Accurate ocean modeling is essential for predicting climate variability and adapting to change. Traditional dynamical models rely on physical parameterizations that limit their accuracy and are slow to exploit growing observational data. Existing machine learning approaches have focused on individual phenomena, such as El Niño, or on limited regions. This work stands apart from neighboring studies in the archive: those either address specific climate modes or the atmosphere, or use AI to accelerate individual components, whereas here a global three-dimensional ocean model is built on a purely data-driven basis for the first time.
ORCA-DL processes the initial ocean state (temperature, salinity, currents) and atmospheric forcing through zonal and meridional wind stress. The architecture includes separate encoders for each ocean variable, a fusion module that accounts for lead time, and decoders. The model was trained on simulations from 20 CMIP6 models, tuned on SODA and ORAS5 data up to 1980, and tested on independent GODAS data from 1980 to 2018. To enhance reliability, an ensemble of 10 models with different random initializations was used.
ORCA-DL reproduces the mean ocean state with less bias than the dynamical model NUIST-CFS1.2: the RMSE for sea surface temperature was 0.63°C, compared with 0.83–1.06°C. The forecast of the Niño 3.4 index at 12 months ahead achieves a correlation of 0.75, overcoming the spring predictability barrier, and significant correlation persists up to 20 months. The model also outperforms dynamical counterparts in predicting upper-ocean marine heatwaves and demonstrates a stable decadal simulation of global sea surface temperature with a correlation of 0.77.
Unlike neighboring works that either specialize in a single phenomenon or use AI as an auxiliary tool within a dynamical model, ORCA-DL shows for the first time that a purely data-driven approach can encompass the global three-dimensional ocean and compete with the best dynamical systems on seasonal and decadal scales. This opens the way to replacing expensive components of climate models with fast emulators.
Promising directions include incorporating external forcings such as CO2 concentrations and additional atmospheric variables (heat fluxes, precipitation, humidity). Transfer learning could be applied to adapt to observational data, and sensitivity experiments could identify key regions for predictability.
It will affect operational seasonal forecasting, risk assessment of marine heatwaves, and decadal climate services.
Add carbon dioxide and other radiative forcings to the model, and explore transfer learning to real data.
The work is directly related to the predictability of El Niño and decadal ocean variability, as well as to reducing systematic biases in climate models.
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
The authors introduced ORCA-DL, the first purely data-driven model of global ocean circulation for seasonal and decadal forecasting. The model was trained on simulations from 20 CMIP6 models, tuned on SODA and ORAS5 data, and tested on independent GODAS data. ORCA-DL outperforms dynamical models in predicting El Niño, marine heatwaves, and subsurface anomalies, and remains stable on decadal timescales.
Where its strength is
The main strength of this work is that it demonstrates for the first time that a purely data-driven approach can capture the global three-dimensional ocean and compete with the best dynamical systems on seasonal and decadal timescales. This paves the way for replacing expensive components of climate models with fast emulators.
Where it could go next
Worth trying
Why this may matter
If data-driven ocean models achieve parity with dynamical ones on all timescales, this would radically accelerate and reduce the cost of climate forecasting, making it accessible to a broader range of adaptation tasks.
Nearby works in our archive
The closest work in the archive sits at 0.61 by meaning — the usual level for our corpus.
ORCA-DL was trained in just 12 hours on four A100 GPUs, whereas a single dynamical model can require weeks of computation on a supercomputer.
In Kim Stanley Robinson's novel 'The Ministry for the Future', AI-driven climate models are used to save the planet; ORCA-DL is a step toward such tools.