A satellite engineer with no background in oceanography has won an international competition to predict where a “message in a bottle” will drift to on the open sea.
Chris Wasson, who is based in southern California, beat 31 teams to scoop the $25,000 top prize in the Forecasting Floats in Turbulence Challenge, organised by the US Defense Advanced Research Projects Agency (DARPA).
“I don’t have any real background in oceanography or weather system forecasting, but the problem was stated as more data- and algorithms-oriented,” says Wasson. “So I registered thinking I’d just see how things went.”
Advertisement
Competitors had to forecast the locations over 10 days of 90 devices drifting in the Atlantic. They were given the previous 20 days of movement and meteorological data on currents, wind and waves.
Wasson won against teams of experienced oceanographers. Second prize went to Deltares, a research institute based in the Netherlands, and third place to the Center for Ocean-Atmospheric Prediction Studies in Tallahassee, Florida.
Wasson modelled the combined effect of wind and surface currents on each float. For the initial 20 days, he compared predictions with the actual position and fine-tuned his model each time, using a combination of machine learning, which is a form of artificial intelligence, and mathematical modelling.
While more advanced machine-learning approaches are fashionable, these may “over-fit” and work well on the training data without generalising to subsequent data. Such systems can also be quite opaque: it is impossible to tell how they are working and whether their workings are realistic. Wasson found a simpler system was easier to understand and the combination with analytical modelling worked well.
“Machine-learning approaches may suggest non-obvious solutions to problems and analytical methods can help to validate and explain those results,” says Wasson.
Wasson located one float to within 4 kilometres of its actual position after 10 days, an achievement only matched by one other team.
He found that ocean currents were the dominant factor, though strong winds could override them. While his model was generally quite accurate, some drifters took off in unexpected directions, possibly because of inaccurate ocean current data.
While he had a reasonable idea how his model would perform, Wasson didn’t know how strong the competition was. “Some other teams also put up some very high-scoring days during the competition, so it was a real nail-biter,” he says.
at the University of Victoria in Canada says technical specialists from outside oceanography have often contributed to the field, so Wasson’s win wasn’t a complete surprise.
Identifying why Wasson’s predictions sometimes went so wrong isn’t straightforward. “That could be a reflection of the chaotic system, or a problem with the data, or a problem with the model – any of the above,” says Aksamit.
DARPA hopes the challenge will enhance predictions for its Ocean of Things (OoT) project, an array of thousands of free-floating sensors that will be deployed this spring to detect ships and submarines. DARPA will make OoT data publicly available to help future drifter models. Drift forecast techniques could be applied to oil slicks, the dispersion of fish larvae and locating shipwreck survivors.
“I would love to work more on ways that the drifter data might be used to validate existing ocean current and wind current forecast models,” says Wasson, but he notes that as he doesn’t work in this field, others are more likely to make the next steps. “I will definitely be on the lookout for future DARPA challenges.”
