NVIDIA Unveils Open Earth-2 Models for AI Weather Forecasting

NVIDIA has launched its Earth-2 family of open-source AI models for weather and climate forecasting at the American Meteorological Society meeting in Houston, Texas. The company said the tools aim to widen access to weather prediction for scientists, firms and governments worldwide.

The Earth-2 platform offers open models, libraries and frameworks that support forecasting from data processing to 15-day global or local storm predictions. NVIDIA said the models reduce compute time and costs compared with physics-based forecasting systems.

https://youtu.be/qo78lSBYi-U?si=zujpStz-dfWUFHwC

NVIDIA said Earth-2 includes new Medium Range, Nowcasting and Global Data Assimilation models, each built on a new architecture. The Medium Range model forecasts up to 15 days across more than 70 weather variables.

The Nowcasting model uses generative AI trained on satellite and radar data to predict storms up to six hours ahead. NVIDIA said it can simulate storm dynamics directly and deliver results within minutes.

“AI-powered weather forecasting saves significant computational time and costs,” the company said. It added that this allows more countries and firms to run tailored forecasting systems.

Example of an Earth-2 Nowcasting forecast demonstrating realistic cloud and rainfall systems, showing how storms develop and organise.

The Global Data Assimilation model generates initial conditions for weather prediction seconds on GPUs. NVIDIA said this replaces hours of processing on supercomputers when paired with other Earth-2 models.

Developers and agencies are already testing the platform. Julian Green, cofounder and chief executive of Brightband, said open models speed innovation and comparison across the sector. 

The Israel Meteorological Service said Earth-2 cut compute time by 90% at high resolution. “NVIDIA Earth-2 models give us a 90% reduction in compute time at 2.5-kilometre resolution compared with running a classic numerical weather prediction model without AI on a CPU cluster,” said Amir Givati, director of the Israel Meteorological Service. 

He further described that, after a recent rainstorm, their AI model trained with CorrDiff was the best of their operational models for a six-hour verification of accumulated precipitation.

These new models add to existing open tools in the Earth-2 stack, including CorrDiff and FourCastNet3. NVIDIA said CorrDiff downscales continental forecasts into high-resolution regional outputs, enabling local predictions up to 500 times faster than traditional methods.

FourCastNet3 delivers forecasts for variables such as wind, temperature and humidity, producing results up to 60 times faster than conventional ensemble approaches. The giant also added that Earth-2 integrates open models from the European Centre for Medium-Range Weather Forecasts, Microsoft, and Google and supports training through its open-source PhysicsNeMo framework.