Nested-EAGLE Application v1.0.0 Release
Release date: 04/22/2026
The Earth Prediction Innovation Center (EPIC) and the NOAA Artificial Intelligence for Numerical Weather Prediction (AI4NWP) Working Group are proud to announce the public release of the nested-Experimental AI Global and Limited-area Ensemble forecast system (nested-EAGLE) v1.0.0, which complements the Environmental Modeling Center’s (EMC) GraphCast-based release of global-EAGLE-solo and global-EAGLE ensemble forecast systems (which are now operational). This initial release of nested-EAGLE v1.0.0 provides a machine-learning (ML)-based forecast workflow integrated with the European Centre for Medium-Range Weather Forecasts (ECMWF) and partner agencies’ Anemoi ML framework. This release covers the end-to-end workflow, from preprocessing training data through forecast generation, verification, and visualization. The nested-EAGLE Application is designed for training an ML model on both a global domain (a choice of ~25 km or 100 km resolution) and a nested-CONUS domain (a choice of ~6 km or 15 km resolution) to produce atmosphere-only forecasts for a range of variables. Figure 1 shows the global and CONUS domain decomposition. The nested-EAGLE v1.0.0 release documentation guides users through a simplified ~100 km global domain and ~15 km nested domain configuration. This initial release targets NOAA employees and affiliates and is supported on NOAA Research and Development High Performance Computing System (RDHPCS) machine Ursa, which provides the GPU resources required for a timely execution of this current iteration of the EAGLE ML model training workflow. While the current release is supported on Ursa, advanced users may replicate nested-EAGLE on other HPC systems. Subsequent releases will broaden the support of the nested-EAGLE application on the native Microsoft Azure cloud and other suitable HPC systems.
Application Description:
The nested-EAGLE workflow includes environment setup, data preprocessing, ML training, inference (forecast generation), postprocessing, verification, and forecast visualization. Preprocessing utilizes NOAA’s Physical Sciences Laboratory (PSL)’s ufs2arco utility which is designed to convert non-NOAA and NOAA-based datasets to Anemoi training-ready data structures(e.g., Unified Forecast System (UFS) to Analysis Ready, Cloud Optimized (ARCO) format). ECMWF and partner agencies’ Anemoi-core performs model training and generates the checkpoint (ML weight) files required for Anemoi-inference, to generate forecast outputs. A postprocessing step then converts the Anemoi forecast output into a format suitable for verification. Verification is performed using NOAA Global Systems Laboratory (GSL)’s wxvx package, which, via MET, computes dozens of statistics and produces spot-check plots of the root-mean squared error (RMSE) and mean error (ME) for standard 2-D and 3-D variables across varying forecast lead times. The eagle-tools utility package, a light weight Python-based toolkit, postprocesses outputs for verification and its visualization capabilities guide users in plotting a variety of model performance metrics. Repository testing includes both optimized end-to-end workflow tests and static code analysis.
User Support Information:
Nested-EAGLE’s overview and user’s guide are available through the nested-EAGLE repository’s README and the ReadtheDocs page, where users can find detailed information about nested-EAGLE including setup, execution, and testing instructions. Data files required to run the nested-EAGLE Application, including those used for verification, are publicly available in AWS S3 buckets. These include the Global Forecast System (GFS) analysis dataset (noaa-gfs-bdp-pds), High Resolution Rapid Refresh (HRRR) analysis dataset (noaa-hrrr-bdp-pds), and the PrepBufr observation files (noaa-reanalyses-pds). Users seeking assistance can submit questions or feedback through the nested-EAGLE’s GitHub Discussions or file an issue through GitHub Issues. Additionally, users can reach out directly through support.epic@noaa.gov.
Contributors:
The inaugural nested-EAGLE Application release is the result of a collaboration among the Earth Prediction Innovation Center (EPIC), Physical Science Laboratory (PSL), Global Systems Laboratory (GSL), and the Environmental Modelling Center (EMC). The code is hosted on GitHub and it incorporates software developed by PSL, GSL, EMC, as well as ECMWF and partner agencies. Publicly available data is provided through an AWS S3 bucket established as part of the NOAA Open Data Dissemination (NODD) Program and NCAR Research Data Archive. Future research will expand the developmental and operational capabilities of nested-EAGLE to support both on-prem and cloud-based computing environments and additional model applications.
Acknowledgements:
This release was funded by NOAA programs including the Atmospheric River Forecast Improvement Project, Weather Program Office’s Earth Prediction Innovation Center Program, NWS Office of Science and Technology Integration, Software Engineering for Novel Architectures project and Joint Technology Transfer Initiative program. The team acknowledges the NOAA RDHPCS program for providing dedicated computing resources for this release.
