Nvidia’s new AI weather models likely saw this storm weeks ago

In the lead-up to the winter storm currently battering much of the United States, the weather forecast for some areas was all over the map, with snowfall forecasts varying widely.

Nvidia could not have timed the release of new Earth-2 models to better predict weather. Or, given how accurate the company’s claims are that the new models are, perhaps it knew something we didn’t?

New AI models promise to make weather forecasting faster and more accurate. Nvidia claims that one model in particular, the Earth-2 Medium Range, outperforms Google DeepMind’s AI-powered weather model, GenCast, on more than 70 variables. gencast, Which was launched by Google in December 2024was itself more accurate than current weather models that were able to generate forecasts for up to 15 days.

Nvidia announced the new tools Monday at the American Meteorological Society meeting in Houston.

“Philosophically and scientifically, it’s a return to simplicity,” Mike Pritchard, Nvidia’s director of climate simulation, told reporters in a phone call before the meeting. “We are moving away from specialized, hand-designed AI architectures and toward a future of simple, scalable transformer architectures.”

Traditionally, most weather forecasts are based on simulations of physics as observed in the real world. AI models are a relatively recent addition. The mid-range Earth-2 model is based on Nvidia’s new architecture called Atlas, which the company said it would release more details about on Monday.

Along with the mid-range, Nvidia’s Earth-2 suite includes a nowcast model and a global data assimilation model.

Nowcasting produces short-range forecasts from zero to six hours into the future and is intended to help meteorologists predict the impacts of storms and other hazardous weather.

“Because this model is trained directly on globally available geostationary satellite observations, rather than region-specific physics model outputs, the nowcasting approach can be adapted anywhere on the planet with good satellite coverage,” Pritchard said. This should help state governments and small nations understand the extent to which extreme weather systems affect their territories.

The global data assimilation model uses data from sources such as weather stations and balloons to produce continuous snapshots of weather conditions at thousands of locations around the world. Those snapshots are then used as starting points for weather models to make their predictions.

Traditionally, these snapshots required massive amounts of computing power before the forecasting work could begin. “It consumes approximately 50% of the total supercomputing load of traditional weather forecasting,” Pritchard said. “This model can do it in minutes on GPUs instead of hours on supercomputers.”

The three new models join two existing models: CorrDiff, which uses coarse forecasts to generate fast, high-accuracy forecasts, and FourCastNet3, which models individual weather variables such as temperature, wind and humidity.

Pritchard said the new models should give more users access to powerful weather forecasting tools, which have historically been the domain of wealthier countries and larger companies, which have the money to pay for expensive supercomputer time.

“This provides the building blocks that are used by everyone in the ecosystem — the National Weather Service, financial services companies, energy companies — and anyone who wants to build and improve weather forecasting models,” Pritchard said. Some tools are already in use. Meteorologists in Israel and Taiwan are using Earth-2 CorrDiff, for example, while The Weather Company and Total Energies are evaluating the nowcast, Nvidia said.

“For some users, it makes sense to subscribe to a centralized, enterprise weather forecasting system,” Pritchard said. “But for others, such as countries, sovereignty is important.” “Weather is a national security issue, and sovereignty and weather are inseparable.”