The Ocean Remembers: Meet Neptune, the AI That Sees 60 Days Into the Sea

Two weeks is the wall. Beyond it, weather forecasts — the ones your phone trusts — dissolve into guesswork, because the atmosphere forgets itself that quickly. The ocean does not. Its currents, its layers of warm and cold water, its slow churning carry memory on timescales of weeks and months, and that memory is exactly what decides whether a drought turns into a crisis, whether a marine heatwave cooks a fishery, whether monsoon rains arrive on time. The problem has always been that simulating the ocean is brutally expensive: physics-based models that solve the governing equations of fluid motion strain the world's fastest supercomputers just to run a single forecast, never mind the hundreds of ensemble members needed to capture uncertainty.

In a new preprint from researchers at the University of Salento, CMCC, and Columbia University, an AI model called Neptune shows this wall can be crossed. Given daily weather as a boundary condition, Neptune emulates the entire global ocean and its sea ice — temperature through 14 depth levels, salinity, currents, sea surface height, ice thickness and concentration — with daily output up to 60 days ahead. Trained on observational reanalysis rather than physical equations, it runs orders of magnitude faster than traditional Ocean General Circulation Models (OGCMs), reproduces climate phenomena like El Niño and the Indian Ocean Dipole, and stays stable over long rollouts. The two-week wall, the authors argue, is not physics — it's computing.

The Science

Subseasonal-to-seasonal (S2S) forecasting is the awkward middle child of climate science: longer than weather, shorter than climate. It sits in the weeks-to-months window where decisions get made — planting, water releases, evacuation orders, energy trading, insurance pricing (Vitart, 2018). The reason this window is predictable at all is the ocean. Because water absorbs and redistributes heat slowly, the ocean retains "memory" of past conditions far longer than the atmosphere, and that persistence is the engine of forecast skill after the first two weeks (Balmaseda, 2026).

Traditionally, scientists represent this memory with Ocean General Circulation Models — computer programs that solve the partial differential equations of fluid dynamics and thermodynamics over a global grid. They work, but they are heavy. Their code is intricate, they are hard to improve, and each forecast costs enormous compute. Running the large ensembles needed for probabilistic S2S prediction is often impractical (Bauer et al., 2015).

The deep-learning revolution that transformed weather forecasting — models like GraphCast, PanguWeather, and GenCast, which became hundreds of times faster than their numerical counterparts — has been slower to reach the ocean. The reasons are real. Land and ice define ocean basins, creating complex coastal dynamics the atmosphere doesn't have. The deep ocean is largely invisible to remote sensing, with historically sparse observations (Amirian et al., 2026). And ocean processes evolve on much longer timescales than atmospheric ones, so AI models optimized for short-term feedback tend to drift into instability (Bracco et al., 2025).

Earlier data-driven ocean models made inroads but with compromises. ORCA-DL (Guo et al., 2025) works at monthly resolution; Samudra (Dheeshjith et al., 2025) smooths its data with a Gaussian filter and five-day averages; both run at a coarse one-degree grid. Neptune's authors set out to close the gap between weather-scale and climate-scale AI, targeting fine spatial and temporal detail at S2S lead times.

The architecture blends two ideas. Convolutional Neural Networks (CNNs) — the workhorse pattern-recognition layers familiar from image processing — capture local features, like water masses and coastal currents. Spherical Fourier Neural Operators (SFNOs) do global math on the sphere, computing spectral transforms that capture long-range cross-scale interactions while remaining computationally efficient (Bonev et al., 2023). This hybrid lets Neptune see both the eddy in the Gulf Stream and the entire Pacific basin at once, in the same forward pass. The model is a U-Net-shaped encoder-decoder, roughly 13 million parameters, that projects ocean and atmosphere fields into a 256-dimensional latent space, passes them through the SFNO blocks, and reconstructs two consecutive predicted ocean states (Donno et al., 2026).

A defining design choice is that Neptune predicts residuals — the change between consecutive ocean states — rather than full fields, which the authors found improves forecasting skill. It also conditions each forecast on the day-of-year (encoding seasonality) and on injected random noise, which lets it generate ensemble forecasts: run the same initial condition many times with different noise, and you get a spread of plausible futures that captures uncertainty.

Neptune was trained on two complementary datasets. The ocean side came from ORAS5, a reanalysis of global ocean and sea ice from the Copernicus Marine Service spanning 1993–2022 at 0.25° native resolution, with 14 selected depth levels reaching to about 5,000 meters. The atmosphere came from ERA5, the gold-standard European reanalysis, with five surface forcing variables: 2-meter temperature and dewpoint, 10-meter wind components, and mean sea level pressure. The model was pre-trained at a coarse 1° resolution (Neptune-1) for 150 epochs, then fine-tuned at the native 0.25° resolution (Neptune-025) for 50 more. Training ran on 8 NVIDIA A100 GPUs on CMCC's Juno supercomputer — about 48 hours for pre-training and 100 for fine-tuning. The researchers evaluated on the 2019–2021 test period, launching a 60-day forecast from each day as an initial condition.

