A Meteorological Foundation Model for Gap-filled High-Resolution Data in Urban Environments

Numerical Weather Prediction (NWP) and machine learning weather prediction models rely on data assimilation (DA) to determine the current state of the atmosphere, a necessary step for accurate forecasting. However, the computational demands of the physics-based algorithms used in operational NWP systems limit the volume of observations that can be assimilated in a timely manner. This, coupled with the coarse resolution of global NWP models, makes it dicult for NWP to represent complex processes in urban areas and other small-scale, high-impact weather phenomena. We previously developed EarthNet, a data-driven DA system that performs global assimilation and a downscaling model that predicts land surface temperature at 70m spatial resolution. We are currently running EarthNet hourly on the commercial cloud, performing global DA 50 times faster than operational systems. Building on this success, we propose EarthNet-Multi-Resolution (EarthNet-MR), a flexible resolution foundation model that will integrate diverse datasets ranging from government and commercial microwave sounders to radar and station data. Through iterative conditioning, EarthNet-MR will progressively incorporate finer-resolution data, ultimately combining kilometer-scale regional observations with global context to enable sub-hourly data assimilation for the Continental United States (CONUS). The robust representations learned during pretraining will enable us to fine-tune EarthNet-MR for severe weather nowcasting with a limited number of examples drawn from storm-resolving simulations over CONUS. The result will be a scalable and data-ecient system for merging sensing systems across scales, with a focus on high-impact weather and urban areas where detailed weather insights are critical. We expect our lower-latency system to capture the types of forecast surprises that NWP often misses, such as the location and timing of convective initiation and severe weather. These phenomena are of great interest in the energy trading markets, the insurance sector, and utility companies such as the Tennessee Valley Authority. To foster transparency and trust, we will pursue open source code development and create a web visualization that continuously validates EarthNet-MR outputs against observations.