Nvidia launches 'Earth-2' AI models for weather forecasting, to be '1000 times' faster than traditional means

In a significant leap for meteorological science, Nvidia has unveiled a new suite of open-source artificial intelligence models designed to make weather forecasting faster, cheaper, and more accurate.

Announced on January 26 at the American Meteorological Society’s annual meeting in Houston, the ‘Earth-2’ family of models aims to replace traditional, resource-heavy supercomputer simulations with agile AI-driven alternatives.

Historically, weather prediction has relied on massive supercomputers running complex physics equations, a process that is both time-consuming and prohibitively expensive for many nations.

Nvidia’s new toolkit, however, leverages graphics processing units (GPUs) to accelerate every stage of the forecasting pipeline. By shifting to an AI-powered approach, the company claims that computational time can be reduced from hours to mere seconds.

The release includes several specialised tools. Earth-2 Medium Range, powered by the new ‘Atlas’ architecture, provides 15-day global forecasts across 70 different variables, including wind, temperature, and humidity.

For more immediate threats, Earth-2 Nowcasting uses generative AI to predict local storm developments at a kilometre-scale resolution within a six-hour window. This is particularly vital for emergency responders and the insurance industry, who require rapid data to mitigate the impact of extreme weather events.

Mike Pritchard, Nvidia’s director of climate simulation research and a professor of earth system sciences at the University of California, Irvine, highlighted the efficiency of the new system, noting that once trained, these AI models are "1,000 times faster" than conventional methods. This allows organisations to run massive "ensembles" (thousands of simultaneous simulations) to identify outlier weather risks that were previously too costly to track.

The models have been made open-source via GitHub and Hugging Face, promoting ‘sovereign’ forecasting where countries can manage their own climate data on local hardware.

Early adopters, including the Israel Meteorological Service and various energy firms, have already reported up to a 90% reduction in computing costs.