Nvidia announced on Monday the release of three open-source artificial intelligence models created for weather forecasting, presenting the software at the American Meteorological Society's annual meeting in Houston. The models form part of the company's wider effort to supply open-source tools that leverage its chips across multiple domains, including conversational agents and autonomous driving.
The suite, collectively referred to as "Earth-2," is intended to serve as an AI-based substitute for conventional numerical weather simulations. Nvidia says these AI models can equal or surpass the accuracy of traditional forecasting approaches while offering significantly reduced run times and lower operating costs after the initial training phase.
One consequence Nvidia highlights is the potential to remove a key constraint in producing detailed assessments of extreme weather. Current practice in many forecasting operations relies on "ensembles" - sets of individual model runs started from slightly different initial conditions - to capture the range of possible outcomes. Detecting rare or outlier events typically requires many ensemble members, and executing each simulation with full numerical detail is slow and costly.
Mike Pritchard, identified as director of climate simulation research at Nvidia and a professor of earth system sciences at the University of California, Irvine, pointed to the insurance sector as an industry likely to benefit from the capability to run much larger ensembles. He noted that after models are trained, AI inference is orders of magnitude faster, enabling insurers and similar users to run ensembles with thousands of members where previously the computational cost was prohibitive.
The three Earth-2 models cover distinct forecasting roles: one model targets forecasts extending up to 15 days; a second focuses on short-term forecasts of up to six hours for severe storm development over the United States; and a third is designed to ingest multiple streams of sensor data to produce improved initial conditions for downstream forecasting systems.
Nvidia positions these models as part of a broader strategy to provide open-source, chip-optimized software across varied applications. The company emphasizes speed and scalability as the primary advantages once the AI systems have been trained, and highlights specific practical use cases such as the ability to run very large ensemble experiments that were previously impractical due to computational expense.