Google AI reportedly outperforms traditional weather forecasts for the first time, but…

Google’s subsidiary DeepMind has introduced a new weather forecasting system. The model outperformed the world’s leading system in 90 percent of the metrics used, the Financial Times reported. These variables also include temperature, pressure, wind speed and direction, as well as humidity at different levels of the atmosphere.

The GraphCast AI model “marks a turning point in weather forecasting,” Google developers DeepMind said in an article published in the journal Science.

As the evaluation shows, the system is more accurate for forecasts made three to ten days in advance than that of the European Center for Medium-Range Weather Forecasts. Plus, it only takes a fraction of the time.

Weather data from the past 40 years

GraphCast makes its predictions based on more than 40 years of historical weather data. This allowed it to learn how weather systems develop and create a ten-day forecast within a minute on one processor. With the conventional ECMWF method, multiple supercomputers take several hours to make a similar prediction.

DeepMind scientists cited Hurricane Lee, which hit the North Atlantic Ocean in September, as an example of successful weather forecasting. GraphCast was able to correctly predict the hurricane’s arrival nine days in advance, compared to six days with conventional approaches. “This gave people three more days to prepare for their arrival,” said Rémi Lam, lead author of the “Science” paper.

Conventional models more reliable during Hurricane Otis

But the system is far from perfect. The prediction of the sudden explosive force of Hurricane Otis on October 25 worked more reliably with the conventional models than with GraphCast. There is also the question of how well the system can predict extreme events affected by climate change if it only has past data.

In the near future, the AI-based model will not replace conventional weather forecasting. For ECMWF, the next step would be to “develop its own AI model and see if it can be combined with its numerical weather prediction system,” says Matthew Chantry, machine learning coordinator at the European Center for Medium-Range Weather Forecasts.

Scientists agree that the combination of different technologies will provide even more detailed weather forecasts in the future – important at a time of dramatic change. Google DeepMind has released the source code of GraphCast, making the technology available to all researchers worldwide.

Sources