How AI is taking the world of meteorology by storm

A model pioneered by the man who founded chaos theory is still being used by weather forecasters but a new method uses similar tech t to ChatGPT. Anthony Cuthbertson reports

- Anthony Cuthbertson

At 2.30am, Michael Fish was awoken by a tree blowing over in his back garden. It was 16 October 1987, and just 13 hours earlier, the weatherman had delivered what is now remembered as the UK’s most infamous forecast. “A woman rang the BBC and said she heard that there was a hurricane on the way,” he’d said during the previous lunchtime’s broadcast. “Well, if you’re watching, don’t worry, there isn’t.”

The hurricane-force winds that ensued ripped out not only Fish’s tree, but 15 million others across the south of England, destroyed billions of pounds’ worth of property, and killed 18 people. The next day, the front page of the Daily Mail featured the headline: “Why weren’t we warned?”

It was a reminder that not only is weather forecasting hard, it can be deadly when wrong. There have been significant improvements in the years since – most notably with the amount of computational power available to carry out weather simulations – yet the underlying prediction models remain the same.

This could be about to change in a big way. A rapid series of breakthroughs have uncovered brand new systems that can deliver unprecedented forecasting accuracy with just a fraction of the computing resources. And all in record time.

It began in 2020, when a scientific paper laid out a new method that was radically different from numerical weather prediction (NWP), which meteorologists have been using for decades to produce forecasts. NWP takes data from weather stations and satellites, and then uses complex equations involving things like fluid dynamics and wind turbulence to simulate how weather patterns might progress over time.