AI can predict the weather 10 days ahead more accurately than current state-of-the-art simulations, says AI firm Google DeepMind – but meteorologists have warned against abandoning weather models based in real physical principles and just relying on patterns in data, while pointing out shortcomings in the AI approach.
Existing weather forecasts are based on mathematical models, which use physics and powerful supercomputers to deterministically predict what will happen in the future. These models have slowly become more accurate by adding finer detail, which in turn requires more computation and therefore ever more powerful computers and higher energy demands.
Rémi Lam at Google DeepMind and his colleagues have taken a different approach. Their GraphCast AI model is trained on four decades of historical weather data from satellites, radar and ground measurements, identifying patterns that not even Google DeepMind understands. “Like many machine-learning AI models, it’s not very easy to interpret how the model works,” says Lam.
To make a forecast, it uses real meteorological readings, taken from more than a million points around the planet at two given moments in time six hours apart, and predicts the weather six hours ahead. Those predictions can then be used as the inputs for another round, forecasting a further six hours into the future.
Researchers at DeepMind ran this process with data from the European Centre for Medium-Range Weather Forecasts (ECMWF) to create a 10-day forecast. They say it beat the ECMWF’s “gold-standard” high-resolution forecast (HRES) by giving more accurate predictions on more than 90 per cent of tested data points. At some altitudes, this accuracy rose as high as 99.7 per cent.
Matthew Chantry at the ECMWF, who worked with Google DeepMind, says his organisation had previously seen AI as a tool to supplement existing mathematical models, but that in the past 18 months it has come to be regarded as something that could actually provide forecasts on its own.
“We at the ECMWF view this as a hugely exciting technology to lower the energy costs of making forecasts, but also potentially improve them. There’s probably more work to be done to create reliable operational products, but this is likely the beginning of a revolution – this is our assessment – in how weather forecasts are created,” he says. Google DeepMind says that making 10-day forecasts with GraphCast takes less than a minute on a high-end PC, while HRES can take hours of supercomputer time.
But some meteorologists have expressed caution about turning weather forecasting over to AI. Ian Renfrew at the University of East Anglia, UK, says GraphCast currently lacks the ability to marshal data for its own starting state, a process known as data assimilation. In traditional forecasts, this data is carefully placed into the simulation after thorough checks against physics and chemistry calculations to ensure accuracy and consistency. Currently, GraphCast needs to use starting states prepared in the same way by the ECMWF’s own tools.
“Google is not going to be running weather forecasts anytime soon, because they cannot do the data assimilation,” says Renfrew. “And the data assimilation is typically half to two-thirds of the computing time in these forecasting systems.”
He says that he would also have concerns about ditching deterministic models based on chemistry and physics entirely and relying on AI output alone.
“You can have the best forecast model in the world, but if the public don’t trust you, and don’t act, then what’s the point? If you set out an order to evacuate 30 miles of coastline in Florida, and then nothing happens, then you’ve blown decades of trust that has been built up,” he says. “The advantage of a deterministic model is you can interrogate it and if you do get bad forecasts, you can interrogate why they’re bad forecasts and try to target those aspects for improvement.”