Reproduce “Leveraging data-driven weather models for improving numerical weather prediction skill through large-scale spectral nudging” on OpenResearch

About this paper

Syed Zahid Husain, Leo Separovic, Jean-François Caron, Rabah Aider, Mark Buehner, Stéphane Chamberland, Ervig Lapalme, Ron McTaggart-Cowan and 4 more

Researchers at Environment and Climate Change Canada developed a hybrid weather prediction system, GDPS-SN, that uses large-scale spectral nudging to integrate AI model (GraphCast) guidance into a physics-based NWP model (GEM). This system improves global forecast skill, yielding an 8-13 hour predictability increase for 500-hPa geopotential height in mid-latitudes by day 7, while preserving fine-scale details and physical consistency.

Key insights

• Data-driven AI models like GraphCast demonstrate superior skill in large-scale atmospheric circulation prediction but exhibit excessive smoothing at fine spatial scales, leading to deficiencies in local variability and physical consistency.
• Spectral nudging is an effective mechanism to integrate skillful large-scale guidance from AI models into physics-based NWP systems, enabling the latter to retain and generate fine-scale features while being constrained by a more accurate large-scale flow.
• The rapid inference speed and demonstrated large-scale accuracy of AI models now provide a viable, timely, and accurate reference for real-time global spectral nudging, which was previously impractical.

Summary by alphaXiv.