Google Deepmind breakthrough could revolutionise weather forecasts, company says

A new artificial intelligence breakthrough could revolutionise weather forecasts, its creators say.

The new technology, built by Google Deepmind, allows for 10-day weather forecasts to be produced in just a minute. And it does so with unprecedented accuracy, Deepmind said.

The forecasts made by the GraphCast system are not only more accurate but produced more efficiently, meaning they can be made more quickly and with fewer resources.

It can also help spot possible extreme weather events, being able to predict the movement of cyclones and provide early alerts of possible floods and extreme temperatures. Google therefore says it could help save lives by allowing people to better prepare.

At the moment, weather forecasts usually rely on a system called Numerical Weather Prediction, which combined physics equations with computer algorithms that are run on supercomputers. That requires vast computing resources as well as detailed expertise by weather forecasters.

The new system is one of a range of technologies that instead use deep learning. Instead of looking at physical equations, it learns from weather data and then uses that to model how the Earth’s weather changes over time.

Creating the model was intensive, since it required training on decades of weather data. But now that it is created it could vastly reduce the resources required for predicting the weather: 10-day forecasts take a minute on one machine, a process that might otherwise take hours and use hundreds of machines in a supercomputer.

In use, the system was able to provide more accurate forecasts than the gold-standard traditional system in 90 per cent of tests, its creators write in a paper newly published in the journal Science.

What’s more, the system is able to spot extreme weather events despite not being trained on it. In September for instance it had predicted the path of Hurricane Lee nine days before it arrived, compared to six days for traditional forecasts.

Deepmind noted that GraphCast’s prediction of extreme temperatures could be particularly useful given the climate crisis. The system can predict areas where the heat will arrive above the historical top temperatures, allowing people to anticipate heat waves and prepare for them.

The company will also open source the system so that it can be used by others. That may help with other new tools and research to help tackle environmental challenges, Deepmind said.