Thunderstorm over Indonesia, seen from the International Space Station
NASA Earth Observatory / International Space Station (ISS)
An AI, we weather program that runs for a single second on a desk, can match the accuracy of traditional forecasts that take hours or days on powerful supercomputers, its creators claim.
Since the 1950s, weather forecasts have taken on physics -based models that extrapolate from observations made using satellites, balloons and weather stations. But these calculations, known as numeric weather prediction (NWP), are extremely intense and are dependent on huge, expensive and energy-hungry supercombus.
In recent years, researchers have tried to streamline this process by using AI. Last year, Google researchers created an AI tool that could replace small chunks with complex code in each cell in a weather model, which cut the required computing power dramatically. Deepmind later took this even longer and used AI to replace the entire prognosis. This approach has been adopted by the European Center for Medium-Drange Weather Prognosis (ECMWF), which launched a tool called Artificial Intelligence Forgnosis System last month.
But this gradual expansion of AI’s role in weather preparation has been Fauln shortly after replacing all traditional number-clips a new model created by Richard Turner at the University of Cambridge and his colleagues seeking to change.
Turner says previous work was limited to predicting, and adopted over a step called initialization, where data from satellites, balloons and weather stations from the world gather, clean, manipulate and merged into an organized grid from which the prognosis can start. “It’s actually half of the calculation,” says Turner.
The researchers created a model called Aardvark Weather, which is replaced for the first time both prognosis and initialization stages. It was only 10 per. Hundreds of the input data, as the existing system does, but can achieve results comparable to the latest NWP forecasts, Turner and his colleagues report in a study that wraps their method.
Generation of a full prognosis that will take hours or even days at PowerCle Supercominut to an NWP forecast can be for approx. 1 second on single desktop computer using Aardvark.
However, Aardvark uses a grid model of the soil surface with cells that are 1.5 degrees square, while the ECMWF’s ERA5 model used a grid with cells as small as 0.3 degrees. This means that Aardvark’s model is too rough to pick up on complex and axed weather patterns, says David Schultz at the University of Manchester, UK.
“There are a lot of unresolved things going on that can blow your forecast,” says Schultz. “They don’t represent the extremes at all. They can solve it on this scale.”
Turner claims that Aardvark can actually beat some existing models to pick up unusual events such as cyclones. But he admits that AI models like his also completely dependent on the physics-based models for training. “It definitely doesn’t work if you take away a training data and just use the observation data to train,” he says. “We tried to do it and go on physics-model-free, but it didn’t work.”
He believes that the future of weather forecasts may be researchers working with increasingly accident physics-based models, which are then used to train AI models that repeat their output faster and with less hardware. Some are even more optimistic in terms of the prospect of AI.
Nikita Gourianov at the University of Oxford believes that AI will eventually be able to create weather forecasts that news surpasses NWP. These will be trained in observation and historical weather data alone, which creates recognized forecasts quite independently of the NWP, he says. “It’s a question of scale, but also a question of cleverly. You have to be wise to how to feed the data in – and how to structure the neural network.”
Topics: