Nested-EAGLE: A Data Driven Global Weather Model with High Resolution over the Contiguous United States

Timothy Smith

NOAA Physical Sciences Laboratory

Tuesday, Sep 23, 2025, 2:00 pm MT
DSRC Room GC402

JOIN THE WEBINAR:

Abstract

NOAA’s suite of Numerical Weather Prediction systems are each uniquely targeted at specific applications that span a wide range of spatiotemporal scales. Each forecast system is uniquely tailored to its application, with tuned physical parameterizations and model resolutions relevant to the situation. While some separation between different modeling systems is practical, computational resource constraints generally prohibit forecast systems from using high resolution to forecast for long lead times.

In this talk, we present a prototype of such an “all in one” approach based on advancements in Machine Learning Weather Prediction (MLWP), with a global model that has high resolution over the area of interest, the Contiguous United States. Building on work from Met Norway, we develop a deep neural network forecast model that is trained on archived NOAA Global Forecast System (GFS) and High Resolution Rapid Refresh (HRRR) data. We present an evaluation of the model’s forecast skill compared to archived HRRR and GFS forecasts. In terms of Mean Squared Error (MSE), the neural network has lower MSE than HRRR over 2 days, and is competitive with GFS at 10 days. Thus, the model shows promise at providing the best of both worlds, capturing both the short and medium range timescales accurately. However, we also show a quantitative analysis of precipitation prediction skill which exposes shortcomings in terms of its ability to represent precipitation patterns. The analysis motivates future work, including the use of either a stochastic or feature-based loss in order to improve the representation of precipitation extremes.

From a broader perspective, the development of this model marks a step forward in terms of our Machine Learning model development capabilities. The model development pipeline employs ECMWF’s Anemoi framework, along with the data processing tool ufs2arco, which was developed in-house. Now, we are at a point where we have access to the many capabilities present in the Anemoi ecosystem, which paves the way for rapid development.

Bio: Timothy Smith is a Research Scientist working in the Modeling and Data Assimilation Division of NOAA's Physical Sciences Laboratory in Boulder, Colorado. He is interested in advancing coupled data assimilation techniques for the next generation of weather forecasting systems. His main focus is in developing Machine Learning methods that can enable strongly coupled data assimilation, so that observations of the atmosphere can impact estimates of the ocean state, and vice versa, directly within the data assimilation framework. More about Timothy