Indonesian storms, viewed from international space station
NASA Earth Observatory / International Space Station (ISS)
AI weather programs running on a desktop, can be a precision of traditional previews on the powerful surveys, scheduling hours, claim its creators.
The weather forecast has since, since the 1950s, has based models based on satellites, balloons and observations made using weather stations. But these calculations, known as numeric predictions (NWP) are very intensive and are based on numerous and expensive supercomputers.
In recent years, researchers have tried to facilitate this process by applying AI. Google scientists created a tool ai last year Replace small parts of the complex code In each weather model cell, it cuts the power of the computer incredibly. Deepmind later used even more and ai To replace the full forecast. This approach has been adopted by the European Center for medium-weather predictions (ECMWF), which launched a tool He called the artificial intelligence forecast system last month.
But the gradual role of the role of the Air prediction has been reduced to the crisps of the traditional number – a new model created by something Richard Turner They want to change at the University of Cambridge and his colleagues.
Turner says that previous work has been limited to the preview that has passed the initial step, for the preview of satellites, balloons and weather stations and weather stations have been planned. “It’s really half of the computational resources,” Turner says.
Researchers created the model called Aardvark weather, for the first time, the prediction and the initial phases. 10% of the incoming data is only used by the existing systems, but can obtain results for comparing the latest NWP predictions, report and their colleagues evaluate their method.
By creating a complete expectation, it would take hours or a few days in a powerful supercomputer for the NWP prediction, it can be done in a single desktop computer using aardvark.
However, Aardvark is using the Earth surface network with cells with 1.5 square degree, ECMWF’s ERA5 model uses a grid with cells As small as 0.3 degrees. This means Aardvark model to receive complex and unexpected weather models, says David Schultz University of Manchester, United Kingdom.
“Many unnecessary things are happening to turn on your forecast,” Schultz says. “They are not represented by the ends. They cannot fix this scale.”
Turner argues that Aardvark can actually pass some existing models to receive unusual events, such as cycles. But his example, his example, its models is based on these models based on physique training. “It doesn’t work if you take their training data away and use only observation data,” he says. “We tried to do that, and we went with a physical physics pattern, but that didn’t work.”
It believes that the future of weather forecasts can be scientists who work in more specific physics-based models, which are used to train AI models with less output and less hardware. Some are even more optimistic about the futures of Air.
Nikita Gourianov He thinks that in Oxford University, in time, AI can create weather forecasts that exceed the NWP. They will only be trained in historical weather data by creating specific predictions that are completely independent in NWP. “The scale question is, but also clearly. You need to have a relationship with how you structure the neural network and how you structure it.”
Themes:
