NVIDIA Launches Earth-2 Family of Open Models — the World’s First Fully Open, Accelerated Set of Models and Tools for AI Weather
- •NVIDIA unveils Earth-2, an open software stack featuring Atlas, StormScope, and HealDA architectures for global weather forecasting.
- •Earth-2 Nowcasting outperforms traditional physics models in predicting local storms and rainfall systems using generative AI.
- •Global Data Assimilation (HealDA) generates atmospheric starting conditions in seconds using GPUs instead of hours on supercomputers.
NVIDIA has fundamentally shifted the landscape of meteorology with the launch of Earth-2, a comprehensive suite of open-source models designed to accelerate climate and weather prediction. This release marks the first fully open, GPU-accelerated software stack for the field, offering tools that range from global 15-day forecasts to hyper-local storm tracking. By moving away from traditional physics-based supercomputing, NVIDIA enables organizations to run sophisticated simulations with drastically reduced computational costs.
The family includes three primary new architectures: Atlas for medium-range forecasting, StormScope for rapid nowcasting, and HealDA for global data assimilation. Atlas provides high-accuracy predictions for over 70 weather variables, while StormScope leverages a generative model to simulate storm dynamics directly from satellite imagery at kilometer-scale resolution. This level of granularity was previously impossible to achieve in minutes, representing a significant leap for emergency response and disaster preparedness.
Complementing these is the HealDA model, which calculates the current state of the atmosphere in seconds. When integrated with existing frameworks for high-resolution downscaling, they form an end-to-end AI pipeline that rivals conventional meteorological methods. This open ecosystem empowers enterprises to fine-tune these foundation model assets on their own infrastructure, democratizing access to high-fidelity weather intelligence.