An AI weather program running for a single second on a desktop can match the accuracy of traditional forecasts that take hours or days on powerful supercomputers, claim its creators.
Weather forecasting has, since the 1950s, relied on physics-based models that extrapolate from observations made using satellites, balloons and weather stations. But these calculations, known as numerical weather prediction (NWP), are extremely intensive and rely on vast, expensive and energy-hungry supercomputers.
In recent years, researchers have tried to streamline this process by applying AI. Google scientists last year created an AI tool that could replace small chunks of complex code in each cell of a weather model, cutting the computer power required dramatically. DeepMind later took this even further and used AI to replace the entire forecast. This approach has been adopted by the European Centre for Medium-Range Weather Forecasts (ECMWF), which called the Artificial Intelligence Forecasting System last month.
Advertisement
But this gradual expansion of AI鈥檚 role in weather prediction has fallen short of replacing all traditional number-crunching 鈥� something a new model created by at the University of Cambridge and his colleagues seeks to change.
Turner says previous work was limited to forecasting, and passed over a step called initialisation, where data from satellites, balloons and weather stations around the world is collated, cleaned, manipulated and merged into an organised grid that the forecast can start from. 鈥淭hat鈥檚 actually half the computational resources,鈥� says Turner.
The researchers created a model called Aardvark Weather that, for the first time, replaces both the forecast and initialisation stages. It uses just 10 per cent of the input data that existing systems do, but can achieve results comparable to the latest NWP forecasts, report Turner and his colleagues in a study assessing their method.
Generating a full forecast, which would take hours or even days on a powerful supercomputer for an NWP forecast, can be done in approximately 1 second on a single desktop computer using Aardvark.
However, Aardvark is using a grid model of Earth鈥檚 surface with cells that are 1.5 degrees square, while the ECMWF鈥檚 ERA5 model uses a grid with cells . This means Aardvark鈥檚 model is too coarse to pick up on complex and unexpected weather patterns, says at the University of Manchester, UK.
鈥淭here鈥檚 a lot of unresolved things going on that could blow up your forecast,鈥� says Schultz. 鈥淭hey are not representing the extremes at all. They can鈥檛 resolve it at this scale.鈥�
Turner argues that Aardvark can actually beat some existing models in picking up unusual events such as cyclones. But he concedes that AI models like his also rely entirely on those physics-based models for training. 鈥淚t absolutely doesn鈥檛 work if you take their training data away and just use the observational data to train off,鈥� he says. 鈥淲e did try to do that, and go completely physics model-free, but that didn鈥檛 work.鈥�
He believes the future of weather forecasting may be scientists working on ever-more accurate physics-based models, which are then used to train AI models that replicate their output faster and with less hardware. Some are even more optimistic about the prospects of AI.
at the University of Oxford believes that, in time, AI will be able to create weather forecasts that actually surpass NWP. These will be trained on observational and historical weather data alone, creating accurate forecasts entirely independent of NWP, he says. 鈥淚t鈥檚 a question of scale, but also a question of cleverness. You have to be clever with how you feed the data in 鈥� and how you structure the neural network.鈥�
Nature
