NVIDIA has announced the launch of three new open-source models as part of the NVIDIA Earth-2 initiative, aimed at simplifying the development of weather forecasting systems across various applications, including data assimilation, forecasting, nowcasting, and downscaling.
The Earth-2 suite comprises a collection of accelerated tools and models that integrate typically separate weather and climate AI functionalities. This open framework allows developers to tailor their simulations to meet specific requirements, utilizing their own data and infrastructure to create personalized weather and climate predictions.
Earth-2 Nowcasting: Precision in Severe Weather Prediction
Now available on Hugging Face, Earth-2 Nowcasting employs a novel model architecture known as StormScope. This model utilizes generative AI to transform country-scale forecasts into kilometer-resolution predictions for local storms and hazardous weather, delivering results in mere minutes. It surpasses traditional physics-based models in short-term precipitation forecasting by directly simulating storm dynamics and leveraging satellite and radar data. The model is trained on globally available geostationary satellite observations (GOES) over the contiguous United States (CONUS), with potential for adaptation to other regions with similar satellite coverage.
Earth-2 Medium Range: Accurate Long-Term Forecasts
Also now on Hugging Face, Earth-2 Medium Range features the Atlas architecture, which enhances accuracy for medium-range forecasts extending up to 15 days. This model covers over 70 weather variables, including temperature, pressure, wind, and humidity. Utilizing a latent diffusion transformer architecture, it predicts incremental atmospheric changes while preserving essential atmospheric structures, thereby minimizing forecasting errors. It has demonstrated superior performance on standard benchmarks compared to leading open models like GenCast.
Upcoming Earth-2 Global Data Assimilation
Coming soon to Hugging Face, the Earth-2 Global Data Assimilation model, powered by the HealDA architecture, will generate initial conditions for weather predictions—offering snapshots of current atmospheric conditions across thousands of locations globally. This model can produce these initial conditions in seconds on GPUs, a significant improvement over traditional supercomputing methods that take hours. When integrated with Earth-2 Medium Range, it promises to deliver highly skillful forecasting predictions through a fully AI-driven pipeline.
These new models join the existing NVIDIA weather and climate models, including FourcastNet3, CorrDiff, cBottle, and DLESym. Developers can leverage the NVIDIA Earth2Studio, an open-source Python ecosystem, to create robust AI weather and climate simulations with ease.
This article was produced by NeonPulse.today using human and AI-assisted editorial processes, based on publicly available information. Content may be edited for clarity and style.
