In September the system predicted about 9 days in advance that Hurricane Lee would make landfall in Nova Scotia
A new Google DeepMind artificial intelligence model is the world’s most accurate global weather forecasting system, according to the London-based company.
GraphCast, as the system is called, promises medium-term weather forecasts of unprecedented accuracy. In a study published yesterday (11/14) in the journal Science, GraphCast was shown to be more accurate and faster than the industry gold standard for weather simulation, the High-Resolution Forecast (HRES). The system also predicted future extreme weather events.
In September the system predicted about nine days in advance that Hurricane Lee would make landfall in Nova Scotia. In contrast, traditional methods predicted the hurricane would hit Nova Scotia only six days in advance. They also provided less consistent forecasts of hurricane timing and location. GraphCast can detect hazardous weather without being trained to do so. After incorporating a simple cyclone detector, the model predicted cyclone movements more accurately than the HRES method.
Such data could save lives. As the climate becomes more extreme and unpredictable, fast and accurate forecasts will provide increasingly vital information for dealing with extreme weather events. Matthew Chantry, machine learning coordinator at ECMWF, believes his industry has reached a tipping point. “More work may need to be done to create reliable business tools, but this is likely the beginning of a revolution,” he told a news conference.
Weather organizations, he added, believed that artificial intelligence would be more useful if it merged with physics. But recent discoveries show that machine learning can also directly predict the weather.
How GraphCast works
Conventional weather forecasts are based on complex physics equations. These are then adapted into algorithms run on supercomputers. This process is difficult and requires specialized knowledge and enormous computing resources. GraphCast leverages a different technique. It combines machine learning with graph neural networks (GNNs), an architecture that can process spatially structured data.
To learn the causes and effects that determine weather changes, the system was trained on decades of weather information. ECMWF fed GraphCast with 40 years of weather data, which included data from satellites, radars and weather stations. When there are gaps in observations, physics-based prediction methods fill them in. The result is a detailed history of global weather. GraphCast uses data to predict the future.
In tests, the results were impressive. GraphCast significantly outperformed the most accurate operational deterministic systems on 90% of 1,380 test targets. The difference was even more significant in the troposphere – the lowest layer of Earth’s atmosphere and the location of most weather phenomena. In this area, GraphCast outperformed HRES on 99.7% of the future weather test variables.
GraphCast is also extremely efficient. It takes less than a minute to complete a prediction on a single Google TPU v4 machine. A conventional approach, by comparison, can require hours of computation on a supercomputer with hundreds of machines.
The role of artificial intelligence in weather forecasting
Despite promising early results, GraphCast could still be improved. In cyclone forecasts, for example, the model proved accurate in tracking movements, but less effective in measuring intensity. “For now, this is one area where GraphCast and machine learning models are still a bit behind physical models. I’m optimistic that this can be improved, but it shows that this is still a very early stage technology,” Chantry said.
These improvements could now come from anywhere, since DeepMind is now open-sourcing the model. Global organizations and individuals can now experiment with GraphCast and add their own enhancements. The potential applications are unpredictable. Forecasts could, for example, inform renewable energy generation and air traffic routing. But they could also be applied in areas we don’t even imagine.
