Google DeepMind uses artificial intelligence models to forecast hurricanes’ tracks and intensity before they arrive, in an effort to reduce computation time and improve the accuracy of warnings directed at communities and emergency officials. The company says WeatherNext models were trained on 50 years of historical weather data and can run on a single TPU processor instead of the massive computing infrastructure used by traditional models.
This was stated in an interview with Ferran Alet, a research scientist at Google DeepMind who participated in developing the WeatherNext models. Alet explains that the terms cyclone, hurricane, and typhoon refer scientifically to the same phenomenon, with the terminology differing according to geographic region.
From physical simulation to AI-assisted forecasting
Hurricane forecasting has historically relied on physical models that simulate the laws of fluid dynamics using data from global sensors. This process required supercomputers the size of shipping containers or buildings several floors high, and the available data could also be incomplete or noisy. According to Google DeepMind, forecast accuracy improved by roughly the equivalent of one additional day every decade, as a result of developing the models and increasing the number of satellites and computing capacity.
WeatherNext adds an artificial intelligence layer that learns patterns from past and current data to help estimate the future. The company says this approach achieved a leap equivalent to what had been expected over an entire decade of progress, within a single generation of models. It also points to improved forecasts of both the location where the hurricane will make landfall and its intensity at landfall.
What is changing in practice?
Weather agencies are considering extending the public forecast horizon from five days to seven days. The additional day could give authorities and communities more time to prepare, without eliminating the need for physical models or human review by meteorological experts.
Google DeepMind cites Hurricane Melissa as an example of the value of early warning. According to the interview, the company supported the National Hurricane Center before the hurricane season, and high-confidence AI forecasts contributed to the issuance of a warning about the hurricane’s rapid intensification after it initially appeared to be Category 1 before reaching Category 5. It also says that authorities in Jamaica gained an additional day for emergency preparations.
Making the models available to researchers
Google DeepMind announced the release of WeatherNext 2 models and WeatherNext hurricane models as open source, along with the interactive Weather Lab website. This availability is intended to enable academics and scientists to experiment with the models and conduct new research into extreme-weather forecasting.
certi.news’s perspective
The most important change here is not merely the use of artificial intelligence in meteorology, but the reduction in stated operating requirements from supercomputers to a single processor, along with the possibility of making the models available for scientific research. This could expand the range of organizations capable of testing hurricane forecasts, but the source is a presentation from Google DeepMind and this article does not provide independent comparative results or quantitative details about the margin of accuracy improvement. Therefore, WeatherNext’s practical effectiveness, especially when dealing with incomplete data or rapidly developing storms, remains a subject that requires assessment by weather agencies and independent researchers.