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

As a significant winter storm strikes much of the United States, Nvidia's newly launched Earth-2 weather forecasting models have emerged at a crucial moment. The timing of the release raises questions about whether the company anticipated the storm's intensity, particularly given the mixed predictions prior to its arrival. Nvidia's Earth-2 models aim to enhance the speed and accuracy of weather forecasting, with the Earth-2 Medium Range model reportedly outperforming Google's GenCast on over 70 forecasting variables. Google’s GenCast, launched in December 2024, had already set a new standard for accuracy among weather models that could predict conditions up to 15 days in advance. During the American Meteorological Society meeting in Houston, Mike Pritchard, Nvidia's director of climate simulation, emphasized a shift towards simpler and more scalable AI architectures. He stated, "Philosophically, scientifically, it’s a return to simplicity." This approach marks a departure from the complex, niche AI systems previously utilized in weather forecasting. Traditionally, weather forecasts have relied heavily on simulations based on real-world physics. AI models, however, represent a modern addition to this field. The new Earth-2 Medium Range model is built on Nvidia's Atlas architecture, further details of which were expected to be unveiled during the event. In addition to the Medium Range model, the Earth-2 suite includes a Nowcasting model, designed for short-term predictions (zero to six hours ahead), which is particularly useful for meteorologists tracking storm impacts. Pritchard highlighted that this model is adaptable worldwide, thanks to its training on global geostationary satellite observations. The Global Data Assimilation model also plays a vital role, utilizing data from weather stations and balloons to create continuous snapshots of weather conditions globally. These snapshots serve as the foundation for subsequent weather predictions. Pritchard noted that this new model dramatically reduces the computing power needed, completing tasks that previously took hours on supercomputers in just minutes on GPUs. This suite of models adds to Nvidia’s existing offerings, including CorrDiff, which generates rapid, high-resolution forecasts, and FourCastNet3, focusing on individual weather factors like temperature and humidity. Pritchard believes that these advancements will democratize access to powerful forecasting tools, which have typically been reserved for affluent nations and corporations with the resources to invest in supercomputing. Meteorologists in countries like Israel and Taiwan are already utilizing the Earth-2 CorrDiff model, while The Weather Company and Total Energies are exploring the Nowcasting model. Pritchard also touched on the significance of weather forecasting in matters of national security, asserting that for many nations, having control over weather data is crucial. As Nvidia continues to innovate in this space, the potential for enhanced weather forecasting is set to transform how governments and organizations prepare for and respond to severe weather events.