Artificial intelligence has significantly advanced weather forecasting capabilities in recent years, producing global forecasts that rival some of the top physics-based models. This progress stems from the availability of extensive weather data, improvements in AI algorithms, and enhanced computational resources. However, despite these strides, AI continues to face substantial challenges when predicting the intensity of hurricanes, particularly during periods of rapid intensification.
At a global scale, AI benefits from decades of comprehensive climate and weather records covering the entire planet. These large datasets allow AI models to detect atmospheric patterns and trends with unprecedented accuracy. Yet, when the focus shifts to the regional level—where hurricanes develop and evolve—the forecasting task becomes considerably more complex. Hurricanes can intensify swiftly, transforming from tropical storms into Category 5 hurricanes within hours, as exemplified by Hurricane Polo’s rapid escalation off Mexico’s Pacific coast in September 2026. Such rapid changes complicate preparedness efforts, as seen in past events like Hurricane Michael in 2018, which caught communities off guard.
Two primary sources of data underpin AI models that aim to predict hurricane intensity: observational measurements and weather model simulations. Observational data include rainfall, wind speeds, and temperature readings collected from weather stations, radars, buoys, and satellites. Although highly detailed, these observations are often unevenly distributed and mostly concentrated near coastlines. Many crucial phases of hurricane development occur over open oceans where measurements are sparse. Satellites help bridge these gaps but cannot fully capture the three-dimensional structure and dynamics of all active hurricanes at all times.
Weather model simulations provide a more comprehensive, three-dimensional view of the atmosphere at high resolution by combining physics-based understanding with observed data. However, these models inherently contain approximations and uncertainties, leaving some fine-scale physical processes unresolved. As a result, neither observational nor simulation datasets in isolation offer the complete, precise data needed to train AI models for accurate hurricane intensity forecasts.
Beyond data limitations, the inherent chaotic nature of hurricanes poses a fundamental challenge. Minor differences in the initial state of a storm can lead to substantial variations in intensity forecasts over time. Research suggests that hurricane intensity fluctuations may follow a chaotic attractor—a range of possible storm states within which evolution is inherently unpredictable. This introduces a tension in AI model training: models designed to minimize forecast errors tend to smooth out chaotic, unpredictable variation, whereas capturing chaos accurately implies accepting a level of uncertainty that limits forecast precision.
Environmental factors such as ocean temperature and wind shear set an upper bound on hurricane potential intensity. Warmer waters typically allow stronger storms but also contribute to more intense fluctuations in intensity, further complicating reliable forecasts.
Improving AI hurricane intensity predictions will require more than just greater data volumes, enhanced algorithms, or faster computing. Researchers emphasize the need for a deeper understanding of hurricane dynamics and the role of chaos in storm evolution. Future forecasting efforts may shift toward probabilistic approaches that convey a range of possible intensities and associated uncertainties rather than focusing on a single predicted value, potentially offering more actionable information for decision-makers and emergency response teams.
Chanh Kieu, an associate professor of atmospheric science at Indiana University, highlights these complexities as essential considerations for the next generation of hurricane forecasting models.