EAGLE

Today, Project EAGLE includes four components: 

• Global-EAGLE-Solo is a demonstration environment for the global deterministic forecast system that complements the numerical weather prediction (NWP) model GFS. The first version of Global-EAGLE-Solo was implemented into operations as AIGFS in December 2025. The model runs on a 0.25-degree latitude-longitude grid (~28 km) and on 13 pressure levels. The model produces 16-day forecasts four times daily at 00Z, 06Z, 12Z, and 18Z (6-hourly) in GRIdded Binary, Edition 2 (GRIB2) format. Major atmospheric and surface fields are available, including vertical velocity, geopotential height, specific humidity, 2-meter temperature, 10-meter wind components, and precipitation. A minor updated version was implemented into operation on July 27, 2026 with significantly improved tropical cyclone intensity and precipitation forecast skill. Recently, the Global-EAGLE-Solo machine learning (ML) framework transitioned to the European Centre for Medium-Range Weather Forecasts (ECMWF) and partner agencies’ Anemoi framework, an infrastructure facilitating AI/ML weather model development and operational deployment. This recent Global-EAGLE-Solo system is pretrained with 45 years of ECMWF Reanalysis dataset, version5 (ERA5) data and then fine-tuned with four years of NOAA’s operational Global Data Assimilation System (GDAS) analyses on a 0.25-degree resolution through Anemoi. The model generates the same suite of products as the operational AIGFS.
• Global-EAGLE-Ensemble complements the deterministic Global-EAGLE-Solo, which is a demonstration environment for ensemble forecast systems initialized with the operational GEFSv12 initial conditions (ICs). Global-EAGLE-Ensemble is also a complement to the existing GEFS physics-based forecast system. The weights for the Global-Eagle-Ensemble members are generated by fine-tuning the original GraphCast weights from DeepMind© with recent GDAS analyses. The resulting weights are effectively trained on the combination of ERA5, ECMWF High RESolution (HRES), and GDAS analysis. Multiple checkpoints were saved to form 31 global ensemble members. Global-EAGLE-ensemble was implemented into operations in December 2025. Since then, efforts have focused on transitioning the Global-EAGLE-ensemble into the Anemoi framework. Recently, the model was pretrained on 42 years of ERA5 data and subsequently fine-tuned using 4.5 years of GDAS data at both 1.0° and 0.25° spatial resolutions. Ongoing development aims to further optimize the model’s forecast performance.    
• HRRRCast is a data-driven model with the same high resolution grid as the HRRR model, developed by NOAA/OAR’s Global Systems Laboratory (GSL) to provide rapid, convection-allowing weather forecasts using ML. Designed as an ensemble forecast system, HRRRCast generates probabilistic guidance by producing multiple stochastic realizations, enabling improved representation of forecast uncertainty. In collaboration with NWS and the Office of Modeling and Development (OMD), the system is executed in near real time over the contiguous United States (CONUS) domain and supports applications in severe weather prediction, with a particular focus on precipitation, reflectivity, and other high-impact variables. HRRRCast aims to complement and enhance traditional NWP by offering a computationally efficient, scalable approach to regional forecasting. For more information, visit GSL’s HRRRCast webpage.
• Nested-EAGLE complements the Global-EAGLE-Solo and is designed as a real-time demonstration environment that runs the nested-EAGLE application, i.e. inference and verification, in NRT. Nested-EAGLE utilizes the ECMWF and partner agencies’ Anemoi ML forecast system (Lang et al., 2024) in its stretch grid configuration (Nipen et al 2026). Nested-EAGLE extends core Anemoi capabilities through ufs2arco, which processes NOAA datasets to train an ML model on both a global domain (~25 km; GFS) as well as a nested-CONUS domain (~6 km or 15 km; HRRR). The pipeline produces atmosphere-only forecasts for 10-meter wind, 2-meter temperature, 2-meter specific humidity,  3-dimensional geopotential height, u- and v-winds, temperature, and specific humidity. Both retrospective and NRT forecasts of variable length are executed on 13 pressure levels, with daily initializations at 00Z, 06Z, 12Z, and 18Z (6-hourly), in netCDF format. Completing the nested-EAGLE pipeline are eagle-tools for postprocessing and visualization and wxvx for verification. For nested-EAGLE code and user support, visit EPIC’s nested-EAGLE Get Code page.