Storm surge forecasts from a global tide-gauge network

Coastal flooding is a complex threat, but its most deadly component is often the storm surge: a wall of water pushed ashore by the wind and pressure of a storm. Unlike a tsunami, which is triggered by an earthquake, a surge is a direct response to the weather. For communities living on the edge of the ocean, knowing when this wall of water will arrive and how high it will rise is a matter of life and death. Yet, the ability to predict these events is not shared equally. While wealthy nations with dedicated funding and advanced computer models can issue reliable warnings, vast stretches of the world's coastline, particularly in developing countries and small island states, remain without any local forecast guidance. The gap between those who have warnings and those who do not leaves millions vulnerable to a hazard that is projected to grow worse as sea levels rise.

For decades, scientists have relied on massive, physics-based computer simulations to model these surges. These models attempt to recreate the entire ocean and atmosphere, calculating how wind and pressure interact with the shape of the seafloor. While powerful, these simulations are computationally expensive and require constant, high-quality data to run. As a result, they often cannot provide the specific, local guidance needed for every coastline on Earth. In recent years, a different approach has emerged: using artificial intelligence to learn patterns from historical data. However, most of these previous attempts were limited. They could only predict surges for the specific locations where they had been trained, or they relied on synthetic data that didn't reflect the messy reality of the ocean. The challenge remained: could a single model learn from the entire world's network of tide gauges and then successfully predict surges at places it had never seen before?

In a new study, researchers Heshan Chen and Andrew Kruczkiewicz set out to answer this question by building a model trained on a global network of 760 tide gauges. They did not just train the model to recognize patterns at these specific locations; they tested its ability to forecast surges at 84 different gauges and in entire regions that contributed absolutely no data to the training process. This is known as a "zero-shot" test, a rigorous standard in prediction science that asks whether a system can generalize its knowledge to entirely new environments. The results were striking. The model successfully forecasted surges at 81 out of the 84 test gauges, beating a simple method called "persistence"—which assumes the water level will stay exactly as it is right now—at nearly every location. This included gauges on every inhabited continent, from Africa to South America, proving that the model had learned the underlying physics of how storms push water, rather than just memorizing local history.

The researchers found that the model's success depended heavily on the diversity of the data it was fed. It was not enough to simply have a large number of training gauges; the gauges had to represent different types of coastlines and storm behaviors. When the team tested the model on regions it had never seen, such as Europe or Japan, the model performed well, but its accuracy was tied to how well those specific coastal environments were represented in the training set. For instance, the model struggled slightly more with the unique wind-driven surges of the North Sea because that specific type of coast was less common in the training data. However, when they increased the diversity of the training set by adding gauges from different parts of the world, the model's ability to predict surges in these unfamiliar regions improved significantly. This suggests that the key to global prediction is not just more data, but data that captures the full variety of how the ocean reacts to the weather.

A critical part of the study involved testing the model under real-world conditions. While the initial tests used perfect historical weather data to see what the model was capable of, the researchers also ran it using actual weather forecasts, which contain errors and uncertainties. Even with these imperfect inputs, the model retained four-fifths of its skill, still outperforming the simple persistence method at 72 out of 79 gauges. This is a vital finding for operational use, as it proves the system can work with the imperfect data that real-world forecasters actually have. The model also demonstrated that having a live feed of local water levels is the single most important piece of information. When the model was given the current water level at a specific gauge, its accuracy jumped dramatically. Without this local observation, the model still performed better than random chance, but it could not match the precision of a system that knows exactly where the water stands at the moment the forecast begins.

The study also compared this new artificial intelligence approach against the best available global physics-based model, which is currently the gold standard for surge forecasting. In a head-to-head test where both systems were given the same local water level observation, the AI model was more accurate for general conditions at an eight-hour lead time. During the most intense storms, the two systems performed similarly, with neither clearly dominating the other. However, the physics-based model still held an advantage in specific environments, such as the semi-enclosed Baltic Sea, where the surge is driven by wind patterns across the entire basin rather than just at a single point. This highlights that while the AI model is a powerful new tool, it does not replace the need for physics-based models in every situation. Instead, the two approaches are complementary: the AI model offers a way to bring high-quality, local forecasts to the many coastlines that currently have none, while physics models remain essential for complex, large-scale ocean dynamics.

Perhaps the most practical aspect of this work is the model's ability to provide not just a single prediction, but a range of possibilities. The researchers trained the model to output a "probability envelope," which tells forecasters the likelihood that the water will reach certain heights. While the model's single-point prediction sometimes underestimated the absolute highest peaks of the most extreme storms, the probability envelope was well-calibrated. It successfully covered the vast majority of hours during a storm, giving emergency managers a reliable sense of the risk. Although the envelope did not capture the absolute highest peak in every single extreme event, it provided a much better picture of the danger than the simple assumption that the water would stay at its current level. This probabilistic information is exactly what is needed to make informed decisions about evacuations and resource allocation.

The implications of this research extend far beyond the technical details of the model. For the first time, there is a proven method to generate skilled, local storm surge forecasts for any coastline in the world, provided there is a tide gauge to provide a starting point. This capability could fundamentally change how early warning systems are built, particularly in the least developed countries and small island states that have historically been left out of the global forecasting network. The researchers emphasize that having a forecast is only the first step; turning that forecast into a life-saving warning requires local governance, funding, and the capacity to act. However, by removing the technical barrier of needing a local, physics-based model for every single coast, this work opens the door for a more equitable distribution of safety. The study concludes that the path forward involves continuing to expand the global network of tide gauges and refining these models to cover the remaining gaps in coastal diversity, ensuring that the next generation of warnings reaches everyone, regardless of where they live.