Towards state-of-the-art nowcasting with a foundational AI model: an ECCC-IBM collaboration

Abstract

Recent progress in AI for nowcasting precipitation (e.g. NowcastNet) and other surface variables (e.g. MetNet-3) has been staggering. The move from advection-based nowcasting using PySteps to the state-of-the-art AI-based nowcasting methods requires an expertise not currently found in ECCC. We present here some results from a collaborative endeavour between ECCC and IBM. First, as a baseline, we compare a PySteps-based radar precipitation nowcasting with a version of NowcastNet optimized for the Canadian domain. Then, we explore how the coupling of NowcastNet with a foundational AI model such as AtmoRep can enhance and expand the nowcasting capabilities by providing guidance on the large-scale dynamic. Finally, we lay the vision on how this approach can be expanded to other predictands (geostationary satellite, lightning, and surface observations) and more predictors (surface variables and clouds). The goal is thus not only to move towards the operationalization of state-of-the-art radar precipitation nowcasting algorithms, but also to push further the nowcasting state-of-the-art with cutting edge AI research.