Measured 2D field error on shapes excluded from training.

Published 4 Aug 2026, updated 14 Nov 2026

A single learned flow-map model trained across 96 randomized 2D shapes and evaluated against a spectral solver on held-out geometries across multiple requested time horizons.

Evaluated 9 Jul 2026

Multi-shape holdout evaluation.

Training on a single canonical shape (such as a circular cylinder) overfits wake separation angles. Training one flow map across 96 randomized geometries cut held-out error by 5.6× and cleared the “nothing changed” persistence floor by 15×.

Held-out RelL2 0.0030 Unseen test shapes
Training diversity 96 Randomized 2D bodies
vs 64-cylinder 5.6× Error reduction
vs Persistence 15× Below static floor
FIG 1.0: Holdout protocol

One model, tested on unseen shapes.

The operator receives the current fluid state and obstacle mask, jumps directly to the requested future horizon, and is scored pointwise against the full numerical simulation.

Plate 1.0: Multi-horizon evaluation summary Measured Multi-shape holdout (9 Jul 2026)
July 9 multi-shape evaluation comparing reference simulation and learned flow-map predictions across requested horizons
Final multi-shape evaluation artifact: Reference spectral simulation versus single-pass flow-map estimates across unseen holdout shapes and requested jump horizons.
Evaluation rule
Zero overlap between the 96 training geometries and the held-out test set
Primary metric
Whole-field Relative L2 against the paired 2D spectral reference solver
Mandatory floor
Hold-last-state persistence (copying the initial fluid frame unchanged to the target time)
Scope boundary
2D periodic domain, bounded Reynolds and shape family
FIG 2.0: Quantitative ledger

Measured results against both baselines.

Every figure in the holdout study is anchored against both the prior single-shape cylinder baseline and the zero-effort persistence floor.

Relative error comparison on held-out geometries Lower bar is more accurate
Multi-shape flow map (96 training geometries) RelL2 0.0030 (1.0× (headline))
Earlier single-family model (64 cylinders) 5.6× higher error (5.6× vs multi-shape)
Hold last state (“nothing changed” floor) 15× higher error (15× vs multi-shape)
0.0030
Held-out field RelL2 across unseen test shapes in the multi-shape evaluation.
96
Randomized 2D training shapes spanning varied aspect ratios, orientations, and bluff cross-sections.
5.6×
Error reduction over the earlier model trained on 64 circular cylinders when scored on non-circular holdouts.
15×
Margin below hold-last-state, verifying genuine unsteady wake transport rather than static frame copying.
Latency & scaling caveat

What the ≈35 ms pass includes, and why 64² is not a speed win.

Constant-cost horizon jumping is an algorithmic property rather than a wall-clock claim over an FFT solver on a 64×64 grid.

On an Apple M3 Pro (PyTorch MPS), one warmed forward pass of the 2D flow-map model averages ≈35 ms across 40 timed calls (eight held-out seeds × five requested jump horizons). Because the operator jumps directly to the target horizon in one pass, inference cost stays flat as the requested jump grows.

Hardware & backend
Apple M3 Pro, PyTorch MPS, 40 warmed calls (8 seeds × 5 horizons)
Single model jump
≈35 ms flat across all five requested horizons (forward pass + MPS sync)
Paired 64² solver
≈0.5 ms per step; ≈8 ms for 16 steps (beyond 16 steps at 64²)
Excluded overhead
Disk I/O, mesh preparation, UI chrome, export, and warm-up frames
Toy-scale caveat

At 64×64 resolution in 2D, the paired spectral FFT solver steps in ≈0.5 ms and completes the tested range faster than a single ≈35 ms model call. This 2D study establishes constant-time horizon jumping and shape generalization rather than desktop 3D runtime speed.

FIG 3.0: Interactive inspection

Divide the Re 301 wake yourself.

Drag the vertical crosshair or toggle Solver Only and Operator Only to inspect vortex rollup and downstream phase drift behind an unseen rectangle 16 steps ahead (t = 0.64).

Geometry
Held-out rectangle (L = 0.70)
Regime
Re = 301, unsteady vortex street
Jump
16 steps, single forward pass (t = 0.64)
Wake RelL2
0.037 (mid-horizon hump damped ≈23%)
Plate 3.0: Held-out rectangle wake split (Re = 301) Measured Wake RelL2 0.068 (unseen aspect ratio & yaw)
Solver reference Model prediction
50%
Solver ground-truth vorticity field for an unseen rectangle at Reynolds 301
Left of reticle: Solver reference; right: Model prediction Native 562×562px lattice render (drag or use ← / → keys)
Unseen Re = 301 rectangle vorticity slice: Primary shear-layer separation and near-wake vortex cores align closely; residual discrepancy concentrates in slight downstream phase offset (wake RelL2 0.037).
FIG 4.0: Error anatomy

Smooth contours can hide phase error.

A neural field can look visually plausible while advecting vortices at the wrong speed or smearing shear layers. Pointwise difference maps and horizon curves expose where accuracy degrades.

Plate 4.0: Pointwise residual & horizon diagnostics Measured Research diagnostic plot
Research evaluation figure showing pointwise error localization and horizon scaling
Pointwise residual inspection: Evaluating full-field RelL2, hold-last-state margins, and spatial error maps together prevents flattering colorbars from masking wake-phase drift. In intermediate shedding transitions (around step 13), consistency training damps the transient wake-error hump by ≈23% without eliminating advection phase sensitivity entirely.
Prior baseline

Earlier 2D studies & audit notes.

Collapsed by default so the July multi-shape record leads, while keeping historical baselines and evaluation-integrity fixes on the record.

Archived baseline 64-cylinder single-family comparison & seed-leak audit
Published 4 Aug 2026

Improvement over the earlier 64-cylinder model (5.6×)

Before training across 96 diverse shapes, our prior 2D baseline trained on 64 circular-cylinder configurations. On the multi-shape holdout suite, expanding shape diversity reduced Relative L2 error by 5.6×. We keep this baseline archived here so improvements remain anchored to a fixed internal reference.

Evaluation integrity audit note

During earlier experimentation, an index bookkeeping leak once caused a test seed to appear cleaner than a true unseen holdout. Catching and auditing that failure established our current non-negotiable rule: strict geometry-family separation, mandatory hold-last-state floors, and zero single-seed cherry-picking.

Operating boundaries

What these numbers do and do not claim.

Every held-out score on this site is paired with the grid, regime, and failure gate that bound it in production.

Qualification

Verification, not certification

Relative L2, force RelL2, and Strouhal checks measure discrepancy against reference solvers on held-out test sets. They do not certify Reyn Studio for flight, road, marine, or safety-critical sign-off.

Grid resolution

Declared 32³ voxel grid

The shipping 3D research operator resolves 32,768 cells. Imported CAD thinner than one cell along any axis is refused before inference rather than extrapolated across empty solid voxels.

Flow regime

Tested Reynolds & horizon scope

Published scores cover laminar/transitional bluff-body wake regimes (Re = 100–355 in 2D; Re = 150 to 128 solver steps in 3D, RelL2 0.052). High-Reynolds RANS/LES boundary layers are outside current bundle scope.

Derived loads

Derived pressure & fluid tractions

Surface Cp is recovered from the predicted velocity field, and diffuse-interface tractions represent fluid loads for relative ranking or downstream FEA mapping, not internal structural stress.

Evaluate a geometry.

Studio packages are public to download; credentials are issued after review for external-flow candidate screening.