On the development of artificial intelligence downscaling applications for medium-range forecasts of weather elements at CCMEP

Abstract

Given the high computational cost of dynamical downscaling, the Canadian Center for Meteorological and Environment Prediction (CCMEP) could greatly benefit from efficient and performant Artificial Intelligence (AI) downscaling techniques. These techniques could be applied both to extend the range of high-resolution forecasts, and to develop high-resolution ensembles. Through a collaboration with IBM, an AI downscaling toolbox is currently being developed at CCMEP, consisting of a modular Python software package applicable to medium-range forecasts of various weather elements (surface winds, surface temperature and precipitation) from CCMEP’s global model, the Global Deterministic Prediction System (GDPS; 15-km resolution). This toolbox will be used to downscale forecasts to a grid-spacing of 2.5 km, the resolution of CCMEP’s High-Resolution Deterministic Prediction System (HRDPS). Considering future development beyond the IBM-CCMEP collaboration and an eventual operational implementation of this downscaling application, certain computational requirements need to be fulfilled. A tuned version of the toolbox, i.e., an AI downscaling model, will be trained and validated using data from GDPS and HRDPS. This presentation will describe preliminary results of training the AI downscaling model for surface winds using GDPS data as the low-resolution dataset, and HRDPS data as the high-resolution dataset. Two different AI architectures are explored: a small AI model based on Generative Adversarial Networks (GAN) and a large AI model based on combining a trained foundation model with an adapter layer to use for downscaling. GANs have been proven successful for downscaling applications and thus represent a reasonable benchmark to validate the large AI model.