South Korean researchers have built an AI model that predicts three-dimensional ocean conditions 200 days into the future in roughly six seconds. The model, called KIST-Ocean, was developed at the Korea Institute of Science and Technology (KIST) and described in a paper published in Science Advances. It produces forecasts that would take hours of supercomputer time using conventional numerical methods, potentially reshaping how climate scientists approach long-range prediction.
The model uses deep learning trained on global ocean simulations from the U.S. Community Earth System Model 2 (CESM2) covering 1850 to 2014, then fine-tuned with ocean reanalysis data from 1982 to 2013. This two-stage training lets KIST-Ocean capture both long-term climate behavior and observed ocean dynamics. Pretraining took about 33 hours and fine-tuning roughly 2.4 hours on a single NVIDIA A100 graphics processor.
How the forecasting works
KIST-Ocean takes three-dimensional ocean conditions and atmospheric boundary conditions as inputs and predicts the ocean's 3D state five days ahead. The model then feeds each predicted ocean state back into itself, repeating this process up to 40 times. The result is a global three-dimensional ocean forecast at five-day intervals stretching out 200 days. The model predicts 62 ocean variables-including sea surface temperature, subsurface temperature, ocean currents, and salinity-down to depths of 600 meters. Six atmospheric boundary conditions, such as wind stress and heat fluxes, account for interactions between the atmosphere and the ocean.
Unlike persistence forecasts that assume current conditions remain unchanged, KIST-Ocean learns the physical relationships governing atmosphere-ocean interactions. In validation tests, it outperformed the North American Multi-Model Ensemble (NMME) for sea surface temperature predictions extending up to six months.
Capturing real-world ocean events
The model accurately reproduced key physical ocean processes, including wind-driven waves, upwelling, and downwelling. During the 2015 super El Niño event, KIST-Ocean successfully captured changes in sea surface temperatures and subsurface heat distribution. The performance suggests the model has internalized not just historical patterns but the underlying physics of how the atmosphere and ocean interact.
Lead researcher Kang Dae-hyun said, "The study shows that AI can reproduce the complex physical interactions between the atmosphere and the ocean while delivering exceptional computational efficiency. We believe it will provide a foundation for next-generation AI Earth system models and strengthen our ability to respond to the climate crisis."
A step toward AI-driven Earth system models
The KIST team said KIST-Ocean represents progress toward AI-powered Earth system models that integrate the atmosphere, oceans, sea ice, and land. Such systems could enable faster, lower-cost analysis of long-term climate trends and a wider range of climate change scenarios. The work adds to a growing body of AI for Science & Research applications where deep learning models tackle problems once reserved for physics-based simulations.
Conventional numerical ocean models solve physical equations on supercomputers, a process that consumes substantial time and energy for each forecast run. KIST-Ocean sidesteps that bottleneck. Once trained, the model generates a 200-day global forecast in six to seven seconds on a single GPU. That speed could let researchers run many more climate scenarios in the same computing window, exploring a wider range of possible futures without waiting days for results.
Why this matters for science and research professionals
For researchers working with climate models, oceanography, or Earth system science, KIST-Ocean signals a shift in how long-range forecasting might be done. A model that produces 200-day forecasts in seconds on a single GPU removes a major computational barrier. The approach is not limited to oceans-the same iterative deep learning framework could extend to other components of Earth system models. Research scientists looking to apply similar methods in their own domains can explore structured training through an AI Learning Path for Research Scientists that covers the techniques behind models like this one.
The KIST-Ocean code and methodology, now published in a peer-reviewed journal, give other research groups a reference architecture to build on. As extreme weather events become more frequent, tools that speed up climate analysis without sacrificing accuracy will matter more. The six-second forecast is not just a benchmark-it is a practical signal that AI-based climate modeling is moving from proof-of-concept toward operational use.
Your membership also unlocks:
