Brest FNO — initial-state effect
Aggregates from two simulation protocols, separated by window count and published at source precision.
Experiments run on 9 and 10 August 2026. Evidence page published on 12 September 2026.
These results compare FNO seed 0 with the B1a baseline in a CROCO surface simulation. They are neither a retraining run nor validation against the real ocean.
| Protocol | Windows | Model | Initial state | Median at 12.5 h |
|---|---|---|---|---|
| Performance | 238 | FNO seed 0 | Exact CROCO state | 15.54 % |
| Performance | 238 | Baseline B1a | Not applicable | 32.56 % |
| Initial condition | 237 | FNO seed 0 | Exact CROCO state | 15.57 % |
| Initial condition | 237 | FNO seed 0 | Harmonic PACK_11 | 31.00 % |
| Initial condition | 237 | Baseline B1a | Not applicable | 32.57 % |
Method and provenance
The protocol covers 570 days of CROCO surface simulation. Lead 50 is 12.5 hours, using 15-minute steps and a stride of 24 steps. The metric is relative L2 over physical u and v on the CROCO water mask, followed by the median across windows; it is not a time average.
The performance protocol contains 238 windows and the initial-condition protocol contains 237. “Exact CROCO” is a simulation state, not an observation. B1a has no dynamic initial state and is therefore marked “Not applicable”. The public JSON retains available source precision: 15.565557261456881% and 31.001187338184856% for the two 237-window FNO arms, and 32.56085470819917% for B1a over 238 windows. The table rounds to two decimals.
Limits and reproducibility
- The 15.54% comparison assumes an exact CROCO initial state. With a PACK_11 harmonic reconstruction, FNO reaches 31.00%, versus 32.57% for B1a over the same 237 windows.
- The 238-window and 237-window protocols use different samples and must not be merged.
- Full reproduction is blocked: model weights, the evaluation dataset and scientific code are private. The packet only supports verification of the aggregates and their CSV transcription.