Nvidia’s AI Predicted This Storm Weeks in Advance
▼ Summary
– Nvidia announced new AI weather forecasting models called Earth-2, aiming to make predictions faster and more accurate than traditional methods.
– One model, Earth-2 Medium Range, claims to outperform Google DeepMind’s GenCast on over 70 variables for forecasts up to 15 days out.
– The suite includes a Nowcasting model for short-term (0-6 hour) storm predictions and a Global Data Assimilation model that creates global weather snapshots using GPU efficiency.
– These AI models are designed to be more accessible, moving weather forecasting away from costly supercomputers and making powerful tools available to more countries and companies.
– The tools address weather as a national security issue, offering sovereign forecasting capabilities and are already being used or evaluated in places like Israel, Taiwan, and by major corporations.
As a major winter storm impacts large portions of the United States, traditional forecasts struggled with significant discrepancies in snowfall predictions. Nvidia’s newly unveiled Earth-2 artificial intelligence weather forecasting platform claims to offer a solution, promising not only greater speed but also superior accuracy. The company announced these models at the American Meteorological Society meeting, suggesting their technology could fundamentally change how we predict weather.
One model within the suite, called Earth-2 Medium Range, reportedly outperforms Google DeepMind’s advanced GenCast model across more than 70 different meteorological variables. Google’s GenCast, released just last December, was already a major leap forward, surpassing the capabilities of conventional 15-day forecast models. Nvidia’s approach represents a strategic shift in design philosophy. According to Mike Pritchard, Nvidia’s director of climate simulation, the move is toward “simple, scalable, transformer architectures” and away from complex, specialized AI systems. While standard forecasting relies heavily on simulating physical laws, AI models like Earth-2 are introducing a powerful new paradigm.
The Earth-2 platform is comprehensive, featuring three core new models alongside two existing ones. The Nowcasting model generates ultra-short-term forecasts for the next zero to six hours, which is critical for predicting the immediate impacts of severe storms. Its design allows for global application because it is trained directly on satellite data rather than regional physics models. This capability could be especially valuable for smaller nations and local governments needing to understand how hazardous weather will affect their specific areas.
Another component, the Global Data Assimilation model, creates continuous global weather snapshots using data from stations and balloons. Historically, generating these foundational snapshots consumed about half of the total supercomputing power dedicated to traditional weather forecasting, a process taking hours. Pritchard notes that this new model can accomplish the same task in mere minutes using GPUs, dramatically reducing the computational barrier.
These new tools join the existing CorrDiff and FourCastNet3 models. The overarching goal is to democratize access to high-powered forecasting. Traditionally, only wealthy nations and large corporations could afford the immense supercomputing resources required for the most accurate weather models. Nvidia’s system aims to provide the essential building blocks for a wider range of users, including national weather services, energy firms, and financial institutions. Some tools are already in operational use or under evaluation by meteorologists in Israel and Taiwan, as well as by companies like The Weather Company and Total Energies.
Pritchard emphasized that while some organizations may prefer a subscription-based forecasting service, for many countries, controlling their own weather prediction is a matter of sovereignty and national security. By making these advanced AI models more accessible and efficient, Nvidia is positioning its technology as a critical tool for anyone whose decisions depend on understanding the atmosphere.
(Source: TechCrunch)
