As a major winter storm impacts much of the United States, the timing of Nvidia’s announcement for its new Earth-2 weather forecasting models could not be more relevant. In the days leading up to the storm, snowfall predictions for some regions varied dramatically, highlighting the challenges of traditional forecasting. Perhaps Nvidia, with its claims of superior accuracy, anticipated this very need.
The new suite of AI models promises to make weather forecasting both faster and more precise. Nvidia states that one model in particular, Earth-2 Medium Range, outperforms Google DeepMind’s AI weather model, GenCast, on more than 70 different variables. Google’s GenCast, released in December 2024, was itself a significant leap in accuracy over existing models capable of generating 15-day forecasts.
Nvidia introduced these tools at the American Meteorological Society meeting in Houston. Mike Pritchard, director of climate simulation at Nvidia, described the shift as a return to simplicity, moving away from specialized AI architectures toward scalable transformer models.
Traditional weather forecasting relies heavily on physics-based simulations of the real world, with AI models being a more recent innovation. The Earth-2 Medium Range model is built on a new Nvidia architecture called Atlas. The full Earth-2 suite also includes a Nowcasting model and a Global Data Assimilation model.
The Nowcasting model generates short-term predictions from zero to six hours into the future. It is designed to help meteorologists forecast the immediate impacts of storms and hazardous weather. Because it is trained directly on globally available satellite observations rather than region-specific physics models, its approach can be adapted anywhere with good satellite coverage. This capability could help state governments and smaller countries better understand how severe weather might affect their territories.
The Global Data Assimilation model uses data from sources like weather stations and balloons to produce continuous snapshots of current conditions at thousands of global locations. These snapshots serve as the starting point for weather models to begin their predictions. Traditionally, creating these snapshots consumed roughly half of the total supercomputing power used in forecasting. Nvidia claims this new model can complete that work in minutes on GPUs instead of the hours required on traditional supercomputers.
These three new models join two existing ones in the Earth-2 ecosystem: CorrDiff, which uses coarse forecasts to generate speedy, high-resolution predictions, and FourCastNet3, which models individual weather variables like temperature and wind.
According to Pritchard, these tools aim to democratize access to powerful weather forecasting, which has historically been limited to wealthier nations and large corporations that can afford expensive supercomputer time. He stated that the models provide fundamental building blocks for national meteorological services, financial firms, energy companies, and anyone looking to build and refine forecasts.
Some tools are already in use. Meteorologists in Israel and Taiwan have been using Earth-2 CorrDiff, while The Weather Company and Total Energies are evaluating the Nowcasting model. Pritchard emphasized that while some users may subscribe to a centralized forecasting service, for others like countries, sovereignty is paramount. He noted that weather is a national security issue, and sovereignty and weather are inseparable.