The potential for artificial intelligence to improve weather forecasting is outlined by studies published by two groups of scientists today.
One method can predict global weather patterns up to a week in advance, whereas the other can make short-term weather predictions, such as extreme downpours, which are expected to become more frequent and intense with global heating.
There are concerns that AI techniques should be properly validated before they can replace existing forecasting methods though a weather forecasting pioneer, Tim Palmer of Oxford University, comments that he expects future forecasting methods to combine AI with conventional so-called ‘physics based’ methods.
At present, supercomputers make weather predictions a few days into the future – which can save lives when it comes to, for example, anticipating destructive storms and tropical cyclones – with reasonable accuracy by constantly updating sophisticated computer models with data from satellites, buoys, aircraft, ships and weather stations.
A typical ‘physics based’ weather forecasting computer model relies on a system of mathematical equations (known as partial differential equations) to simulate whether it is going to rain or shine. In all, it takes hundreds of equations for each location to model the planet down to a resolution of, typically, around 60 to 25 kilometres. Moreover, today’s forecasters do not just make one forecast but ‘ensembles’ of many, to take into account uncertainties, which makes them even more reliable, though they are couched in probabilities.
However, these traditional weather forecasting methods demand a lot of computational power – the Met Office uses a Cray supercomputer – and are often slow. Instead of making predictions on the basis of an understanding of physics, AI methods forecast weather patterns that are statistically plausible in the light of historical measurements. In recent years, these AI-based methods have shown potential in significantly accelerating weather forecasting, although the accuracy is usually lower than physics-based methods.
Today, in Nature, Qi Tian and colleagues from Huawei in Shenzen, China, present a promising AI-based weather forecasting system called Pangu-Weather that can predict global weather up to a week in advance.
The system, a so-called deep neural network that consists of simple processing nodes that are densely interconnected, is trained for forecasting, in this case using 39 years of data produced by physics-based methods.
Pangu-Weather produced forecast results with accuracy comparable to the world’s best numerical weather prediction system, run by the European Centre for Medium-Range Weather Forecasts, ECMWF, while being more than 10,000 times faster.
The Huawei team believe it should be quicker to use it for ensemble forecasts too, though Palmer commented: “Until Pangu Weather have developed their ensemble system, the ECMWF ensemble system will be inherently more reliable (and hence more skilful) than the Pangu deterministic system.”
In a separate study in Nature, Mike Jordan of the University of California, Berkeley, Jianmin Wang of Tsinghua University, and colleagues present NowcastNet, which blends physics-based forecasting, based on fluid flow equations, with deep learning to provide ‘nowcasting’ of precipitation. “NowcastNet is precisely a method that combines AI with physics-based methods,” says Jordan.
Nowcasting is weather forecasting over a very short term up to six hours ahead, and is seen as valuable in anticipating extreme downpours and flooding, for example. Based on radar observations from the United States and China, NowcastNet predicted precipitation with high resolution over regions of 2,048 km × 2,048 km up to three hours in advance.
The forecast skill and value of different models for extreme precipitation was evaluated by 62 meteorologists; NowcastNet ranked first in around 70% of cases against other methods and the team say its accuracy can be honed by introducing more physics, more observations, for instance by satellites, and the quantification of uncertainty.
“They should be able to say, “I’m not sure”, and have that option be produced under some kind of calibration,” says Jordan. “Hurricane forecasting has very usefully provided such uncertainties in recent years, and I think that people have found them very helpful.”
Stephen Belcher, Science Museum Trustee and Met Office Chief Scientist, adds: “These papers show clearly the rapid and impressive progress in the application of AI to weather prediction and open up interesting possibilities when combined with our current physics-based models.”
A combination of AI and current methods in what he calls ‘Big AI’ is necessary because AI has important limitations, comments Peter Coveney, Director of Computational Science at UCL. “Among the shortcomings of the AI approach is that it is energy intensive to train AI, being comprised of enormous numbers of fitting parameters whose uncertainty influences predictions in unknown ways, an AI might not have been trained on rare-but-extreme events, and AIs are unreliable in conditions they have not encountered before. The latter point is important because the global climate is also entering uncharted waters, with the hottest day since records began, when there is an increasing risk of tipping points. Finally, AIs are based on algorithms that assume digital computers can handle all the complexity of the natural world, although we know of situations where that is not the case.”