Downstream Applications on Prithvi WxC: A Foundation Model for Weather and Climate

Abstract

The use of deep learning for weather and climate analysis is rapidly gaining attention due to its ability to generate swift and precise weather forecasts, enhance climate models, and uncover new scientific insights. This shift towards data-driven forecasting and analysis prompts the question of how well these systems can adjust to varying temporal and spatial scales. The challenge lies in the variation of model inputs and outputs depending on the specific tasks and scales. We present downstream applications for Prithvi WxC, a foundation model for weather applications, trained on the Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2) dataset. The validation of Prithvi WxC extends over downstream tasks such as gravity wave flux parameterization derived using ERA5 U, V, T, and P variables, downscaling of weather and climate datasets, hurricane estimation, and the insertion of off-grid observational data. Additionally, we validate to which degree the model learns the underlying physics by adding several perturbations to the input data and comparing the response to that predicted by theory.