AI accurately predicts monster waves

Monster waves were long regarded as a sailor's yarn: sudden, huge waves that can destroy entire ships. It was only when digital instruments first recorded a 26 metre high wave hitting the Norwegian oil drilling platform "Daupner" in 1995 that there was solid proof of these gigantic outliers. Since then, they have been extensively researched. Now a team led by computer scientist Dion Häfner from the Niels Bohr Institute in Copenhagen has developed a model with the help of a lot of data and machine learning that explains the formation of monster waves - and also indicates the probability of their occurrence. The work was published in the scientific journal "PNAS" on 20 November 2023.

A shipping accident that occurred in December 1978 illustrates just how dangerous the giant breakers can be. Back then, the German freighter "München" was caught in a storm in the North Atlantic. The crew remained relaxed - a storm was not supposed to harm the 261 metre long ship. But shortly after a distress call was made, the freighter and its 28-strong crew disappeared. Only a few lifeboats and shipping containers were later salvaged. At the time, the German Navy was unable to determine the cause of the sinking. Today, it is assumed that the freighter was the victim of a monster wave.

In order to protect shipping from such wave outbreaks, researchers are trying to understand how monster waves form - and which conditions favour them. To this end, Häfner and his colleagues have analysed freely available data on sea movements, swell and the corresponding water depths. They also collected data from buoys located at 158 different locations around the US coasts and in overseas territories. Taken together, the researchers had information on more than one billion waves at their disposal - which corresponds to the ocean movements of around 700 years.

Monster waves occur quite frequently
The scientists were surprised to discover that monster waves occur more frequently than expected. "We have registered 100,000 monster waves in our data set. That corresponds to about one such wave per day," explains computer scientist Johannes Gemmrich, co-author of the current study. However, not all of them were of extreme size (a wave with a height of more than 20 metres is considered "extreme"). In general, the researchers categorised a wave as a monster wave if it was at least twice as high as the usual swell.

Häfner and his colleagues used several AI methods to analyse the huge amount of data. While conventional AI systems simply spit out results without justifying them, the researchers were interested in causal relationships. They wanted to use machine learning to find out what causes monster waves - and not just calculate a probability for their occurrence without knowing how the AI model arrives at the result. They therefore developed a so-called "causal" model, the results of which can be checked and interpreted. The algorithm and the data collected by the researchers are freely available and can be used by anyone.

In fact, the results of Häfner and his colleagues break with a common explanation of monster waves. Some experts assumed that the outliers occur when several waves combine for a short time and one wave absorbs all the energy - and travels on as a monster wave. However, as Häfner's team discovered, the predominant factor appears to be what is known as "linear superposition": If wave crests propagate at different speeds, it can always happen that many wave crests overlap at one point and amplify each other to such an extent that a monster wave is created.

The algorithm developed by the researchers aims to predict when the perfect conditions prevail for such a linear overlap to occur. "Basically, it's just bad luck when one of these giant waves occurs. They are caused by a combination of many factors," says Häfner. So let's hope that the researchers' algorithm works - shipping companies could already be testing it.

Spectrum of science

We are partners of Spektrum der Wissenschaft and want to make well-founded information more accessible to you. Follow Spektrum der Wissenschaft if you like the articles.