Learning skillful medium-range global weather forecasting

Area of Science:

• Meteorology and Climate Science
• Artificial Intelligence
• Dynamical Systems Modeling

Background:

• Accurate medium-range global weather forecasting is vital for societal and economic sectors.
• Traditional numerical weather prediction models improve accuracy with more compute power but do not leverage historical data directly.
• Existing methods face limitations in speed and direct utilization of historical weather patterns.

Purpose of the Study:

• To introduce GraphCast, a novel machine learning (ML) based weather forecasting method.
• To demonstrate GraphCast's capability to predict global weather variables for the next 10 days with high resolution and speed.
• To evaluate GraphCast's performance against state-of-the-art operational weather forecasting systems.

Main Methods:

• Developed GraphCast, an ML model trained directly on historical weather reanalysis data.
• Utilized a graph neural network architecture to process and learn from spatio-temporal weather patterns.
• Evaluated forecast accuracy across 1380 verification targets for a 10-day forecast horizon.

Main Results:

• GraphCast achieved global forecasts of hundreds of weather variables at 0.25° resolution in under one minute.
• The ML model significantly outperformed the most accurate operational deterministic systems on 90% of verification targets.
• GraphCast demonstrated superior capabilities in predicting severe weather events, including tropical cyclones, atmospheric rivers, and extreme temperatures.

Conclusions:

• GraphCast represents a significant advancement in accurate and efficient medium-range global weather forecasting.
• The ML-based approach offers a powerful alternative to traditional numerical weather prediction, leveraging historical data effectively.
• This work highlights the potential of machine learning for modeling complex dynamical systems like Earth's weather.