Abstract
We introduce a learned weather simulator—called “GraphCast”—which outperforms the most accurate operational medium-range weather forecasting system in the world. GraphCast operates on a0.25°latitude-longitude grid, and uses graph neural networks and a novel high-resolution multi-scale mesh representation, to autoregressively predict the 10-day temporal trajectories of 227 key dynamic variables that represent the state of the atmosphere, at 6-hour time intervals. Our results show GraphCast is significantly more accurate than the European Centre for Medium-Range Weather Forecasts (ECMWF)’s deterministic forecasting system, “HRES”, on89.3%of the 2760 target variables and lead times we evaluated. It also outperforms the most accurate previous machine learning weather forecasting model on98.8%of the 252 targets it reported. GraphCast is orders of magnitude faster than ECMWF’s operational systems, and can generate a 10-day forecast (35 gigabytes on disk) in under 60 seconds on Cloud TPU hardware. Together these results represent a key step forward in improving weather modeling with machine learning, open new opportunities for fast, accurate forecasting, and help realize the promise of machine-learning based simulation in the physical sciences.
Authors
Remi Lam, Alvaro Sanchez, Matthew Willson, Meire Fortunato, Peter Wirnsberger, Alexander Pritzel, Timo Ewalds, Zach Eaton-Rosen, Ferran Alet, Jacklynn Stott, Shakir Mohamed, ravuris , Oriol Vinyals, George Holland
Venue
Science