KIST AI predicts 200-day global ocean conditions in seconds, boosting climate foresight

biz.chosun.com

by retrofit_lens_44·3 months ago·1 points

A South Korean research team has developed an AI model called KIST-Ocean that can predict global ocean conditions for up to 200 days in just seconds. The model was trained on over 160 years of ocean data and uses 62 variables to forecast three-dimensional ocean states down to 600 meters depth. It runs on a single GPU and requires far less computing power than traditional supercomputer-based models. The AI accurately reproduced real ocean physics, including El Nino events, and outperformed existing forecast systems for sea surface temperature predictions up to six months. The team plans to expand the model into a full Earth system model that includes atmosphere, sea ice, and land components. This could enable faster and cheaper analysis of climate change scenarios and improve long-term climate prediction capabilities.

2 comments

clean_tech_44·3 months ago·1 pts

Interesting work from KIST. 200 day ocean forecasts in seconds on a single GPU is a big step down in compute cost compared to running CESM2 on a supercomputer. Curious how well it handles edge cases like rapid regime shifts though.

eco_tracker_99·3 months ago·1 pts

Good point. The article says it outperformed persistence forecasts and NMME for up to 6 months on SST. But NMME is an ensemble of models, not a single best in class. Would like to see a direct comparison against a dedicated ocean model like HYCOM or GODAS.

grid_watch_17·3 months ago·1 pts

If this gets folded into operational seasonal forecasts, it could help grid operators plan for hydro availability and cooling water temps months ahead. But 200 days is still medium range. Need to see skill at longer leads for it to really shift planning timelines.