A Korean research team developed an artificial intelligence (AI) model that learns 160-plus years of global ocean data and predicts ocean conditions for the next 200 days in a matter of seconds. Because it enables long-range ocean forecasting with far fewer computing resources than traditional supercomputer-based numerical models, it is drawing attention as a foundational technology to improve responses to climate change and the accuracy of long-term climate forecasts.
Korea Institute of Science and Technology (KIST) Climate and Carbon Cycle Research Center researchers Kang Dae-hyeon and Kim Jeong-hwan said on the 26th that they developed an AI-based global 3D ocean prediction model, KIST-Ocean. The findings were published in the journal Science Advances.
As climate change brings more frequent extreme weather such as heat waves, heavy downpours, droughts, and typhoons, the importance of long-range climate forecasting is growing. In particular, the ocean, which covers about 70% of Earth's surface, stores and circulates vast amounts of heat and carbon and plays a key role in climate variability such as El Niño and La Niña.
But existing ocean prediction models must compute complex physical equations, requiring massive supercomputing resources and long run times. That has limited rapid forecasts and analysis of various climate change scenarios.
To overcome these limits, the team pre-trained a Deep Learning AI model, KIST-Ocean, on global large-ensemble simulation data from 1850 to 2014 generated by the Community Earth System Model 2 (CESM2), and fine-tuned it with ocean reanalysis data from 1982 to 2013 to reflect ocean characteristics.
KIST-Ocean uses 62 ocean variables—including not only sea surface temperature but also water temperature, currents, and salinity—to predict 3D ocean states down to a depth of 600 meters at five-day intervals. It is also designed to incorporate six surface boundary conditions transmitted from the atmosphere to the ocean, including wind stress, longwave and shortwave radiation, and latent and sensible heat fluxes.
The team said KIST-Ocean completed pre-training in about 33 hours and fine-tuning in about 2.4 hours using just one Nvidia A100 GPU, and generated global ocean forecasts for 200 days in 6–7 seconds, greatly reducing computation time and expense compared with numerical models.
In virtual experiments that verify how accurately it reproduces real ocean physics, KIST-Ocean also showed that ocean waves and upwelling/downwelling driven by atmospheric changes matched established ocean physics theory. In the 2015 super El Niño case, it successfully reproduced the rise in sea surface temperatures in the equatorial Pacific and changes in internal ocean heat distribution.
The team said the results show that KIST-Ocean goes beyond simply learning past data and can effectively capture the complex physical interactions between the atmosphere and the ocean.
In experiments using actual atmospheric conditions, KIST-Ocean achieved higher accuracy in 200-day forecasts than a persistence forecast, which assumes the current ocean state remains unchanged. For sea surface temperature forecasts, it outperformed the North American Multi-Model Ensemble (NMME) for up to six months.
The team plans to advance the current ocean-only model into an AI-based Earth system model integrating the atmosphere, ocean, sea ice, and land. They expect this will enable faster and lower-expense analysis of various climate change scenarios as well as long-term climate forecasts.
Kang Dae-hyeon said, "This study shows that AI can not only deliver excellent computational efficiency but also realistically reproduce the complex physical relationships between the atmosphere and the ocean," adding, "It will lay the groundwork for developing AI-based Earth system models and strengthening capabilities to respond to the climate crisis."