Sixty Days of Ocean, With Conditions

TL;DR for operators

Neptune is a global ocean and sea-ice emulator that produces daily states out to 60 days. On held-out 2019-2021 reanalysis data, its 1° model generally beats training-set climatology for most variables and preserves large-scale temperature, sea-surface-height, ocean-heat-content, ENSO, and Indian Ocean Dipole structure well into the forecast horizon.

The qualification problem begins when the forecast is examined by variable, region, and spatial scale. Currents and mixed-layer depth deteriorate more rapidly, energetic current systems remain difficult, fine-scale velocity variability is smoothed, and some sea-ice regions become less reliable at long lead times. The 0.25° version produces higher-resolution output but does not uniformly improve forecast diagnostics.

For operational users, the strongest near-term case is therefore not “replace the ocean model.” It is to use faster emulation where many daily 60-day scenarios are valuable, while qualifying each output against diagnostics relevant to the downstream decision. This remains a conditional pathway: Neptune is driven by prescribed atmospheric forcing, relies mainly on one ocean-reanalysis family for training and testing, and its high-resolution variant was not evaluated as an ensemble.

A 60-day forecast is only useful if the right ocean behavior survives

Suppose a planning system can generate two months of global ocean states cheaply enough to run many scenarios instead of one expensive simulation. Average forecast error would be an obvious first test. It would not be enough.

A fisheries decision may depend on regional currents. An energy or insurance workflow may care about large-scale thermal conditions. A polar operation may be sensitive to the location of an ice edge. A model can perform well in aggregate while being much weaker on one of these quantities.

That is the practical tension in Neptune: An AI model for Global Ocean Subseasonal Prediction.¹ Neptune-1 generally improves on a climatological baseline for most predicted fields, yet its errors grow unevenly as forecasts approach 60 days. Currents, mixed-layer depth, sea-ice thickness, energetic boundary-current regions, and other rapidly varying structures are harder cases.

The relevant question is therefore not simply whether the emulator is accurate. It is which parts of ocean evolution remain usable for a particular decision.

Neptune combines local structure with global interaction

Ocean forecasting has two spatial requirements that are difficult to satisfy with one representation. Local structure matters for fronts, currents, coastlines, and other geographically concentrated features. At the same time, ocean states evolve through interactions that span very large distances.

Neptune addresses both within a 13-million-parameter encoder-operator-decoder architecture. Convolutional blocks process local multiscale structure. Between them, two Spherical Fourier Neural Operator blocks represent global interactions through spectral operations defined for spherical geometry.

The model does not directly predict an entire future ocean state from scratch. Given two recent ocean states and two atmospheric-forcing states, it predicts residual changes for the following two ocean states. These predictions are then rolled forward autoregressively.

Atmospheric forcing includes near-surface temperature, dewpoint, winds, and mean sea-level pressure. Latitude, longitude, day of year, and day length provide geographic and seasonal context. Gaussian-noise conditioning allows Neptune-1 to generate stochastic ensemble members rather than only one trajectory.

That ensemble matters because a 60-day planning system should not be judged solely by how close one path lands to the realized future. It should also be tested on whether the distribution of possible trajectories remains informative. The paper therefore evaluates Neptune-1 with both conventional error measures and CRPS, a probabilistic score for forecast distributions.

Large-scale ocean memory survives better than fast dynamics

The held-out benchmark uses 2019-2021, with every test day serving as the initial condition for a 60-day rollout. The climatological comparison is constructed only from the 1993-2016 training period, avoiding leakage from the test years.

Across this benchmark, Neptune-1 generally has lower RMSE than climatology for most output variables. Its ensemble also produces better CRPS than climatology for most variables through much of the forecast window.

But the aggregate result conceals a strong hierarchy of predictability.

Temperature is the clearest case. Its anomaly correlation remains above 0.75 across all evaluated depths through 30 days and above 0.5 after 60 days. Sea-surface-height structure is also persistent in important large-scale regions: its anomaly correlation remains above 0.9 after 60 days in the Indian Ocean and equatorial Pacific.

Currents and mixed-layer depth degrade more substantially. Skill is also spatially weaker in highly chaotic systems including western boundary currents, the Antarctic Circumpolar Current, and the North Atlantic subpolar gyre.

This distinction becomes clearer when the paper tests ocean behavior rather than gridpoint error alone.

Upper-ocean heat content remains strongly coherent through the rollout. The lowest reported correlation between Neptune and the reference heat-content time series is 0.988 at day 60. Yet the same analysis shows progressive damping of extremes and regional heat biases. High correlation therefore indicates preservation of broad temporal evolution, not exact reproduction of every anomaly.

