Buffio.

A field journal of machine-made science

← All discoveries

Climate & weather · 14 November 2023

Skillful 10-day global weather forecasts from a learned model

GraphCast produced 10-day global forecasts at 0.25° resolution in under a minute and outperformed the leading operational deterministic system on 90% of 1,380 verification targets, with better severe-event prediction.

Field
Climate & weather
Date
14 November 2023
AI system
GraphCast
Institution
Google DeepMind
Published in
Science
Identifier
10.1126/science.adi2336

Traditional numerical weather prediction improves by using more computing to run a physics-based model. It does not, in the usual sense, learn from the historical weather record to revise the model itself. GraphCast, published in Science on 14 November 2023 (print issue 22 December 2023), is a machine-learning method trained directly on reanalysis of past atmospheric conditions.

The model predicts hundreds of weather variables, globally, up to ten days ahead, at 0.25° resolution, and produces a forecast in under one minute. The authors report that GraphCast significantly outperforms the most accurate operational deterministic systems on 90% of 1,380 verification targets. Its forecasts also support better severe-event prediction, including tropical-cyclone tracking, atmospheric rivers and extreme temperatures.

GraphCast is a forecasting system, not a newly observed climate law. It is included here because a learned model, trained on the observational record, produced medium-range global forecasts that surpassed the leading physics-based operational deterministic model on the paper’s verification suite.