AI set to take weather forecasting by storm

Weather forecasting blunders live long in the memory – Michael Fish and Britain’s 1987 Great Storm, Hurricane Katrina, and the 2003 European heatwave. Now hopes have been raised of more reliable predictions thanks to a revolution in meteorology driven by AI.

Papers released by technology companies Google DeepMind, Huawei and Nvidia appear to show impressive forecasting results driven by machine learning.

In 1987, during a forecast for the BBC, Fish mentioned that a woman had phoned in to say that a hurricane was on the way. ‘‘Well, if you’re watching, don’t worry, there isn’t!’’ he blithely told viewers. Had Fish had the benefit of AI, perhaps he would not have been so bold.

During the past few months, a quiet revolution has been brewing. Peter Dueben, head of earth system modelling at the European Centre for Medium-Range Weather Forecasts, described Google DeepMind’s efforts in particular as ‘‘quite astonishing’’.

‘‘It has really surprised the weather and climate community how good those predictions can be in principle,’’ he said. ‘‘There’s a lot happening.’’

This month, Atmo, a San Francisco-based startup, released what it claimed was the first live global medium-range weather forecast, purely based on an AI model, that was as good as the gold standard.

The difference? It is cheaper, faster and learns from its mistakes – and could herald street-by-street forecasts, according to experts.

Traditional forecasts rely on physics-based models. They solve equations based on the flow of air and moisture and the thermodynamics of the atmosphere.

The new AI models are prediction engines based on previous data – and they need a lot of it. Atmo’s model uses decades of data that amounts to 30,000 terabytes, equivalent to the memory required to store 15 million films or 10 billion photos.

The AI models were more agile, said Alexander Levy, the co-founder of Atmo.

‘‘Since this model can run in a few seconds, instead of hours, you get the information about the forecast much earlier. I think that is the main breakthrough.

‘‘The computers necessary to produce a fast forecast would before cost billions of dollars, whereas this can be done more cheaply or much more quickly, or both, or in greater detail.’’

Cheaper and faster means greater accuracy because weather forecasters can create a bigger ‘‘ensemble’’ forecast, which is where they re-run the model with different permutations to get a better prediction.

Kieran Hunt, an expert in tropical meteorology at the University of Reading in England, said hyperlocal forecasting was also on the horizon. ‘‘I think with these AI-based forecasts, we’re able to scale up to that kind of hyper-local forecasting where the quality will improve more quickly than I think it would if we stayed along a physics-based trajectory.’’

Atmo appears to have demonstrated this with a forecast for San Francisco that can predict medium range at the scale of 300m by 300m.

Those wanting forecasts beyond two weeks, however, may be disappointed. ‘‘The limit of 10 to 14 days is still almost an impossible wall to break,’’ said Johan Mathe, cofounder and chief technology officer of Atmo.