NVIDIA Launches Earth-2 Open AI Weather Forecasting Model and Tools

NVIDIA has introduced Earth-2, a comprehensive suite of AI models and tools designed to make weather forecasting more accessible than ever before for scientists, startups, developers, businesses, and governments worldwide.

The availability of accurate weather predictions is critical, as it impacts a wide range of human activities, from daily routines to major strategic decisions.

For instance, a logistics manager who learns of an impending heavy rainstorm affecting port operations can proactively adjust staffing and routing before the first drop falls. Similarly, government officials responsible for critical infrastructure, upon seeing a forecast of a polar vortex bringing six feet of snow within four days, must treat planning not as optional but as an urgent matter requiring immediate action, resource deployment, and sustained response.

The Earth-2 open technology portfolio includes pre-trained models, frameworks, customization options, and inference libraries. These resources enable users to accelerate every stage of the forecasting process, from processing observational data to generating 15-day global or localized storm predictions.

Historically, production-ready weather forecasting has relied on supercomputer-based models. Generative AI is now disrupting this paradigm by reducing the need for computationally intensive and costly modeling simulations, thereby making advanced forecasting more attainable for both startups and nations.

The toolset features Earth-2 Mid-Range, a 15-day high-precision forecasting model powered by a novel architecture called Atlas for medium-range predictions of over 70 weather variables, including temperature, pressure, wind, and humidity. Earth-2 Nowcasting is driven by a new model named StormScope, which utilizes generative AI to transform national-level forecasts into kilometer-resolution, zero-to-six-hour predictions for local storms and hazardous weather events in just minutes.

Earth-2 Global Data Assimilation, powered by HealDA, generates the initial conditions for weather forecasts. These are snapshots of current atmospheric conditions at thousands of locations globally. This assimilation process can produce these initial conditions in seconds using graphics processing units, as opposed to the hours traditionally required on supercomputers. NVIDIA states that when combined with Earth-2 models, this leads to highly accurate forecasts.

Earth-2 CorrDiff can downscale continental forecasts to high-resolution regional weather for granular local predictions. FourCastNet3 delivers high-accuracy forecasts for various weather variables like wind, temperature, and humidity, outperforming leading traditional models while providing answers up to 60 times faster.

Forecasting Partners

Numerous partners have joined NVIDIA in building with Earth-2. This includes AI weather tool provider Brightband, which runs the Mid-Range model to publish daily global forecasts.

"Brightband was among the first to operate the Earth-2 Mid-Range model. The open-sourcing of the model accelerates innovation and makes it easier for other members of the weather enterprise to benchmark and enhance their work," said Julian Green, Co-founder and CEO.

Power companies, including Eni, GCL Technology Holdings Limited, and S&P Global Commodity Insights, are utilizing Earth-2 tools for predictive weather analytics to optimize their operations.

"NVIDIA Earth-2 represents a significant leap forward in enabling advanced weather intelligence at scale," said Emmanuel Le Borgne, Climate and Weather Forecasting Product Manager at TotalEnergies SE.

NVIDIA announced that the full-featured family of open models and training tools for diagnostics will be licensed for both commercial and non-commercial use. They will be made available on GitHub and Hugging Face upon release.