Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
During hurricane season, meteorologists use a variety of different forecast models. As this season draws to a close, experts are taking stock of which companies have performed well and which haven’t, and Google’s rookie model has left them absolutely dumbfounded.
Although Google DeepMind’s Weather Lab only began releasing forecasts in June, it was by far the best model for predicting hurricane track and intensity this season, according to a preliminary analysis by Brian McNoldy, a meteorologist and senior researcher at the University of Miami. Meanwhile, America’s flagship weather model – the Global Forecast System – performed the worst.
The National Hurricane Center will release official data on each model’s performance in a few months, but this initial assessment suggests a turning point in hurricane forecasting. As the incredible superiority of AI-based models becomes apparent, it may be time to start phasing out traditional, physics-based models.
“Going forward, it’s safe to say we’ll be relying heavily on Google and other AI weather models, which are likely to improve in the coming years because they’re relatively new and have room for improvement,” Houston-based meteorologist and space reporter Eric Berger wrote for Ars Technica.
The rise of AI forecasting has begun
McNoldy’s analysis includes two charts: one that compares the track forecast accuracy for all 13 named storms in the Atlantic basin this season, and one that compares the intensity forecast accuracy for all 13 storms. The different colored lines represent different forecast models, which are indicated by the legend on the right edge. The lower a line is, the better the model performs.
The GFS – in this case referred to as AVNI – is shown in orange at the top of the charts. NOAA developed this model in the early 1980s, and the National Weather Service continues to use an updated version as its primary forecasting system today.
“The GFS was particularly dire in its forecast for Melissa, with an average 5-day track error rising to over 500 miles (800 kilometers) and insisting on a sea divergence that never materialized,” said Miami-based meteorologist and hurricane specialist Michael Lowry wrote in a recent blog post.
Unlike Google’s predictive model, the GFS is based on traditional physics and advanced supercomputers. The difference between them can be clearly seen in these diagrams. Google’s model ranks at the bottom, indicating superior performance to all other models evaluated – especially the GFS.
“The beauty of DeepMind and other similar data-driven, AI-based weather models is that they produce a forecast much more quickly compared to their traditional physics-based counterparts, which require some of the most expensive and advanced supercomputers in the world,” Lowry wrote. “In addition, these “intelligent” models with their neural network architectures have the ability to learn from their mistakes and correct on the fly.”
Better forecasts are urgently needed
Hurricane Melissa – which one devastated the Caribbean last week– is just one example of how rising sea surface temperatures are intensifying storms. As climate change causes hurricanes to become more deadly and damaging, it is important that forecasters have the best possible tools to predict their path and intensity.
AI-based models could help forecasters adapt to a warming world. DeepMind’s stunning debut has certainly caught their attention and could mark the start of a new era in hurricane forecasting.