One jump. New shapes. Beat “nothing changed.”

Published 21 Aug 2026

The model estimates a later velocity field from the current one in a single jump. That estimate is scored against a full 3D simulation on shapes held out of the test set, on a 32³ Cartesian grid in a periodic box. The required floor is hold last state: copy the initial field forward as if nothing changed.

Held-out irregular body: mid-plane vorticity from the simulation, the learned map, and hold last state at 32, 64, and 128 steps — Irregular body · 32³ mid-plane vorticity · simulation, learned map, hold last state

FIG 2.0 Headline numbers

Measured results from the same study.

Published 21 Aug 2026

Whole-field relative L2 on velocity. Occupied cases and empty-box cases are split at 128 steps only.

0.052: relative L2 at 128 steps on the held-out set
4.0×: margin over hold last state at 128 steps
0.020 / 0.028: relative L2 at 32 and 64 steps (hold last state 0.084 / 0.136)
0.062 / 0.003: occupied-case and empty-box means at 128 steps
48: held-out cases on a 32³ periodic box

FIG 3.0 Horizon and families

The gap to hold last state holds across horizons and families.

Published 21 Aug 2026

The public chart shows only the learned map and hold last state. Family means at 128 steps are eight cases each.

Line chart of whole-field relative L2 versus horizon for the learned map and hold last state — Learned map versus hold last state · 32, 64, and 128 steps · 48 cases

Bar chart of 128-step relative L2 by geometry family for the learned map and hold last state — Family means at 128 steps · eight cases each · learned map versus hold last state

Family means at 128 steps
Family	Learned map	Hold last state
single body	0.041	0.187
irregular body	0.052	0.207
ground-adjacent	0.063	0.280
multiple bodies	0.077	0.301
irregular multiple	0.076	0.271
empty box	0.003	0.006

FIG 4.0 Inspection

Mid-planes stay coarse. That is the grid.

These slices are 32³. They are not upscaled. Orthogonal cuts at 128 steps sit beside single-body and multiple-body comparisons across the same horizons.

Orthogonal xy, xz, and yz vorticity slices at 128 steps for an irregular body, simulation versus learned map — Irregular body · orthogonal vorticity slices · 128 steps

Single-body mid-plane vorticity comparing simulation, learned map, and hold last state at 32, 64, and 128 steps — Single body · mid-plane vorticity · 32 / 64 / 128 steps

Multiple-bodies mid-plane vorticity comparing simulation, learned map, and hold last state at 32, 64, and 128 steps — Multiple bodies · mid-plane vorticity · 32 / 64 / 128 steps

FIG 5.0 Study limits

What this study does not cover.

Published 21 Aug 2026

This evaluation is 32³, a periodic box, held-out obstacle and empty-box families, velocity RelL2, and a one-jump map to 128 steps. It does not certify Reyn Studio, engineering loads, other physics, or finer grids. RelL2 is a diagnostic, not engineering approval.

Research 3D evidence is the 32³ one-jump obstacle operator (RelL2 0.052 at 128 steps, 32,768 cells). CAD Run voxels the case to that declared grid. It does not resample this operator onto a 64³ field.

Measured 3D field error on held-out obstacle cases.

Held-out 3D obstacle-flow map

One jump. New shapes. Beat “nothing changed.”

Measured results from the same study.

The gap to hold last state holds across horizons and families.

Mid-planes stay coarse. That is the grid.

What this study does not cover.

Earlier studies

3D solver and obstacle evaluations

3D has solver evidence. Studio keeps the gate.

Limits of these evaluations.

Request access.