From regional to global: a typhoon-aware adaptive graph transformer for cross-basin tropical cyclone-induced storm surge forecasting
This paper introduces ST-GNNFormer SurgeCast, a globally adaptive graph transformer framework that successfully unifies tropical cyclone-induced storm surge forecasting across all ocean basins into a single model, demonstrating robust generalization in data-sparse regions and achieving high predictive accuracy on both reanalysis and independent in-situ observational data.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Every year, tropical cyclones sweep across the world's oceans, gathering strength and speed before crashing into coastlines. When these storms hit land, they do not just bring wind and rain; they push massive walls of seawater onto the shore, creating storm surges that can drown low-lying communities and destroy infrastructure. For decades, scientists have tried to predict exactly how high these waters will rise. The traditional way involves building complex computer models that simulate the physics of the ocean and the atmosphere. While accurate, these models are incredibly heavy to run, requiring supercomputers and hours of calculation time, which makes them difficult to use for rapid, widespread warnings. In recent years, researchers have turned to artificial intelligence to find a faster path, teaching computers to recognize patterns in past storms and predict future surges. However, most of these smart models have a significant flaw: they are trained to work only in specific places, like the Gulf of Mexico or the coast of Japan. If a storm hits a region where the model was never trained, or where data is scarce, the system often fails, leaving those vulnerable areas without reliable forecasts.
A team of researchers has now developed a new approach that breaks this regional barrier. They created a single, unified artificial intelligence system capable of predicting storm surges anywhere in the world, from the typhoon-prone waters of the Western Pacific to the cyclone tracks of the South Pacific. The system, named ST-GNNFormer SurgeCast, was trained on a massive global dataset covering twenty-six years of storm history, from 1993 to 2018. Instead of learning the unique geography of just one bay, the model learned to understand the universal rules of how storms interact with coastlines. It does this by treating the world's coastlines as a connected network of observation points, or tide gauges, rather than isolated locations. By analyzing how water levels change at one station in relation to others, and how those changes lag behind the wind and pressure of a passing storm, the model can infer what is happening even in places where it has never seen a storm before.
The core of this new system is a clever mechanism that allows it to adapt to every single storm it encounters. When a tropical cyclone forms, the model does not just look at the storm's location; it dynamically reshapes its internal understanding of which coastal stations are most relevant. It recognizes that a storm in one part of the ocean might affect a specific group of tide gauges, while a storm in a different basin affects a completely different set. This flexibility allows the model to focus its attention on the areas that matter most for that specific event, effectively learning to "zoom in" on the danger zone without needing to be retrained for a new region. The researchers tested this global model on thousands of storm events across five different ocean basins. The results showed that the system could predict water levels with remarkable accuracy, achieving an average error of less than five centimeters across the entire globe. This level of precision held true even in data-sparse regions like the North Indian Ocean and the South Pacific, where traditional models often struggle due to a lack of local training data.
To ensure the model was not just memorizing the data it was trained on, the researchers compared its predictions against real-world measurements from tide gauges that were not part of the training set. They found that the artificial intelligence performed nearly as well as the most advanced physics-based numerical models, but with a fraction of the computational cost. The system also proved capable of looking further into the future, providing reliable forecasts up to forty-eight hours ahead, though the accuracy naturally decreased slightly as the forecast window extended. In specific case studies of catastrophic storms, such as Super Typhoon Megi in 2010, the model successfully captured the extreme peaks of the surge that often catch emergency planners off guard. It managed to predict the timing and height of the water with a high degree of confidence, outperforming simpler models that only looked at space or only looked at time.
Despite these successes, the researchers are careful to note the boundaries of their work. The model is designed specifically for tropical cyclones and does not yet apply to the winter storms that batter the coasts of Europe and North America, which are driven by different atmospheric forces. Additionally, while the model handles broad geographic patterns well, it sometimes struggles in areas with extremely complex coastlines, such as the intricate island chains of the Zhejiang and Fujian provinces in China, where local geography plays a massive role in how water behaves. Nevertheless, this work represents a significant shift in how we approach disaster forecasting. By moving from a collection of isolated, regional tools to a single, globally adaptive system, the researchers have demonstrated that artificial intelligence can learn the universal language of storms. This opens the door to providing high-quality, life-saving warnings to coastal communities around the world, including those in developing nations that have historically lacked the resources to build their own sophisticated forecasting systems.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.