How to Build Regional Weather and Climate Models Using NVIDIA Earth-2

Deploying Open Weather AI Models

Developers and researchers can leverage the NVIDIA Earth-2 family of open AI models and tools to customize global weather predictions, regional forecasts, and environmental research with their own data. Operating on NVIDIA computing infrastructure, these technologies enable organizations to transition from broad global forecasting to localized insights essential for energy producers, insurers, and transport operators.

Combining Models for Regional Workflows

Practitioners can construct specialized workflows by combining different Earth-2 architectures to match specific geographic and operational needs:

• Short-Term Forecasting: Utilize StormScope and its regional European variant, StormScope EU, to predict evolving cloud conditions using EUMETSAT satellite observations at 10-minute intervals.
• High-Resolution Detail: Apply the CorrDiff model to add local detail to coarser weather data.
• Custom Environmental Modeling: Train model architectures for new applications, such as air-quality modeling, using the open-source NVIDIA PhysicsNeMo framework.

Managing Real-Time Results and Data Pipelines

Teams can manage and serve real-time results without downloading entire datasets by integrating forecasting workflows with data platforms. For example, developers can use Earth2Studio to run forecasts—such as those showcased in live pipelines utilizing EUMETSAT observations—and serve predictions as map tiles for interactive exploration.

Integrating Proprietary Data and Domain Expertise

The PhysicsNeMo training and customization tools allow technical teams to incorporate proprietary data and domain expertise directly into model development. Organizations have adapted these capabilities for diverse use cases:

• Study regional cyclone risk beyond historical records by running simulations across thousands of seasons using FourCastNet 3.
• Enhance weather intelligence platforms by combining models with proprietary observations, lightning data, and software-as-a-service offerings.
• Scale environmental modeling from national supercomputers down to local desktop systems, as demonstrated by researchers running adapted CorrDiff models on NVIDIA DGX Spark hardware.
• Evaluate solar-radiation nowcasts at subhourly intervals for photovoltaic sites.

Source

Official announcement or documentation