Compare a numerical reference with a trained CNN's prediction on held-out test pairs. See exactly where they disagree.
Height shows vorticity in the first two plots and absolute error in the third. These are 2D fields drawn in perspective, not 3D fluid simulations.
| Prediction | Relative L2 | MAE | Enstrophy Δ |
|---|---|---|---|
| Persistence baseline | |||
| CNN · replay | |||
| CNN · live · browser CPU |
Persistence reuses the input field unchanged. These values describe one held-out sample; they are not held-out benchmark performance. The 2000-sample table is in the Benchmark section.
The dataset's own test file (2000 pairs) is never read for statistics, thresholds or training. Validation was carved from the train file by instance ID; the archive exposes no trajectory grouping.
32 × 32 fields are a stride-4 decimation of the 128 × 128 archive. One mean/std fitted on training pairs normalises inputs and targets; predictions are inverted to stored units before display and metrics.
Spatial errors, the persistence baseline and enstrophy are diagnostics, not proofs: enstrophy agreement does not establish a PDE solution, and a 2D surrogate says nothing about 3D Navier–Stokes.
| Model | Params | Rel. L2 mean | Rel. L2 worst | MAE | Enstrophy Δ mean | Enstrophy Δ worst | p50 ms | p95 ms |
|---|
| Model | Rel. L2 mean | Rel. L2 worst | MAE | Enstrophy Δ mean |
|---|