Research and validation
Specification, model design, how we measure skill and what is still to prove. Written for technical evaluators; for the overview, see the home page.
Specification
Every forecast covers 24 levels and four variables, plus sea level, on the same grid as the data the model learned from.
| Coverage | Global ocean, 75°S to 75°N |
|---|---|
| Grid | 0.1° (about 11 km at the equator), 3,600 × 1,499 cells |
| Depth | 24 levels from 0.5 m to about 1,000 m; land and seafloor masked per level |
| Output | 97 fields: eastward (u) and northward (v) current, potential temperature (T) and practical salinity (S) on 24 levels, plus sea-surface height |
| Forecast | Mean of the next five days, one step ahead |
| Data | Long-term data from a global ocean model, five-day means: trained on 1981–2015, validated on 2016–2019, 2020–2023 held out for testing |
| Model | Residual U-Net, 13.4 million parameters; a global step is 119 tile evaluations on a single GPU |
| Status | Research preview, interim validation |
Model design
The nowcast rebuilds today's 3D ocean from what can be measured at the surface. The forecast steps that state five days forward. Both use the same grid, levels and normalisation, so the nowcast's output is exactly the forecast's input.
Model 1
Model 2
The forecast learns from 35 years of consistent 3D ocean data and never touches raw observations; only the nowcast does. A new satellite product means retraining the nowcast alone, and each model can be validated, and used, on its own: a 3D nowcast by itself serves drift and acoustics work.
An earlier prototype, trained on a different ocean model (MOM5), reconstructed temperature and salinity in the top 500 m from surface fields alone, with R² of 0.98 and 0.96 on held-out years. That is feasibility evidence: model data in and out, and no currents. The new nowcast is built and tested on synthetic data; its training is next.
Validation results
Persistence assumes the next five days look like the last five. It is the baseline every ocean forecast has to beat.
| Variable | MSE ratio | RMSE ratio | Skill score |
|---|---|---|---|
| Eastward current (u) | 0.267 | 0.52 | 0.73 |
| Northward current (v) | 0.212 | 0.46 | 0.79 |
| Temperature | 0.739 | 0.86 | 0.26 |
| Salinity | 0.922 | 0.96 | 0.08 |
| Sea level | 0.751 | 0.87 | 0.25 |
Ratio = model ÷ persistence (below 1 beats persistence). RMSE ratio = √(MSE ratio). Skill score = 1 − MSE ratio.
Interim checkpoint after 54 of 100 training epochs (1 October 2026). Scored on a fixed sample of 36 forecast start dates spread across 2016–2019, years excluded from training. The starting state and the truth both come from the same global ocean model data, so these scores measure how well the model has learned the ocean's five-day evolution, not yet the skill of the full real-time chain. Errors are pooled over all ocean cells and 24 levels in normalised units, weighted by cell area, and divided once. Where checked so far, skill is highest where currents change most: the high-variability regions of western boundary currents, the Antarctic Circumpolar Current and the equator.
Validation plan
Each step adds a harder test. Results are published with the method, including the ones that don't flatter the model.
3,119 global five-day fields from 1981 to 2023, standardised to 24 levels. Training uses 1981–2015, validation 2016–2019, and 2020–2023 is held back for the final test.
DoneCurrent errors well below persistence by the 21st training epoch, with no sign of overfitting: training and validation loss are equal.
DoneA control model trained on currents alone matched the main model within about 1%. The forecast keeps temperature, salinity and sea level, the state needed for longer forecasts.
DoneError by variable, depth and region in physical units (m/s, °C, psu, m), against persistence, on years the model has never seen.
In progressForecasts started from the nowcast, compared with forecasts started from global ocean model data and with persistence. If the chain loses the skill, the nowcast has failed, however good it looks alone.
PlannedSatellite temperature, salinity and altimetry with operational wind stress. Currents checked against drifting buoys; temperature and salinity against Argo floats.
PlannedThe forecast model has completed 54 of 100 training epochs and is still improving. Since epoch 21, the current error relative to persistence (MSE ratio) has fallen from 0.303 to 0.267 for the eastward component and from 0.239 to 0.212 for the northward.
A control model trained on currents alone came within about 1% of the main model, so we stopped it to save compute and kept the full 3D forecast. Next: finish training, then run the one-time test on 2020–2023.
A pilot scores Oceryn3D for your region on years it has never seen.
