AI Weather Models Show Promise in Hurricane Forecasting

This past hurricane season, AI-assisted weather models were tested, with experts suggesting they may become essential for future hurricane forecasting. At the start of the 2025 Atlantic hurricane season, the National Hurricane Center (NHC) announced a partnership with Google DeepMind to evaluate its latest AI weather model within the center's tropical cyclone forecasting workflow.

Michael Brennan, the director of the NHC, stated, "This collaboration between NOAA and Google will ensure that NOAA's National Hurricane Center is able to quickly evaluate new tropical cyclone forecasting technology as it arises."

NOAA reported that after integrating DeepMind's model into forecasters' tools, it outperformed traditional models in some instances. A notable example was the forecast for Hurricane Melissa, where DeepMind's model and its European counterpart provided forecasters with a high level of confidence that the storm would rapidly intensify into a major Category 5 hurricane before impacting Jamaica.

Matt Lanza, managing editor of The Eyewall blog, expressed his admiration for the DeepMind model's performance, particularly in handling rapid intensification. He remarked, "I was really impressed with [the DeepMind model's] ability to handle rapid intensification, as that has been a thorny issue with a lot of these types of models. The work it did during Melissa was unquestionably critical in terms of sounding the alarm on very high-end risk."

Overall, NOAA found that DeepMind's model was the most accurate for storm track and intensity, with only the NHC's official forecasts surpassing it in accuracy. James Franklin, a former branch chief at the NHC, referred to the results as a "banner year" for the Google DeepMind model. However, a spokesperson from Google DeepMind cautioned against drawing conclusions based on a single storm or metric.

Traditional weather models, such as the European Model and the American Global Forecast System, rely on complex simulations using atmospheric and physics-based equations, which require extensive time and computing resources. In contrast, AI weather models like DeepMind's are faster and require less computing power. They learn to forecast by analyzing historical weather data and identifying patterns from past storms, enabling them to generate forecasts in seconds.

Ryan Torn, a professor at the University of Albany specializing in weather modeling, explained, "AI/Machine Learning models use information from past weather fields, typically 40+ years of how the atmosphere looks at 6-hour intervals, to 'learn' how the atmosphere evolves with time. Once the AI/Machine Learning model learns the atmosphere, you can give it information about what the atmosphere looks like at the current time to create a forecast."

Google claims that its DeepMind AI model can produce hundreds of different weather scenarios from a single starting point in minutes, while traditional models may take hours for the same task. Despite the advantages observed during the 2025 hurricane season, experts emphasize the need for further research and testing before AI models can dominate forecasting.

Lanza noted that models like DeepMind still need to demonstrate their effectiveness with Gulf storms, especially since this hurricane season was relatively quiet. He stated, "When you think about extreme weather and climate change, you need to also think that events outside the bounds of what's expected will occur, and AI modeling may not capture that risk. That's where traditional physics-based modeling may remain essential."

Torn also highlighted the importance of traditional models, noting that AI models might smooth out differences caused by sudden weather changes or data gaps, which could lead to inaccuracies. He said, "Spreading out the differences over a larger area is not physically meaningful and will hurt the forecast later on. Our fully physics models tend to do much better at this because we have incorporated the physical laws into the models."

A spokesperson from Google DeepMind reiterated that while AI weather models do not process complex physical and atmospheric equations like traditional models, they complement these models by using them for training and initial conditions, thereby combining speed and accuracy to enhance predictions.