Modern neural network weather models already outperform traditional numerical schemes in accuracy and speed. But for seasonal and multi-year climate forecasts, you need the ocean: its thermal inertia sets the memory for the whole system. This work sits at the fringe: there is almost no one else nearby, with the closest neighbors either tackling isolated aspects of machine learning for geophysics or narrow tasks like predicting El Niño from time series. KIST-Ocean takes a step toward a full three-dimensional ocean model that could become part of a neural network Earth model.
The model is built on a U-shaped architecture with visual attention (VAN), which efficiently captures multi-scale structures with a small number of parameters. The input consists of 62 oceanic variables (temperature, salinity, currents at 15 depth levels) and 6 surface boundary conditions (wind stress, heat fluxes). Training proceeded in two stages: pretraining on 320 years of CESM2 Large Ensemble simulations, then fine-tuning on 32 years of GODAS/OISST/ERA5 reanalysis. To combat distribution drift in autoregressive forecasting, adversarial training (GAN) with a PatchGAN discriminator was used. Partial convolution allowed correct handling of coastlines by ignoring land.
KIST-Ocean outperformed the persistence forecast for 81.3% of targets in terms of anomaly correlation coefficient (ACC) and for 86.5% in terms of RMSE. When forced with real boundary conditions, the model maintained significant skill up to 200 days for potential temperature, up to 85 days for zonal current at 145 m depth, up to 40 days for meridional current, and up to 100 days for salinity. Comparison with five NMME models showed that KIST-Ocean gives better sea surface temperature forecasts up to six months ahead under real forcing, and under climatological forcing it matches the best dynamical models for one to two months. In idealized experiments, the model reproduced eastward propagation of Kelvin waves and westward propagation of Rossby waves, with Rossby wave speeds matching theoretical predictions for different latitudes. Cyclonic wind stress caused upwelling of cold water and surface cooling, while anticyclonic wind stress caused downwelling and warming, as predicted by Ekman theory.
This work shows that a neural network can learn not only statistical patterns but also the physical mechanisms of ocean dynamics. That sets it apart from neighbors that either model only individual variables or do not verify physical consistency. KIST-Ocean is the first demonstration that deep learning can reproduce wave and Ekman responses to atmospheric forcing, opening the path to neural network coupled ocean–atmosphere models.
The next logical step is to fully couple KIST-Ocean with a neural network atmosphere model so that ocean and air exchange fluxes at every step. This would allow forecasting El Niño and other climate phenomena years ahead without expensive numerical models. Extension to sea ice and biogeochemistry is also possible.
Improved seasonal and interannual climate forecasts, especially for El Niño and La Niña events, which affect droughts, floods, and hurricanes worldwide. The model could also aid fisheries and shipping through current forecasts.
Quantitatively assess wind stress responses and compare them with analytical solutions. Develop a fully coupled neural network ocean–atmosphere model and test it on El Niño forecasts.
For oceanography, an open question remains how well neural network models can reproduce long-period variability such as the Atlantic Meridional Overturning Circulation or the Pacific Decadal Oscillation. KIST-Ocean was trained on 32 years of reanalysis, which may be insufficient to test these slow processes.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
KIST-Ocean is a deep neural network that models three-dimensional ocean circulation with a five-day time step. It is trained on reanalysis and climate simulations and validated on forecasts up to 200 days. It reproduces the physics of Kelvin and Rossby waves and the response to wind stress according to Ekman theory.
Where its strength is
The model does not just predict; it reproduces known physical mechanisms of ocean dynamics, as confirmed by comparison with theory. This paves the way for coupled neural network models of the ocean and atmosphere.
Where it could go next
Worth trying
Why this may matter
If neural network ocean models can reliably reproduce key physical processes, this will accelerate the development of fast and accurate climate forecasts, which are important for adaptation to change.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.53, against a usual 0.64. It looks like few have drilled nearby.
The model has only 6.6 million parameters, tens of times fewer than large language models, yet it is enough to simulate the ocean on a 1°×1° grid with 15 depth levels. On a single A100 GPU, a 200-day forecast takes about 7 seconds.
In Kim Stanley Robinson's novel 'The Ministry for the Future', AI-driven climate models are used to predict catastrophic heat waves. KIST-Ocean is a step toward such systems: a neural network that understands ocean physics rather than just fitting statistics.