Amidst a backdrop of recent winter storms causing wildly varying weather forecasts across the U.S., Nvidia has unveiled its new Earth-2 AI weather forecasting models, which promise to deliver significantly faster and more accurate predictions. The company chose the American Meteorological Society meeting in Houston to announce these groundbreaking tools, hinting at a potential knowledge of upcoming weather patterns that underscores the models' claimed precision.

Nvidia's Earth-2 Medium Range model, in particular, boasts superior accuracy, outperforming Google DeepMind’s AI weather model, GenCast, on over 70 variables. GenCast, itself a significant leap forward when released in December 2024, was already noted for being considerably more accurate than traditional weather models capable of forecasting up to 15 days out. Mike Pritchard, director of climate simulation at Nvidia, highlighted a philosophical shift towards simplicity, emphasizing a move away from niche AI architectures to scalable, transformer-based designs, which form the core of these new systems.

Traditionally, weather forecasting has relied heavily on physics simulations of real-world observations, with AI models being a more recent integration. The Earth-2 Medium Range model is built upon a new Nvidia architecture dubbed Atlas, with further details expected to be released by the company. This model is part of a comprehensive Earth-2 suite that also includes Nowcasting and Global Data Assimilation models.

The Nowcasting model is designed for short-term predictions, covering zero to six hours into the future, and aims to assist meteorologists in forecasting the immediate impacts of storms and other hazardous weather events. Uniquely, Nowcasting is trained directly on globally available geostationary satellite observations rather than region-specific physics model outputs, making its approach adaptable to any location on the planet with adequate satellite coverage. This global applicability is expected to greatly aid governments of various sizes in understanding severe weather systems' potential effects on their territories.

The Global Data Assimilation model addresses another critical aspect of forecasting by utilizing data from diverse sources like weather stations and balloons to create continuous snapshots of global weather conditions at thousands of locations. These snapshots serve as crucial starting points for weather models to generate their predictions. Historically, creating these detailed snapshots demanded immense computing power, consuming roughly 50% of the total supercomputing loads for traditional weather forecasting. Nvidia's new model can accomplish this task in minutes using GPUs, a dramatic improvement over the hours required by traditional supercomputers.

These new Earth-2 models integrate with two existing Nvidia forecasting tools: CorrDiff, which generates speedy, high-resolution predictions from coarse-grained forecasts, and FourCastNet3, designed to model individual weather variables such as temperature, wind, and humidity. This comprehensive suite aims to democratize access to powerful weather forecasting capabilities, which have historically been exclusive to wealthier nations and large corporations due to the high costs associated with supercomputer time.

Pritchard stressed that these tools provide fundamental building blocks for various entities within the ecosystem, including national meteorological services, financial service firms, and energy companies, enabling them to build and refine their forecasting models. He also highlighted the geopolitical importance of weather, stating that "weather is a national security issue, and sovereignty and weather are inseparable." Some components of the Earth-2 suite are already in use, with meteorologists in Israel and Taiwan deploying Earth-2 CorrDiff, and The Weather Company alongside Total Energies evaluating the Nowcasting model.