Preliminary Testing of a Machine Learning Limited-Area Model (ML-LAM) over Continental North American Domain

One of the major requirements from the 3D Real Time Mesoscale Analysis (RTMA) project is timely delivery of analysis, performed at 2.5 km in 15 min time intervals, to stakeholders for nowcasting and other applications. With the goal to speed up the efficiency in producing the analysis, we started development of a new paradigm in which, instead of first running a numerical model, presently Rapid Refresh Forecast System (RRFS) at 3 km, and then downscaling its results to 2.5 km to get background fields for the variational analysis - a machine learning limited-area model (ML-LAM) is run directly at resolution of 2.5 km to produce background fields.

The Environmental Modeling Center (EMC) has rich experience with the development of limited-area numerical models as well as with application of ML to weather forecasting. We looked at several possible routes toward the development of our new paradigm: replacing only some of the physics components in the current RRFS with its ML surrogates; building a whole ML model from scratch following the Graph Neural Network approach used in the Google’s GraphCast and applied at EMC in the Global Forecasting Version (GFS) version, or adopting ideas from the Physics Inform Neural Networks (PINNs) for solving the dynamics equations of a numerical model. However, in the first stage, we decided to simply apply a Graph Neural Network based Machine Learning Limited Area model (Graph ML-LAM), developed by Oskarsson et al. (2023) - https://arxiv.org/abs/2309.17370 on machines used at EMC, at different resolution and over different domain, using both the RRFS and the Model for Prediction Across Scales (MPAS) for training.

This presentation will discuss challenges and issues encountered during this development and present the first results of ML model runs derived initially at resolution of 3 km over Continental United States (CONUS) domain.