GraphCast: Weather forecasting model
2 min readNov 16, 2023
DeepMind has developed an AI model for weather forecasting: GraphCast.
It can complete 10-day weather forecasts in less than a minute, with accuracy exceeding the European Center for Medium-Range Weather Forecasts (ECMWF)’s High-Resolution Weather Modeling System (HRES), which is recognized as a high standard in the industry.
It can also predict extreme weather events such as hurricanes and floods in advance.
Main features of GraphCast:
- High-precision weather forecast: GraphCast can provide weather forecasts up to 10 days, and its accuracy exceeds the industry standard High-Resolution Weather Modeling System (HRES), produced by the European Center for Medium-range Weather Forecasts (ECMWF).
2. Early warning of extreme weather events: GraphCast can predict extreme weather events earlier, such as accurately predicting the path of cyclones, identifying atmospheric rivers associated with flood risks, and predicting the occurrence of extreme temperatures. This ability helps save lives through better preparation.
3. Weather prediction system based on deep learning: GraphCast is a weather prediction system based on machine learning and graph neural networks (GNNs). Through training, GraphCast learns to recognize weather patterns and trends in this data. For example, it can learn to identify specific climate conditions that lead to storms or high temperatures.
4. Global coverage: It predicts at a high resolution of 0.25 degrees longitude and latitude on a global scale, covering more than one million grid points on the earth’s surface. The ability to provide weather forecasts on a global scale is useful for international travel, global business operations, and climate research.
5. Efficient prediction model: Although the training process of GraphCast is computationally intensive, the final prediction model is very efficient. A 10-day forecast with GraphCast takes less than a minute, whereas traditional methods like HRES can require hours of supercomputer calculations.
6. Continuous learning and adaptation: Over time, GraphCast can continue to learn from new meteorological data, continuously improving the accuracy and reliability of its predictions.
7. Wide application: GraphCast is used by several weather agencies, including ECMWF, which has run real-time experiments on model predictions on its website. https://t.co/r3RVQ7qPS4
8. Open source code: In order to make AI-based weather forecasting more popular, DeepMind has open sourced the code of the GraphCast model, so that scientists and forecasters around the world can benefit from it.
Detailed introduction: deepmind.google/discover/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/
Science Paper: science.org/doi/10.1126/science.adi2336
Paper PDF: https://t.co/storage.googleapis.com/deepmind-media/DeepMind.com/Blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/Learning_skillful_medium-range_global_weather_forecasting.pdf
open source code: github.com/google-deepmind/graphcast