What They Found

The headline result is stability plus skill. Across a battery of statistics — Root Mean Square Error (RMSE, how far predictions stray from truth), CRPS (a probabilistic score that rewards well-calibrated uncertainty), and Anomaly Correlation Coefficient (ACC, how well the pattern of deviation from normal matches reality) — Neptune outperforms the naive climatology baseline and reproduces the spatiotemporal evolution of ocean fields through 60 days. For the fine-tuned 0.25° version, resolution is sharp enough to represent mesoscale structure — the eddies and jets that carry heat and nutrients (Donno et al., 2026).

Physical coherence, not just raw numbers, was a central test. The study tracked Ocean Heat Content (OHC), Eddy Kinetic Energy (EKE), and an ice-specific Brier score, checking that the model conserves heat, maintains realistic turbulent energy, and gets sea-ice probabilities calibrated. Neptune preserved OHC across the forecasting horizon, a sign that it isn't silently leaking energy. It faithfully captured large-scale climate modes — the El Niño–Southern Oscillation, the depth of the 20°C isotherm (the Z20 metric, a key ENSO thermometer), and the Indian Ocean Dipole — meaning the model produces the phenomena that actually drive seasonal impacts, not just smooth averages that look right (Donno et al., 2026).

Forecast error at 60-day lead: Neptune vs. climatology

Label	Value
Neptune-1 RMSE	0.24
Climatology RMSE	0.37
Neptune-025 RMSE	0.2

Approximate RMSE values illustrating Neptune's better-than-climatology skill at 60-day lead times (Donno et al., 2026).

The authors emphasize that earlier global emulators sacrificed detail: monthly timesteps, smoothed fields, coarse grids. Neptune-025 operates at daily frequency and 0.25° resolution — a genuine bridge from the weather to the climate regime. The ensemble capability, enabled by noise conditioning, is not cosmetic: probabilistic S2S prediction requires running many perturbed forecasts, and an emulator that produces them cheaply is precisely what operational centers need.

Why This Changes Things

The most important implication is economic. S2S forecasts are not an academic curiosity; they are a decision tool for agriculture, water management, fisheries, energy planning, disaster risk reduction, and insurance (Troccoli, 2010; Meehl et al., 2021). But a tool is only useful if you can run it when you need it. OGCM-based S2S systems are so compute-hungry that quick evaluations for end users and large forecast ensembles are often impractical. A model that emulates the ocean hundreds of times faster changes the calculus — you can afford the ensemble, the sensitivity runs, the what-if scenarios, the running updates that a responsive service demands.

The contrast with existing AI ocean emulators also matters for how science progresses. Models like Samudra that smooth their inputs and average over five days remove the very high-frequency detail — the daily swings of currents and ice — that can matter for marine and coastal applications. Neptune's daily output at fine resolution keeps that detail while remaining stable, which the authors attribute to the residual-prediction formulation and the hybrid CNN-SFNO architecture.

There is a broader institutional angle worth noting. The work is a collaboration between a university (Salento), a leading European climate center (CMCC), and Columbia's LEAP center — an example of the research infrastructure that data-driven climate science increasingly needs. The model was built and evaluated on CMCC's own supercomputer, a reminder that training and validating these models is a substantial computing enterprise in its own right, even if inference is cheap.

Crucially, the authors are careful about what Neptune is not. It is an ocean and sea-ice component — an emulator of one part of an S2S system — driven by prescribed atmospheric forcing, not a self-contained forecast model. And like all deep-learning climate models, it trades some physical interpretability for speed: the physics is learned, not solved, so the model is best understood as a fast, accurate approximation whose internal reasoning is a black box (Bauer et al., 2015). For a system whose dynamics are this complex and this consequential, that trade-off deserves acknowledgment, not dismissal.

What's Next

The paper is careful to point out what remains open. The authors note that, due to computational constraints during evaluation, they could not run the full ensemble for the higher-resolution Neptune-025 model — only for Neptune-1. Extending the ensemble to the fine resolution is an obvious next step. So is coupling: Neptune was forced by prescribed atmospheric fields, so a natural direction is to pair it with an AI atmosphere model into a fully data-driven coupled S2S system, the oceanic mirror of the atmosphere-only revolution. The longer-term stability demonstrated by models like Samudra (centuries) hints that rolling Neptune forward well past 60 days deserves investigation, though the authors are appropriately modest about extending claims beyond the S2S window for which it was designed.

The deeper significance is architectural. For years, the argument for data-driven climate prediction has been that it reproduces the average behavior of numerical models. Neptune is part of a newer, more ambitious wave claiming it can stand inside a production forecasting pipeline — at high resolution, at daily frequency, with calibrated uncertainty, out to a horizon that actually matters for decisions. If that claim holds up to broader evaluation, the role of physics-based OGCMs shifts from the workhorse of every forecast to a validation and training substrate — the thing you use to teach the fast model, then let the fast model run the world's drills.

Two weeks was never the limit of our knowledge. It was the limit of our patience with brute-force computation. Neptune is evidence that a faster, cheaper, daily-resolution ocean emulator — one that remembers El Niño and the Indian Ocean Dipole and the slow churn of the deep — can cross it, and bring the decisions on the other side within reach.