Google’s new weather model impressed during its first hurricane season

As the Atlantic hurricane season comes to a close, meteorologists are reflecting on the effectiveness of forecasting models used throughout the year. Notably, Google DeepMind's Weather Lab, which only began issuing cyclone track forecasts in June, has emerged as a standout performer. In stark contrast, the Global Forecast System (GFS), managed by the US National Weather Service and reliant on conventional physics and powerful supercomputers, has struggled significantly. While the official performance data from the National Hurricane Center will not be released for several months, preliminary analyses by Brian McNoldy, a senior researcher at the University of Miami, reveal striking results. A chart summarizing the accuracy of track forecasts for all 13 named storms in the Atlantic Basin this season shows mean position errors over varying forecast hours, from 0 to 120 hours. The lower the line on this chart, the better the model's performance. The findings are clear: the GFS, represented in the chart as AVNI, is the least effective model, while Google DeepMind (GDMI), shown in maroon, consistently ranks as the top performer across nearly all forecast hours. The error margin for Google's model was notably lower, recording an error of 165 nautical miles at the five-day forecast mark, compared to 360 nautical miles for the GFS—more than double the error. Such discrepancies highlight the potential for forecasters to favor one model over another based on performance. Moreover, Google's model often surpassed the official forecasts generated by the National Hurricane Center, which are crafted by human experts analyzing a range of model data. It also outperformed respected consensus models like TVCN and HCCA, indicating a significant advancement in AI-driven forecasting technology.