Learning skillful medium-range global weather forecasting

ABSTRACT

Global medium-range weather forecasting has long been dominated by massive, compute-intensive numerical weather prediction (NWP) models governed by atmospheric physics equations. We present GraphCast, an artificial intelligence model based on graph neural networks trained directly from decades of historical weather data. GraphCast generates 10-day global forecasts at 0.25° resolution in under one minute on a single Google TPU. Comprehensive evaluations show that GraphCast significantly outperforms the industry gold-standard NWP system (HRES from ECMWF) across 90% of test variables and lead times, demonstrating a profound leap in meteorological prediction capability.