Especially in light of climate change, modelling medium-range (next 10 days) weather is becoming increasingly important. The current industry gold-standard is the High Resolution Forecast (HRES), produced by the European Centre for Medium-Range Weather Forecasts (ECMWF). It leverages Numerical Weather Prediction (NWP), which are complex physical equations, translated into algorithms run by super computers. It takes several machines hours to run the NWP.
Researchers at Google DeepMind recently launched GraphCast, an artificial intelligence (AI) and machine learning (ML) based model to make medium-range weather forecasts. The model can make the predictions, using a single Google TPU v4 machine, in under a minute and is more accurate then HRES in 90% (overall) to 99.7% (only troposphere – the most important forecasting layer) of the 1380 tested cases. Furthermore, it can also make accurate predictions of extreme weather events, like the paths of cyclones or increased heat risks.
GraphCast’s model is trained on more than 40 years of ECMWF weather data, including satellite images, radar and weather stations, complemented by a classical NWP to fill in missing observations. As the code is open source, researchers and weather agencies around the world can leverage the model; even the ECMWF is currently running a live test of the model on their website (ECMWF | Charts).
Leveraging AI and ML to make more accurate predictions about future weather can make a significant difference, saving millions (if not billions) of lives in the long-run and preventing catastrophic economic losses. For example, in September 2023, the ECMWF used GraphCast to track Hurricane Lee. While HRES could predict landfall in Nova Scotia only six days in advance, GraphCast could accurately predict landfall nine days in advance, leaving citizens and the government three more valuable days to prepare.
Research Paper: Learning skillful medium-range global weather forecasting | Science
Google DeepMind Article: GraphCast: AI model for faster and more accurate global weather forecasting – Google DeepMind