NVIDIA is pleased to announce three new open source models as part of the NVIDIA Earth-2 family. This makes it easier than ever to build weather forecasting capabilities across the entire weather stack, including tasks such as data assimilation, forecasting, nowcasting, and downscaling. Additionally, developers can quickly start building weather and climate simulations using NVIDIA open source software: Earth2Studio to create inference pipelines and Physics Nemo to train models.
NVIDIA Earth-2 consists of a set of accelerated tools and models that enable developers to integrate typically disparate weather and climate AI capabilities. Earth-2 is completely open, allowing developers to use their own data and their own infrastructure to build sovereign weather and climate predictions that they completely own and control, customizing and fine-tuning their simulations to suit their specific needs. Earth-2:
A suite of leading open weather and climate models Easy to use thanks to an ecosystem of open source software You can create your own sovereign capabilities
Earth-2 Nowcasting: Kilometer-scale Severe Weather Forecasting
We launched Hugging Face: Earth-2 Nowcasting, powered by a new model architecture called StormScope. Use generative AI to predict national forecasts with kilometer resolution and localized storms and hazardous weather from 0 to 6 hours in just minutes. By directly simulating storm dynamics, Earth-2 Nowcasting can generate the first forecasts that outperform traditional physically-based weather prediction models in short-term precipitation prediction. Use AI to predict directly from satellite and radar data.
This version is trained directly on globally available geostationary satellite observations (GOES) over the continental United States (CONUS). However, this method can be applied to train versions of the model in other regions with similar satellite coverage.
Your browser does not support the video tag.
Research paper: Learning accurate storm size evolution from observations
Earth-2 Medium Range: Highly accurate 15-day global forecast
Hugging Face: Earth-2 Medium Range has been released with a new model architecture called Atlas. This enables high-precision weather forecasting across more than 70 weather variables, including temperature, pressure, wind, and humidity, with medium-range forecasts or up to 15 days in advance. Predict incremental changes in the atmosphere using a latent diffusion transformer architecture to preserve important atmospheric structure and reduce prediction errors. On standard benchmarks, it outperforms leading open models such as GenCast on the most common predictor variables measured in the industry.
Your browser does not support the video tag.
Research paper: Solving the mysteries of data-driven probabilistic medium-range weather forecasting
Earth-2 Global Data Assimilation: An End-to-End AI Pipeline
Coming soon is Hugging Face: Earth-2 Global Data Assimilation, which leverages a new model architecture called HealDA. It produces the initial conditions for weather predictions: a snapshot of the current atmosphere including temperature, wind speed, humidity, and pressure at thousands of locations around the world. Earth-2 Global Data Assimilation can generate initial conditions in seconds on GPUs instead of hours on supercomputers. When combined with Earth-2 Medium Range, an open, complete AI pipeline produces the most sophisticated predictions.
Your browser does not support the video tag.
Research paper: HealDA: Highlighting the importance of initial error in end-to-end AI weather forecasting
These models join established open NVIDIA weather and climate models such as FourcastNet3, CorrDiff, cBottle, and DLESym.
NVIDIA Earth2Studio is an open source Python ecosystem for rapidly creating powerful AI weather and climate simulations. This provides all the inference tools needed to start a new model checkpoint for Hugging Face. It’s as simple as:
introductory video
Company Blog: NVIDIA Introduces Earth-2 Family of Open Models — World’s First Completely Open Set of Models and Tools for AI Weather
Announcement video: NVIDIA Earth-2: The future of AI weather forecasting is unlocked
Webpage: Earth-2
Earth-2 nowcasting hug face package
Research paper: Learning accurate storm size evolution from observations
Earth-2 mid-range hug face package
Research paper: Solving the mysteries of data-driven probabilistic medium-range weather forecasting
Research paper: HealDA: Highlighting the importance of initial error in end-to-end AI weather forecasting
