Technology company NVIDIA announced three new open-source models within its Earth-2 family, aimed at streamlining the development of weather forecasting applications across the full meteorological workflow, including data assimilation, forecasting, nowcasting, downscaling, and related tasks.
Developers can leverage NVIDIA’s open-source software, such as Earth2Studio for building inference pipelines and Physics Nemo for model training, to quickly construct weather and climate simulations.
The Earth-2 suite consolidates a range of accelerated tools and models, uniting capabilities that have traditionally been dispersed across the weather and climate AI landscape. Its fully open framework allows developers to customize and refine simulations according to their own requirements, utilizing proprietary data and infrastructure to produce self-managed, sovereign weather and climate predictions.
With this approach, Earth-2 provides a leading set of open weather and climate models, supported by an accessible ecosystem of software, enabling the creation of tailored, fully controlled forecasting capabilities.
Released on Hugging Face, Earth-2 Nowcasting introduces the model architecture called StormScope that leverages generative AI to produce high-resolution, short-term forecasts. The system generates kilometer-scale predictions of local storms and hazardous weather up to six hours in advance, delivering results in minutes.
By simulating storm dynamics directly and predicting satellite and radar outputs, Earth-2 Nowcasting is capable of outperforming traditional physics-based models in short-term precipitation forecasts. The current version is trained on geostationary satellite observations over the contiguous United States, though the methodology can be adapted for other regions with comparable satellite coverage.
In parallel, Earth-2 Medium Range, which is also available on the platform, is built on a new Atlas model architecture designed for medium-range forecasting up to 15 days ahead. The system predicts over seventy atmospheric variables, including temperature, wind, pressure, and humidity, using a latent diffusion transformer architecture that models incremental changes in the atmosphere to preserve structural coherence and minimize errors. Benchmark testing indicates that Earth-2 Medium Range surpasses other open forecasting models, including GenCast, across commonly used industry variables.
In addition, coming soon to Hugging Face, Earth-2 Global Data Assimilation will feature the HealDA architecture, which generates initial atmospheric conditions at thousands of locations worldwide, including temperature, wind speed, humidity, and air pressure. By producing these snapshots in seconds on GPUs rather than hours on traditional supercomputers, HealDA enables faster and more precise initialization for forecasts. When combined with Earth-2 Medium Range, it forms a fully AI-driven pipeline capable of delivering highly accurate predictions at scale.
These additions expand NVIDIA’s existing suite of open weather and climate models, which includes FourcastNet3, CorrDiff, cBottle, and DLESym. Supporting this ecosystem, NVIDIA Earth2Studio provides an open-source Python framework for quickly building AI weather and climate simulations, offering tools for inference and integration with the new model checkpoints.
The post Nvidia Expands Earth-2 Suite With Three New Open-Source Models For Advanced Weather And Climate Simulations appeared first on Metaverse Post.
Also read: Global Games Show Abu Dhabi 2025 Sets Direction for Gaming and Web3
