ETH Zurich Feeds 100 Petabytes of NASA Data Into a Supercomputer to Speed Up Disaster Forecasting

A revolution in forecasting natural disasters is under way, according to researchers in Switzerland who are training AI models on vast troves of NASA climate data to produce potentially life-saving forecasts at far greater speed. By feeding NASA's data archive into one of the world's most powerful supercomputers, they hope artificial intelligence can speed up and expand weather and climate forecasting, and spot patterns scientists could not see before.

Researchers at Switzerland's Federal Institute of Technology Zurich (ETH Zurich) said they copied around 100 petabytes of publicly available NASA data, roughly six billion files, onto servers next to Alps, one of the world's most powerful supercomputers, housed at the Swiss National Supercomputing Centre (CSCS) in Lugano. The transfer took about a year, said Reto Knutti, a climate physics professor who heads ETH's Center for Climate Systems Modeling (C2SM); he compared the data volume to roughly 20 million feature-length films, or about a million times the storage of a typical computer.

"This is a huge scientific opportunity," said Thomas Schulthess, an ETH computational physics professor and head of the CSCS, adding that having NASA's climate data plugged directly into the machine is "really enabling scientists to do things we would not even have thought of before." New AI-generated statistical models, Knutti said, are far faster than the traditional approach of solving mathematical equations to simulate the ocean and atmosphere: instead of hours, a statistical model based on pattern recognition can run a global weather forecast for multiple days -- the whole globe -- in a minute or so.

The models are expensive to train but cheap to run, Knutti said, so researchers can run far more iterations to check whether a specific weather pattern is forming, enabling early-warning systems "to save lives." Not every disaster is foreseeable, but he said geological hazards such as landslides and glacier collapses can often be spotted early in satellite data if there is capacity to recognise the pattern.

He pointed to the Swiss village of Blatten, wiped out by a glacier collapse in May last year: the risk was visible in satellite data more than a year beforehand, and close monitoring let Swiss authorities evacuate the village a week before the collapse, avoiding mass casualties. Closer monitoring, Knutti said, could similarly have flagged the area as a hotspot before the devastating Aug. 26 glacial collapse on the Nepal-China border, where satellite image analysis reportedly showed some warning signs beforehand.