Exploring the Benefits of the NASA/IBM Atmospheric Foundation Model in Deriving High-Resolution Precipitation Estimates from MERRA-2 Data

Precipitation plays a key role in the global, continental, and regional hydrologic cycles. However, the resolution of precipitation data derived from passive satellite sensors or reanalysis model such as the Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG, 10-km resolution) and the Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2; ~50-km resolution) is often too coarse to represent fine-scale precipitation features. This paper explores the capability of the recently developed NASA/IBM Atmospheric Foundation Model (FM) in enabling the conversion of low-resolution atmospheric reanalysis data into higher-resolution precipitation fields. We will utilize data from the MERRA 2 and collocate it with high-resolution precipitation estimates from the GPM Combined Algorithm (CORRA, 5-km resolution). A comprehensive dataset consisting of the FM-derived encodings of the MERRA-2 data and the associated CORRA precipitation estimates will be compiled. The underlying premise of the compiled dataset is that the NASA/IBM FM has learned the core principles of precipitation processes. In addition, this research will develop a regional precipitation downscaling model covering the Continental United States (CONUS) from 1 June 2014 to 31 May 2024 to fine-tune the FM. The proposed downscaling model will be based on an enhanced generative adversarial network (GAN), which has demonstrated effectiveness in similar applications. The compiled dataset will be used to train and evaluate the deep learning model, which will learn to predict high-resolution CORRA-like precipitation estimates based on the FM encodings of the MERRA-2 data.