AI Predicts Weather Instantly Without a Supercomputer

Its creators claim that AI weather programs running for a second on the desktop can match the accuracy of traditional predictions that take hours or days on a powerful supercomputer.

Weather forecasts rely on physics-based models that extrapolate from observations made using satellites, balloons and weather stations since the 1950s. However, these calculations, known as numerical weather forecasts (NWPs), are highly concentrated and rely on vast, expensive, energy-hungry supercomputers.

In recent years, researchers have tried to streamline this process by applying AI. Last year, Google Scientists created an AI tool that could replace a small chunk of complex code in each cell of a weather model, dramatically reducing computer power. DeepMind later went further by doing this, using AI to replace the entire prediction. This approach is adopted by European Medium-Range Weather Forecast (ECMWF). The tool has been launched Last month it was called the Artificial Intelligence Prediction System.

However, this gradual expansion of the role of AI in weather forecasting has not replaced the calculation of all traditional figures – the new model created by Richard Turner Cambridge University and his colleagues are looking for change.

Turner says that previous work was limited to prediction and passed a step called initialization. There, data from satellites, balloons and weather stations around the world is collated, washed, manipulated and integrated into an organized grid where predictions can begin. “It’s actually half the computational resource,” Turner says.

The researchers created a model called Aardvark Weather. This replaces both the prediction and initialization stages for the first time. It uses only 10% of the input data that existing systems make, but achieves results comparable to the latest NWP predictions. Turner and his colleagues report in a study assessing the method.

Generating a perfect prediction that takes hours or days on a powerful NWP prediction supercomputer can be done in about a second on a single desktop computer using Aardvark.

However, Aardvark uses a grid model of the Earth’s surface with a square cell of 1.5 degrees, while ECMWF’s ERA5 model uses a grid with cells. 0.3 degrees smaller. This means that Aardvark’s model is too rough to pick up complex and unexpected weather patterns, David Schultz At the University of Manchester, UK.

“There are a lot of unresolved things that could blow up predictions,” Schultz says. “They don’t represent any extremes at all. They can’t solve it on this scale.”

Turner argues that Aardvark can actually beat some existing models. However, he acknowledges that AI models like him also rely entirely on these physics-based models. “It’s absolutely not working just to steal training data and train with observational data,” he says. “We tried to do that and did a complete modelless physics, but it didn’t work.”

He believes the future of weather forecasting could be scientists working on more accurate physics-based models. This is used to train AI models that replicate output faster and with less hardware. Some are even more optimistic about the AI ​​outlook.

Nikita Gouryanov At Oxford University, we believe that AI will eventually be able to produce weather forecasts that actually exceed NWP. They are trained solely on observational and historical weather data, and produce accurate predictions that are completely independent of the NWP, he says. “It’s a matter of scale, but also a matter of smartness. You have to be smart about how you deliver data and how you build the structure of a neural network.”