AI Gets Better at Global Weather Forecasting, but Struggles to Predict Tropical Storm Explosions

The ability of artificial intelligence to forecast global weather has surged in recent years, even rivaling the best physics-based prediction systems. However, when asked to predict tropical storm intensity, AI models still show fundamental limitations that cannot be solved simply by adding more data or computing speed.

This gap has come into focus after several storms underwent explosive intensification within hours. Hurricane Polo, for instance, went from an ordinary tropical storm to a Category 5 hurricane within 24 hours in September 2026, with wind speeds reaching 180 mph. A similar event occurred with Hurricane Michael in 2018, which intensified drastically before slamming into Florida, leaving communities with minimal evacuation time.

The difference in scale is the root of the problem. Global weather AI is trained on historical climate and weather data covering the entire Earth, allowing models to recognize large-scale atmospheric patterns. However, storm intensity is a regional phenomenon involving rapid and extreme processes, requiring far more detailed data than is currently available.

In training AI for storm intensity prediction, scientists typically rely on two data sources. First, direct observations from weather stations, radar, buoys, and satellites. This data is detailed in coastal areas but sparse over the open ocean where storms grow. Modern satellites do help, but they cannot simultaneously scan the three-dimensional structure of storms worldwide.

Second, weather model simulations that combine atmospheric conditions and physical laws to produce high-resolution three-dimensional pictures. Unfortunately, these simulations are also imperfect because they contain approximations and uncertainties due to incomplete understanding of Earth's atmosphere. As a result, there is no truly adequate three-dimensional dataset to train AI to predict storm intensity.

However, data is not the only obstacle. Even if scientists could measure every storm every second for thousands of storms, perfect prediction would still be difficult to achieve. The cause is chaos: small differences in a storm's initial conditions can rapidly grow over time. Recent research by Chanh Kieu, an associate professor of atmospheric science at Indiana University, and his colleagues shows that storms contain chaotic elements that limit long-term prediction accuracy.

In a supportive environment, a tropical storm can intensify toward a maximum limit called potential intensity, which is determined by sea surface temperature, wind shear, and other environmental factors. The warmer the ocean, the greater the intensity fluctuations. Recent studies propose that these fluctuations are not entirely random, but occur within what is called a chaotic attractor—a set of storm states that allow for unexpected evolution.

"AI models trained to minimize prediction error may actually smooth out unexpected fluctuations, thereby losing important chaotic aspects," Kieu wrote in his analysis.

There is a fundamental dilemma: on one hand, AI wants to minimize prediction error; on the other, AI must also capture the intrinsic chaos of storms. These two goals conflict. If AI captures chaos, error cannot be suppressed indefinitely. If AI only pursues accuracy, it will ignore fluctuations that are crucial for early warning.

For Indonesia, the implications of this limitation are very real. As an archipelago surrounded by warm waters, Indonesia is vulnerable to tropical cyclones that can develop rapidly, although their frequency is relatively low compared to other regions. The Meteorology, Climatology, and Geophysics Agency (BMKG) has adopted various numerical models, but AI integration for storm intensity prediction is still in its early stages. Limited observational data over the open ocean and the chaotic nature of storms pose particular challenges for weather forecasters in the country.

Going forward, improving prediction is not just about adding data, building more sophisticated neural networks, or speeding up computers. Equally important is understanding storm behavior and how chaos emerges. The evaluation approach for next-generation prediction systems also needs to shift from a single intensity number to a probability range of possible intensities. The question is, will large investments in weather AI be enough to overcome the fundamental limit of chaos, or do we need a new paradigm that combines physics and machine learning more deeply?