Leveraging Data-Driven Weather Forecasting for Improving Numerical Weather Prediction Skill Through Large-Scale Spectral Nudging

Operational weather forecasting has traditionally relied on physics-based numerical weather prediction (NWP) models, but the rise of AI-based weather emulators is reshaping this paradigm. However, most data-driven models for medium-range forecasting still face limitations, such as a narrow range of predicted variables and low effective spatiotemporal resolution. This presentation will compare the strengths and weaknesses of these two approaches, using Environment and Climate Change Canada’s Global Environmental Multiscale (GEM) model and Google DeepMind’s GraphCast model. It will demonstrate that GraphCast outperforms GEM in predicting large-scale features, particularly for longer lead times.

Building on these findings, we propose a new hybrid NWP-AI system, in which GEM’s large-scale state variables are spectrally nudged towards GraphCast’s inferences, while GEM continues to generate fine-scale details critical for weather extremes. Results show that this hybrid system improves GEM’s forecast accuracy, reducing RMSE for the 500-hPa geopotential height by 5-10% and extending predictability by 6-12 hours in the extratropics, peaking at day 7 of the forecast. It also yields significant improvements in tropical cyclone trajectory prediction without degrading intensity forecasts. Unlike state-of-the-art AI-based models, the hybrid system ensures meteorologists retain access to all forecast variables, including those critical for high-impact weather. Preparations are currently well underway for the operationalization of this hybrid system at the Canadian Meteorological Centre.