The velocity fields reveal a different failure mode. Eddy kinetic energy measures variability associated with horizontal current anomalies. Neptune preserves large spatial structures and spectral components around wavelengths of roughly $10^3$ km, but increasingly smooths the highest-frequency energetic variability. A forecast can therefore retain the broad circulation pattern while losing dynamically important fine structure.

Physical diagnostics turn accuracy into a qualification problem

The paper’s evaluation is more operationally informative because it asks several different questions of the same forecast.

The Indian Ocean Dipole illustrates the value of this broader validation. Its spatial correlation is about 0.99 at day 10 and 0.89 at day 60. The model also retains useful Z20 structure associated with equatorial Pacific thermocline behavior, although correlations are stronger in the western Pacific than in the more wind-driven eastern region.

These tests are not separate demonstrations of unrelated capabilities. Their likely purpose is diagnostic: they ask whether low statistical error corresponds to preservation of ocean quantities that downstream users may actually depend on.

For a business deployment, Cognaptus would treat that evaluation style as a model-qualification template. A user should identify the physical quantity tied to the decision, then validate that quantity at the relevant lead time and geography rather than inheriting confidence from a global headline metric.

A 0.25° grid is not automatically a better forecast

Neptune also tests whether its architecture can move from 1° pre-training to 0.25° output. Neptune-025 is fine-tuned for 50 epochs on native 0.25° data while retaining a 1° latent working representation through spherical resampling.

The result demonstrates resolution transfer: the model can produce substantially finer output without rebuilding its global latent architecture.

It does not demonstrate that increasing nominal resolution uniformly increases forecast fidelity.

Neptune-025 broadly preserves the spatial behavior of Neptune-1, but the paper reports slightly higher RMSE in several cases, a small cooling and declining-ocean-heat-content tendency, and continued under-representation of high-frequency eddy kinetic energy. The authors associate these results partly with limited fine-tuning and model capacity.

There is also an evaluation asymmetry. Neptune-1 is tested as an eight-member stochastic ensemble. Neptune-025 is not, because of computational constraints. Its higher grid resolution therefore comes with less evidence about probabilistic performance.

For an operational buyer, these are three separate attributes: output resolution, predictive fidelity, and uncertainty quality. None should be inferred automatically from another.

The business case is cheaper scenario generation, subject to qualification

The paper does not measure business outcomes, so the commercial interpretation is necessarily an inference rather than a reported result.

The plausible advantage is computational. If neural emulation makes daily 60-day ocean trajectories sufficiently cheap, organizations in fisheries, energy planning, insurance, water management, agriculture, or disaster-risk workflows could potentially examine larger scenario sets than would be practical with expensive physics-based runs alone.

The qualification layer is equally important. A fisheries workflow dependent on currents should not inherit confidence from temperature skill. A sea-ice use case should inspect polar-region behavior and ice-edge error. A climate-risk workflow centered on large-scale thermal anomalies may find the heat-content and climate-index results more relevant than mesoscale fidelity.

This is where Neptune’s evaluation design may be as consequential as its architecture: it provides an example of testing an emulator against physical behavior that maps more directly to downstream decisions.

The current benchmark stops short of operational S2S forecasting

The largest boundary is atmospheric forcing. Neptune receives prescribed atmospheric fields rather than forecasting them jointly. Its 60-day ocean skill therefore does not include errors that would arrive from an imperfect future atmosphere, nor does it establish performance for a fully coupled atmosphere-ocean system.

The evidence is also concentrated in one primary ocean-reanalysis family, with ERA5 atmospheric forcing and a three-year held-out period. External reanalyses or observational regimes could expose different errors.

Finally, the model remains weaker where ocean dynamics are fast and fine-scaled. Fourteen depth levels and relatively coarse upper-ocean vertical discretization limit mixed-layer representation, while rectangular-grid convolutions in the encoder and decoder coexist with spherical operators only in the latent core.

These are not reasons to dismiss the emulator. They specify where additional validation is required before a Neptune-like model becomes part of a consequential forecasting workflow.

Faster forecasts change the validation burden

Neptune makes a credible case that a compact data-driven model can preserve substantial large-scale ocean structure through daily 60-day autoregressive forecasts. It also makes clear that long-horizon skill is not one number.

Temperature, heat content, sea level, climate indices, currents, mesoscale variability, and sea ice do not degrade at the same rate. Moving from 1° to 0.25° does not remove that heterogeneity.

For operators, faster emulation therefore shifts rather than eliminates the validation problem. Once generating scenarios becomes cheaper, the critical task is deciding which modeled quantities are reliable enough, at which lead times and locations, to enter a real decision process.

Cognaptus: Automate the Present, Incubate the Future.