Research · 2D fields

Measured 2D field error on shapes the model did not see in training.

Published 4 Aug 2026

Latest study Evaluation 9 Jul 2026

Multi-shape evaluation

Published 9 Jul 2026 · Updated 4 Aug 2026

One model trained on 96 shapes, then checked on shapes it had never seen. Numbers, figures, and failure notes from that study follow.

One model, tested on new shapes.

Published 9 Jul 2026 · Updated 4 Aug 2026

This study compares model estimates with a full fluid simulation across several requested time horizons. It was not tested on every geometry or flow regime.

July 9 multi-shape evaluation comparing reference and learned flow-map results
Final multi-shape evaluation · reference and prediction comparison
FIG 2.0 Headline numbers

Measured results from the same study.

Published 9 Jul 2026 · Updated 4 Aug 2026

Each value below comes from the multi-shape holdout evaluation.

0.0030
relative L2 error in the held-out multi-shape evaluation
96
randomized geometries used for training
5.6×
improvement over the earlier 64-cylinder model
15×
reported margin over the “nothing changed” baseline
FIG 2.1 Runtime note and limits

What the ≈35 ms measurement includes.

Published 9 Jul 2026 · Updated 4 Aug 2026

This is an average for one forward pass of the learned flow-map model on a 64×64 2D periodic case. It was measured on an Apple M3 Pro with PyTorch MPS across 40 warmed calls: eight held-out seeds at five requested horizons.

Timed
model forward pass + MPS synchronization
Excluded
loading, data preparation, UI, import, export, and reference solve
Scope
research benchmark · not end-to-end Studio latency

The average stayed flat across the five tested horizons because each request used one model pass. The paired reference solver was faster through the entire tested range at 64². This benchmark therefore does not establish a speed advantage over that solver or over conventional CFD.

FIG 3.0 Direct inspection

Divide the field yourself.

Published 9 Jul 2026 · Updated 4 Aug 2026

Solver ground truth and the model’s one-forward-pass prediction for an unseen rectangle at Re 301, sixteen requested steps ahead (t=0.64). Drag the divider or use the arrow keys to inspect where the wake phase drifts.

Case
held-out rectangle · Re 301 · L 0.70
Horizon
16 steps · one forward pass
Wake RelL2
0.037 at t=0.64
Solver ground-truth vorticity field for an unseen rectangle at Reynolds 301

Vorticity panels from wake evaluation figure · held-out geometry study

Research evaluation figure showing measured flow-map performance
Evaluation figure · research artifact · not a UI rendering
FIG 4.0 Error inspection

Pictures can look right while the numbers are wrong.

Published 9 Jul 2026 · Updated 4 Aug 2026

The evaluation compares full fields, the “nothing changed” baseline, and visible wake errors. That combination catches failures that a convincing image can hide.

Earlier studies

Earlier study Prior geometry baseline

Published 4 Aug 2026

Improvement over the earlier 64-cylinder model

The July multi-shape evaluation reports a 5.6× improvement over the earlier 64-cylinder model on the published holdout comparison. That earlier baseline is kept here so the latest number stays comparable, not so the old run replaces the current record.

Evaluation integrity note: a bookkeeping leak once made one seed look better than a true holdout. That class of mistake is why this page leads with held-out shapes and a “nothing changed” floor, not with a single flattering image.

Limits of these evaluations.

Published 4 Aug 2026 · Updated 21 Aug 2026

These evaluations do not certify Reyn Studio, prove accuracy outside the tested equations and shapes, or qualify every model bundle. Relative L2 and load errors are diagnostics, 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 each case to the model's declared grid. This hub covers Reyn's published fluid-model research, not the permanent scope of the company.

Request access.

Studio packages are public. Credentials are issued after review. These evaluations do not qualify the desktop preview as production CFD.

Request beta access