Skip to main content
Rase AI Intelligence Newsletter

Research & education Research

GraphCast used AI to forecast global weather ten days ahead

In a 2023 Science study, Google DeepMind’s GraphCast produced a ten-day global forecast in under a minute on a Google TPU v4. It was more accurate than the comparison system on about 90% of the study’s 1,380 verification targets. [3][6]

Research output
A machine-learning model that produces a single global weather forecast up to 10 days ahead, with published research code and weights. [3][5]
Stage of evidence
Peer-reviewed research published in Science. [3][6]

What the researchers found

GraphCast learned from historical weather data and used recent atmospheric conditions as input. The team compared its forecasts with HRES, the European Centre for Medium-Range Weather Forecasts’ system, across variables, pressure levels and forecast lead times. [3][6]

Why it matters

Forecast centres gained evidence that a model trained on historical weather data could compete on a broad set of forecasting tests. Publishing the research code and weights let others repeat comparisons and investigate which conditions favoured each approach. [1][3][7]

Who can use the findings

Forecast researchers and weather centres can use GraphCast as a reference model for comparing global forecasts. People making weather-sensitive decisions should continue to use their local authorities’ forecasts and warnings. [3][5]

Key dates

  1. First preprint submitted [2]
  2. Science publishes GraphCast online [3][6]
  3. ECMWF ends its experimental GraphCast runs [4]
  4. Google updates the model weight license [5]

Practical uses

Treat the reported speed as the time for a research model run on the hardware used in the study. An operational forecast also needs a current estimate of the atmosphere, checks of forecast quality and a way to communicate warnings. Google DeepMind’s August 2026 update permits commercial use of its published GraphCast weights under CC BY 4.0. [3][5][6]

Limits of the evidence

GraphCast produces a single forecast from each starting state. Estimating probabilities requires additional methods. ECMWF found that changing the starting weather analysis reduced its accuracy on some measures. [3][7]

Sources and verification

Catalog evidence checked How we verify records

  1. Primary source Google DeepMind WeatherNext research code, including GraphCast
  2. Primary source GraphCast first preprint, arXiv:2212.12794 v1
  3. Primary source Science paper describing GraphCast, DOI 10.1126/science.adi2336
  4. Primary source ECMWF implementation notice for forecast cycle 50r1
  5. Primary source Google DeepMind’s GraphCast README and August 2026 weights-license update
  6. Primary source GraphCast: AI model for faster and more accurate global weather forecasting — Google DeepMind
  7. Independent source ECMWF’s account of its experimental external forecast models

Related profiles

Research & education

NVIDIA Nemotron for Science

NVIDIA

The collection packages technical components for researchers and scientific computing teams.

High confidenceEvidence checked

Research & education

Khanmigo

Khan Academy

It provides guided practice and supports teaching tasks through an education-focused chat interface.

High confidenceEvidence checked

Research & education

OpenScientist

Washington University in St. Louis / Lawrence Berkeley National Laboratory (C-BRAIN)

Helps researchers organize evidence from literature and analyze datasets as part of supervised research workflows.

High confidenceEvidence checked