Google DeepMind's GraphCast AI weather predictor looks fascinating on paper but ...
... Even its creators say it ain't 'a replacement for traditional forecasting methods'
Google DeepMind claims its latest AI model is capable of generating ten-day weather forecasts in under a minute and is just as accurate as traditional predictive models running on supercomputers.
In a paper published in Science on Tuesday, researchers described GraphCast: a graph neural network made up of 36.7 million parameters. Trained on 39 years of data collected from 1979 to 2017 by the European Centre for Medium-Range Weather Forecasts (ECMWF) – a research institute that crunches global numerical weather predictions 24/7 – the system produces a ten-day weather forecast, split into six-hour increments.
The ECMWF's models rely on numerical weather prediction methods that run mathematical simulations modelling the motion of the atmosphere and oceans with fluid dynamics equations. GraphCast, however, looks at the weather patterns in satellite images, radar, and measurements from meteorological stations to make its predictions.
The model splits global maps containing information about the atmospheric and oceanic data into grids, and is trained to learn the relationships between the different weather variables that lead to specific events – like tropical cyclone tracks, atmospheric rivers, and heat waves. GraphCast predicts factors like the temperature, wind speed and direction, humidity, and air pressure at 37 different altitudes to help forecast the weather.
"In a comprehensive performance evaluation against the gold-standard deterministic system, [the ECMWF's High Resolution Forecast], GraphCast provided more accurate predictions on more than 90 percent of 1,380 test variables and forecast lead times," declared Remi Lam, lead author of the paper and a staff research scientist at Google DeepMind.
"When we limited the evaluation to the troposphere, the 6–20 kilometer high region of the atmosphere nearest to Earth's surface where accurate forecasting is most important, our model outperformed HRES on 99.7 percent of the test variables for future weather," he added.
- If it's going to rain within the next 90 mins, this very British AI system can warn you
- Here comes an AI that can predict hurricane strength. Don't worry, NASA made it so it probably actually works
- What could power an early-warning system for harmful radiation storms in Earth's Van Allen belts? AI? Let's see, say Los Alamos boffins
Although the training process for GraphCast is computationally intensive and required running a cluster of 32 Google's Cloud TPU v4 chips over four weeks, the final trained model can be run on a single Google TPU v4 machine (which includes four TPU v4 chips). A ten-day forecast can be generated in under a minute, beating the hours it typically takes numerical weather prediction models running on supercomputers.
Like all AI models, GraphCast's performance is dependent on the quality of its data. "GraphCast is now the most accurate ten-day global weather forecasting system in the world, and can predict extreme weather events further into the future than was previously possible. As the weather patterns evolve in a changing climate, GraphCast will evolve and improve as higher quality data becomes available," Lam claimed.
GraphCast's forecasts aren't perfect, however. The data it generates is sometimes incomplete, and spatial blurring occurs in areas of uncertainty – meaning its predictions might not be useful when trying to calculate probabilities of different weather events, known as ensemble forecasts. It also struggles to generate predictions for atmospheric data high in the stratosphere as well as the ECMWF's High Resolution Forecast system.
So can AI replace older, clunkier numerical weather prediction models? Not really, unfortunately.
The researchers admitted that GraphCast relied on traditional methods to obtain quality data in the first place, and that the ECMWF's High Resolution Forecast system can produce other types of forecasts that AI cannot yet.
"Our approach should not be regarded as a replacement for traditional weather forecasting methods, which have been developed for decades, rigorously tested in many real-world contexts, and offer many features we have not yet explored," they concluded in their paper.
"Rather our work should be interpreted as evidence that [machine learning-based weather prediction] is able to meet the challenges of real-world forecasting problems and has potential to complement and improve the current best methods."
Google DeepMind has released the code for GraphCast here. ®