A geophysically constrained Transformer model for tropical cyclone track and intensity simulation over the western North Pacific and its contribution analysis

Tropical cyclones are among the most powerful and destructive forces on Earth, spinning up over warm ocean waters to unleash fierce winds and torrential rain upon coastal communities. Predicting exactly where these storms will go and how strong they will become is a constant struggle for meteorologists. Traditional methods rely on massive computer simulations that solve complex equations of physics, but these systems are slow, require immense computing power, and often struggle to update quickly enough as a storm evolves. In recent years, scientists have turned to artificial intelligence to speed up the process, training computer programs to recognize patterns in historical weather data. However, many of these digital models act like black boxes; they can make predictions, but they do not explain why they made them, and they sometimes ignore the fundamental physical laws that govern how a storm moves and intensifies.

A team of researchers from Beijing Normal University and Monash University has developed a new approach that bridges this gap. They created a sophisticated computer model that combines the pattern-recognition power of modern artificial intelligence with the hard rules of geophysics. This model, designed specifically for the western North Pacific—the most active region for tropical cyclones on the planet—learns from decades of historical storm records and atmospheric data. Unlike previous attempts that treated the storm as a simple sequence of numbers, this new system is "geophysically constrained," meaning it is guided by the actual physical relationships between the storm and its environment. The researchers tested this model against two other common types of prediction tools: one that learns from sequences of data over time and another that builds decision trees based on past examples. Their results show that by respecting the physical reality of the storm, their model can forecast the path and strength of a tropical cyclone with greater accuracy and stability, particularly in the critical first few days of a storm's life.

The researchers focused on the western North Pacific, a vast ocean basin where more than a third of the world's tropical cyclones form each year, frequently threatening countries in East Asia. To train their model, they fed it a massive archive of information, including the recorded paths and intensities of storms from 1949 to 2024, alongside detailed snapshots of the atmosphere and ocean conditions surrounding those storms. The model learned to look at the current state of a storm—its location, wind speed, and central pressure—and the surrounding environment, such as sea surface temperatures and wind patterns at different heights in the sky. It then used this information to predict where the storm would be and how strong it would be in the future, step by step, for up to 36 hours ahead.

When the team put their new model to the test, it outperformed the standard tools used for comparison. For short-term forecasts six hours into the future, the model predicted the storm's location with an average error of about 200 kilometers, a significant improvement over the other methods. Even when looking further ahead to 36 hours, the model maintained a high level of reliability, with an average error of roughly 357 kilometers. In terms of predicting intensity, which is measured by the speed of the strongest sustained winds, the model was equally impressive. Six hours ahead, it estimated wind speeds with an average error of just 2.59 meters per second, and at 36 hours, the error grew to 7.54 meters per second. These results were notably better than the other models, which made larger mistakes in both location and wind speed, especially in the crucial early hours of a forecast.

What makes this work particularly valuable is not just the accuracy, but the clarity of how the model reaches its conclusions. The researchers used a method to peel back the layers of the artificial intelligence and see which pieces of information mattered most. They found that the model relied heavily on two specific factors: the air pressure at the center of the storm and the speed of the winds. These two elements accounted for more than half of the model's decision-making power. This aligns perfectly with what meteorologists have long known: a storm's intensity is tightly linked to its internal pressure and wind structure. The model also considered larger atmospheric patterns and ocean temperatures, but these played a secondary role, acting as supporting actors rather than the main drivers. This finding suggests that the model is not just memorizing data but is actually learning the physical logic of how storms behave.

The team also tested their system on two specific, real-world storms to see how it handled complex situations. In one case, a storm moving steadily across the open ocean, the model tracked its path with high precision. In another instance, a storm that intensified rapidly near the coast, the model successfully identified the trend of strengthening, though it slightly underestimated the peak wind speed. These real-world tests confirmed that the model could handle diverse conditions without producing erratic or unrealistic jumps in its predictions. While the model is not perfect and still faces challenges with the most extreme and rapidly changing storms, it represents a significant step forward. By combining the speed of artificial intelligence with the reliability of physical science, this new tool offers a promising way to improve early warnings and help communities prepare for the powerful storms that threaten their shores.