AI model provided the most accurate hurricane forecasts in 2025

The 2025 Atlantic hurricane season, which concluded on November 30, demonstrated the growing impact of artificial intelligence in predicting storms. Experts agree that this technology is now an integral part of hurricane forecasting.

This year, the National Hurricane Center (NHC) supplemented its traditional forecast models with AI-based systems, including one developed by Google DeepMind. Meteorologists reported that this AI tool allowed them to achieve what has long been considered the holy grail of hurricane prediction: accurately forecasting sudden changes in storm intensity.

While the accuracy of hurricane path predictions has improved over the years using conventional models, intensity forecastsespecially rapid intensificationhave remained less reliable. The DeepMind AI model enabled forecasters to confidently predict that Hurricane Melissa would intensify rapidly about three days before it struck Jamaica as a catastrophic Category 5 storm. This marked the first time the NHC predicted a storm would escalate to Category 5 at the time of its initial formation as a Category 1 hurricane.

As climate change increases the frequency of storms that intensify quickly, the ability to forecast such rapid changes becomes increasingly critical. In both real-time forecasts and end-of-season analyses, NHC officials credited the DeepMind model with providing greater confidence in their predictions.

Wallace Hogsett, science and operations officer at the NHC in Miami, explained that forecasters are becoming more comfortable relying on AI outputs, even if the inner workings of these models are not as fully understood as traditional physics-based systems. Traditional models, such as the European Model and the American Global Forecast System, rely on complex mathematical equations to simulate atmospheric processes and require powerful supercomputers to produce results.

In contrast, AI models leverage historical weather data to estimate a storms likely path and intensity. These models can run at high resolution in the cloud or even on personal computers within minutes. The DeepMind AI ensemble executes multiple simulations simultaneously, each with slightly varied initial conditions, providing forecasters with statistical insight into the reliability of predictions.

Hogsett noted, AI models will be a key part of hurricane forecasting moving forward. The Google DeepMind model, in particular, stands out as one of the top-performing tools.

However, he emphasized that performance over an entire season is more important than results from a single storm. Google DeepMind representatives also cautioned against over-interpreting results from one event. Still, former NHC branch chief James Franklin found that the DeepMind ensemble outperformed all other models in overall accuracy for track and intensity forecasts during the 2025 season, despite uneven performance across individual storms.

Forecasters acknowledge that AI is a black box compared to physics-based systems, as it identifies patterns from historical data rather than simulating atmospheric physics. This makes assessing reliability more challenging and highlights the importance of running multiple simulations in parallel.

Despite these challenges, NHCs end-of-season review highlighted clear improvements in forecasting rapid intensification. Senior hurricane specialist John Cangialosi noted that the field is experiencing its fastest period of advancement in two decades. He also emphasized that AI tools complement rather than replace traditional models or human expertise, pointing out that AI is prone to errors like any forecasting method.

While companies like Google, Microsoft, and Nvidia have led AI weather modeling, the National Oceanic and Atmospheric Administration and international partners are also developing AI systems, which will undergo testing in the next Atlantic hurricane season.

Author: Ethan Caldwell