Extreme weather is becoming more frequent and intense as the climate warms, but one of the planet’s most important drivers of climate variability remains difficult to predict: the ocean. Covering roughly 70 percent of Earth’s surface, the ocean absorbs and redistributes enormous quantities of heat and carbon, shaping atmospheric conditions from one season to the next. Phenomena such as El Niño and La Niña are closely linked to these ocean–atmosphere interactions, yet conventional ocean forecasting systems often require powerful supercomputers and lengthy calculations to solve complex physical equations. A new artificial intelligence model developed in South Korea could dramatically accelerate that process.
Researchers at the Korea Institute of Science and Technology (KIST) have developed KIST-Ocean, a data-driven global ocean prediction model designed to reproduce three-dimensional ocean conditions and forecast how they will evolve. The system learns from decades of atmospheric and oceanic observations, as well as simulated data, rather than calculating every physical process from first principles each time a forecast is produced. According to the research team, the model can generate an ocean forecast extending approximately 200 days into the future in only a few seconds using a single graphics processing unit, or GPU.
KIST-Ocean is trained to work with multiple physical variables that describe the state of the global ocean, including sea-surface temperature, salinity, currents and subsurface heat distribution. It predicts how these variables will change over five-day intervals and resolves ocean conditions down to a depth of 600 meters. The model receives a three-dimensional ocean state and atmospheric boundary conditions as its initial input. It then predicts the ocean’s condition five days later, feeds that prediction back into the system, and repeats the process up to 40 times. This iterative approach produces a global forecast covering nearly seven months at regular five-day intervals.
The researchers say the model’s speed could transform how scientists investigate climate risk. Traditional numerical ocean models use detailed equations governing fluid motion, heat transfer and other physical processes. Although these systems are scientifically powerful, their calculations are computationally demanding, particularly when researchers need to run hundreds or thousands of simulations to examine possible climate scenarios. KIST-Ocean replaces much of that repeated calculation with a trained neural model that has learned statistical relationships embedded in historical and simulated ocean data. The result is a system capable of rapidly generating forecasts and large ensembles at a fraction of the usual computational cost.
Speed alone, however, does not guarantee scientific value. To test whether the artificial intelligence system had learned meaningful ocean dynamics rather than merely reproducing familiar patterns, the research team conducted experiments involving atmospheric forcing. In a virtual wind-generation experiment, changes in wind produced ocean responses including waves, upwelling and downwelling. These processes are central to ocean physics: upwelling carries colder, nutrient-rich water toward the surface, while downwelling transports surface water and heat into deeper layers. The behavior generated by KIST-Ocean was consistent with established physical theories, suggesting that the model captured important links between the atmosphere and the ocean.
The team also tested KIST-Ocean against the development of the 2015 Super El Niño, one of the most powerful El Niño events recorded. During El Niño, unusually warm surface waters spread across the equatorial Pacific, altering atmospheric circulation and influencing weather patterns across much of the world. The model reproduced key features of the event, including the warming of the equatorial Pacific and changes in the distribution of heat beneath the surface. These results provided evidence that the system can represent both visible surface changes and the hidden subsurface processes that help drive long-lasting climate variability.
The significance of the technology extends beyond faster ocean maps. Seasonal and annual forecasts depend heavily on the ocean because seawater changes more slowly than the atmosphere and can preserve climatic information for months. A model that can rapidly update three-dimensional ocean conditions could help researchers explore the likelihood of prolonged heatwaves, droughts, heavy rainfall or shifts in typhoon behavior. It could also support early-warning systems by allowing scientists to test many possible atmospheric and oceanic developments rather than relying on a small number of expensive simulations.
KIST-Ocean may also become a building block for broader artificial intelligence-based Earth system models. Such systems would combine the atmosphere, ocean, land surface, ice and carbon cycle in a unified framework. Integrating these components is technically challenging because each operates on different timescales and interacts through complex feedbacks. A fast ocean component could make it easier to conduct the repeated experiments needed to study those connections, while reducing the computing resources required for climate research. The researchers believe this could lower barriers for institutions that do not have access to the largest supercomputing facilities.
The team cautions that artificial intelligence does not eliminate the need for observations, physical understanding or continued model evaluation. AI forecasts depend on the quality and range of the data used during training, and unusual conditions outside that historical experience can test the limits of any data-driven system. For that reason, the researchers evaluated whether KIST-Ocean reproduced recognized physical mechanisms, not just whether its numerical predictions matched past datasets. Dr. Kang Daehyun, who led the work at KIST’s Center for Climate and Carbon Cycle Research, said the results show that AI can achieve both computational efficiency and a realistic representation of atmosphere–ocean relationships. The team now plans to refine the model as a practical forecasting tool aimed at improving preparedness for climate-related disasters and reducing their social and economic costs.
Subject of Research: AI-based global ocean forecasting and atmosphere–ocean dynamics
Article Title: Data-driven global ocean model resolving atmospherically forced ocean dynamics
News Publication Date: 12-Jun-2026
Web References: https://doi.org/10.1126/sciadv.aed1225
References: Science Advances, DOI: 10.1126/sciadv.aed1225
Image Credits: Korea Institute of Science and Technology
Keywords
KIST-Ocean, artificial intelligence, ocean forecasting, climate prediction, El Niño, ocean dynamics, climate change, machine learning, Earth system models, seasonal forecasting
