The Atlantic hurricane season is winding down, and as the tropics calm for winter's slumber, forecasters' attention turns to assessing what worked and what didn't in the previous season.
This year the answers are obvious. Although the Google DeepMind weather laboratory cyclone trajectory forecasts have just begun to be published The company's AI forecasting service performed exceptionally well in June. In contrast, the Global Forecast System model, run by the US National Weather Service, is based on traditional physics and runs on powerful supercomputers, but performs at abysmal rates.
Official data comparing the forecast model's performance won't be released by the National Hurricane Center for several months. However, Brian McNoldy, a senior research fellow at the University of Miami, has already drawn some conclusions. digital pre-processing.
The results are stunning:
Hurricane model performance for the 2025 Atlantic season on track accuracy.
Credit: Brian McNoldy
A little help reading charts wouldn't hurt. This chart summarizes the forecast accuracy for all 13 named storms in the Atlantic Basin this season, measuring the average position error at various forecast hours from 0 to 120 hours (five days). In this graph, the lower the line, the better the model performs.
New champion
The dashed black line shows the average forecast error for official forecasts for the 2022–2024 seasons. What stands out is that the leading US global model, GFS (denoted here as AVNI), is by far the worst model. Meanwhile, at the bottom of the chart, in maroon, is Google's DeepMind (GDMI) model, which performed best during almost all forecast hours.
The difference in errors between the US GFS model and Google's DeepMind is striking. Over five days, Google's forecast had an error of 165 nautical miles, compared to 360 nautical miles for the GFS model, more than twice just as bad. This is the kind of error that causes forecasters to completely ignore one model in favor of another.
But that's not all. Google's model was so good that it regularly beat the National Hurricane Center's official forecast (OFCL), which is compiled by experts looking at a wide range of model data. The AI-based model also outperformed highly rated “consensus models” including products from TVCN and HCCA. For more information about the different models and their designations, see see here.
