Google LLC today detailed a new artificial intelligence model, GraphCast, that it says can generate weather forecasts faster and more accurately than traditional algorithms.

The neural network was developed by the company’s recently formed Google DeepMind research unit. DeepMind has made the source code for GraphCast available on GitHub. According to Google, multiple weather agencies are already using the model to support their work.

GraphCast is a so-called GNN, a specialized neural network optimized to process graphs. A graph is a data structure that can store multiple snippets of information as well as record how they relate to one another. Graphs’ ability to store large, interconnected datasets makes them useful for describing complex phenomena such as weather events.

Google trained GraphCast on nearly four decades’ worth of weather observations collected by the European Centre for Medium-Range Weather Forecasts, or ECMWF. The training dataset included information from satellites, radar systems and other sources. According to Google, ECMWF is one of the weather organizations that have adopted GraphCast to support their work.

GraphCast can forecast the weather up to ten days in advance. Moreover, it does so with a high degree of granularity. Google says GraphCast can predict temperature, humidity levels, wind speed and other variables at dozens of different altitude levels.

In an internal test, the company compared GraphCast with a weather forecasting algorithm called HRES that is known for its reliability. GraphCast achieved higher accuracy than HRES across more than 90% of the 1,380 weather variables that were analyzed during the test. The AI system performed even better when the scope of the task was limited to the troposphere, the part of the atmosphere that is located directly above the Earth’s surface.

“When we limited the evaluation to the troposphere, the 6- to 20-kilometer high region of the atmosphere nearest to Earth’s surface where accurate forecasting is most important, our model outperformed HRES on 99.7% of the test variables for future weather,” Google DeepMind staff research scientist Rémi Lam detailed in a blog post. 

During the test, Google’s researchers also determined that GraphCast can outperform traditional forecasting algorithms in another area: anticipating extreme weather events. Though it wasn’t trained to do so, GraphCast managed to predict a cyclone’s movements more accurately than HRES. Moreover, Google believes that the AI system could prove useful for forecasting floods.

Traditional weather forecasting algorithms require a significant amount of hardware to run because they rely on complex physics equations to generate predictions. As a result, such algorithms are often deployed on supercomputers. Even with a supercomputer’s hardware resources, generating a forecast can take several hours in some cases.

Google says that GraphCast requires significantly less infrastructure. According to the company, the AI system can generate 10-day weather forecasts using only a single Google Cloud TPU v4 instance. Moreover, it’s capable of generating those forecasts in under a minute.

“Pioneering the use of AI in weather forecasting will benefit billions of people in their everyday lives,” Lam wrote. “But our wider research is not just about anticipating weather – it’s about understanding the broader patterns of our climate. By developing new tools and accelerating research, we hope AI can empower the global community to tackle our greatest environmental challenges.”

Photo: Pixabay