Why AI has trouble predicting the fury of hurricane intensity
Hurricane Irma rapidly intensified before hitting the Caribbean and Florida in 2017, but intensity is still hard to forecast. NASA Artificial intelligence has revolutionized weather forecasting in just a few years, with global AI weather models now able to produce forecasts that rival some of the world’s best physics-based prediction systems.
Artificial intelligence (AI) has significantly improved weather forecasting in recent years, with global AI models now matching some of the best traditional physics-based prediction systems. This progress is driven by three key factors: vast amounts of weather data, advanced AI models, and powerful computing resources. Global weather data, collected over decades, provides millions of examples of atmospheric changes, enabling AI to identify patterns that were previously impossible to detect. For instance, datasets like those from the National Oceanic and Atmospheric Administration (NOAA) contain decades of climate records covering the entire Earth. However, this success is more pronounced at the global scale. When focusing on smaller regions, such as hurricane intensity forecasting, AI faces greater challenges due to the need for more detailed and localized data.
Hurricane intensity forecasting is a regional problem, requiring detailed data about extreme weather events like heavy rainfall, thunderstorms, or hurricanes. These events often develop rapidly or move quickly, making them difficult to capture in AI models. Current global datasets lack the granularity needed for accurate regional predictions. Scientists typically rely on two types of data to train AI models for hurricane intensity: direct observations and weather model simulations. Observations include measurements from weather stations, radars, buoys, and satellites, but these are often limited to coastal areas and unevenly distributed. Satellites can help fill some gaps, but they cannot provide a complete three-dimensional view of every hurricane globally. Weather model simulations offer a more complete picture but include approximations and uncertainties due to incomplete knowledge of atmospheric physics.
AI models struggle with hurricane intensity forecasting because they lack a full, three-dimensional dataset to train on. Direct observations, such as those from NOAA’s Atlantic Oceanographic and Meteorological Laboratory, provide detailed but sparse data, often missing critical stages of hurricane development over the open ocean. Satellites, like those managed by NASA, can estimate rainfall, wind speeds, or cloud temperatures but cannot scan the entire structure of a hurricane simultaneously. For example, satellite data in 2016 provided partial insights into Hurricane Matthew’s intensity as it approached Florida, but this was not enough for comprehensive forecasting. Weather model simulations, while more complete, still miss fine-scale processes due to computational and knowledge limitations, leaving gaps in the data needed for AI training.
Even with perfect data, AI models may struggle to predict hurricane intensity due to the chaotic nature of these storms. Tiny differences in initial conditions can lead to vastly different outcomes over time, a phenomenon known as the butterfly effect. Research suggests that hurricanes may contain an element of chaos that prevents AI from making perfect intensity predictions, even if all storm data were measured perfectly. For instance, a tropical storm can intensify toward a maximum possible strength, called its potential intensity, which depends on environmental factors like ocean temperature or wind shear. Warmer oceans increase this potential intensity, but small disturbances can cause unpredictable fluctuations. Studies indicate that these fluctuations may follow a chaotic attractor, a set of possible storm states that evolve unpredictably.
AI models face a fundamental dilemma when predicting hurricane intensity: they must balance accuracy with capturing chaos. On one hand, AI models aim to minimize forecast errors by learning patterns from data. On the other hand, they need to account for the inherent unpredictability of hurricanes. If an AI model captures this chaos, its error cannot be reduced indefinitely, as it must account for unpredictable fluctuations. For example, an AI trained to minimize error might learn the most likely hurricane evolution while smoothing out unpredictable changes. This competition between goals limits the accuracy of hurricane intensity forecasts, which worsen after just a few days. The National Hurricane Center notes that forecast accuracy declines significantly beyond this timeframe.
Improving AI’s ability to predict hurricane intensity requires more than just better data, neural networks, or computers. It also demands a deeper understanding of hurricane behaviors and how chaos emerges in their intensity. Scientists must distinguish between predictable and unpredictable aspects of hurricanes to train AI models effectively. This distinction will define the next generation of forecasting systems, which should focus on providing a range of possible hurricane intensities and their probabilities rather than a single predicted value. For instance, future systems might offer probabilities of a hurricane reaching Category 3, 4, or 5 intensity, rather than predicting a single category. This approach aligns with the inherent uncertainties in hurricane behavior and could improve preparedness and response efforts.
