Nvidia releases three open-source AI models for weather forecasting
Nvidia has announced three new AI models for weather forecasts. They are said to be faster and more energy-efficient than conventional supercomputers.
Nvidia has presented three open-source AI models for weather forecasting. They are said to be significantly superior to traditional supercomputers.
(Image: Nvidia)
With the accuracy and speed of AI weather models, Google had already amazed the scientific community. Google DeepMind demonstrated on the example of tropical storm forecasts where meteorology might be heading. Now, at the American Meteorological Society Conference in Houston, Nvidia has also presented three new open-source AI models for weather forecasting. As the company announced on its Earth-2 platform, the models are said to operate significantly faster and more energy-efficiently than conventional supercomputer-based systems.
The three new models cover different timeframes: Earth-2 Medium Range forecasts weather events up to 15 days in advance, analyzing over 70 weather variables such as temperature, air pressure, wind, and humidity. The model uses the so-called Atlas architecture and, according to Nvidia, is said to surpass Google's DeepMind GenCast in the number of variables. Earth-2 Nowcasting covers the short-term range up to six hours and uses generative AI to predict satellite and radar data. As the first AI model, it is said to outperform traditional physics-based systems in simulating storm dynamics, according to Nvidia.
Calculations in seconds instead of hours
The third model, Earth-2 Global Data Assimilation, is said to generate precise initial atmospheric conditions on GPUs in seconds – a task that takes hours on supercomputers. The model is based on the HealDA architecture and will be released later this year. The first two models are available via Nvidia's Earth2Studio, Hugging Face, and GitHub; Earth-2 Global Data Assimilation will follow later.
Videos by heise
In addition to the new models, Nvidia presented the generative AI model cBottle (Climate in a Bottle) at GTC Washington. Announced in June 2025, the model generates high-resolution climate simulations on a kilometer scale and is said to be thousands of times faster than conventional supercomputer-based simulations, according to Nvidia. cBottle was trained on just four weeks of high-resolution climate simulation data and can supplement missing climate data, correct distorted climate models, or improve low-resolution data.
Another presented method is "score-based data assimilation." In this process, an AI model transforms limited weather observation data into high-resolution visualizations. The technique fills gaps between data points using a generative model. According to Alex Philp from MITRE Corporation, which uses the method, the model marks a "turning point for weather forecasting." Traditional data assimilation techniques require millions of CPU hours per year on supercomputers to create an extremely high-resolution weather map of the USA. In contrast, the approach with Nvidia's PhysicsNeMo framework requires only one hour on a single GPU. The US National Weather Service is testing the new Earth-2 models to improve its operational workflows.
International weather services rely on Earth-2
The Earth-2 models are also finding international application. The Taiwan Central Weather Administration is among the first users of the new Earth-2 cloud APIs and creates high-resolution simulations for more accurate typhoon forecasting. The Israeli Meteorological Service is testing Earth-2 CorrDiff. Energy companies such as TotalEnergies use Earth-2 Nowcasting to improve short-term risk assessment and decision-making in energy systems. Insurance companies such as AXA and JBA Risk Management use the tools to simulate extreme weather events for risk analysis.
In terms of energy efficiency, Nvidia promises significant improvements: the CorrDiff model is said to be 500 times faster and 10,000 times more energy-efficient than CPU-based systems.
While Nvidia aims to score with its models, the competition is not idle. In December 2024, Google DeepMind presented GenCast, an AI weather model that creates 15-day ensemble forecasts in eight minutes on a single Google Cloud TPU v5, surpassing the European Centre for Medium-Range Weather Forecasts. The German Weather Service is also working on its own AI models and opened an AI center in Offenbach in August 2025.
(mki)
