In a paper published in Science, the GraphCast team presents an AI model that can predict weather conditions up to 10 days in advance with unprecedented accuracy and speed. The model, trained on decades of historical weather data, utilizes machine learning and Graph Neural Networks to make forecasts at a high resolution covering the entire Earth’s surface. Compared to the industry gold-standard weather simulation system, GraphCast provides more accurate predictions on over 90% of test variables and forecast lead times. It can also offer earlier warnings for extreme weather events such as cyclones and floods. The open-sourcing of the model’s code allows scientists and forecasters worldwide to benefit from its capabilities. Overall, GraphCast represents a significant advancement in AI for weather prediction, enabling better decision-making and potentially saving lives.
https://deepmind.google/discover/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/