Earth system forecasts, which predict a range of different things, including weather, air quality, ocean currents, sea ice, and hurricanes, are used to provide early warnings of extreme events.

Existing methods analyse decades worth of data and require a huge amount of computing power. That data has been fed into Aurora, and the researchers say the AI outperformed seven forecasting centres in predicting the route of cyclones five days into the future in 100% of cases, and in 92% of 10-day weather forecasts. Training Aurora took around two months, whereas existing methods can take years. However, it was only possible to train the AI so quickly because the data used to train it had already been collected by existing Earth forecasting centres, the researchers say.

Paris Perdikaris and colleagues note that Aurora is a foundation model for the Earth system and could be adapted for uses beyond weather forecasting. They conclude by saying that Aurora represents a development in efficient Earth system forecasting and highlights the potential of AI technology to provide wider access to weather and climate information.