Trust the machine: artificial intelligence may be the answer to creating rapid warning systems for volcanic eruptions

Cosmos Magazine

Cosmos

Cosmos is a quarterly science magazine. We aim to inspire curiosity in ‘The Science of Everything’ and make the world of science accessible to everyone.

By Cosmos

In 2019 a volcano on Whakaari/White Island in New Zealand erupted, killing 22 people.

The eruption spurred volcano researchers Dr Alberto Ardid and Dr David Dempsey, of the University of Canterbury, NZ, to see if they could help reduce the risk of people being caught out by a volcano ever again. In a paper published overnight in the journal Nature Communications, the pair propose a new type of rapid warning system that utilises machine learning.

How to know when it will blow?

Ardid compared over 40 years of seismic data from six volcanoes, three in New Zealand and three in Alaska. He wanted to see if there was a common signal shared by the volcanoes before they blew.

“New Zealand volcanoes all share a common signal that occur before eruptions,” Ardid says. “We see common patterns in all of the five past Whakaari eruptions in the last 12 years, we also see them to some extent, in the Ruapehu eruptions in 2006 and 2007 and also in Tongariro in 2012.”

volcano
New Zealand’s White Island volcano erupted 9 December 2019. Credit: Stone / Getty

Before each of these eruptions a seal was forming on top of the volcano allowing pressure to build, eventually leading to an eruption. The artificial intelligence that Ardid used was able to identify when this process was happening.

Despite getting a clear signal for the New Zealand volcanoes, their Alaskan counterparts were more unpredictable.

Ardid explains: “Volcanoes in New Zealand have a well-developed hydrothermal system beneath the ground that is beneath the crater – we call them wet volcanoes. And this method seems to be working really well for them.”

Warning signs of an eruption started to show up around three weeks before the event with the strongest signal occurring a few days before, explains Dempsey.

“What we haven’t done is looked at the question yet of how far in advance is the window of increased concern of eruption.”

How was this pattern found?

Dempsey explains that the machine learning algorithm is very effective when you don’t know what patterns you’re looking for in the data, but it suspects a pattern is there.

“It becomes a brute force technique, where it [the artificial intelligence] has this huge library of possible patterns and it just starts crunching through all of them, checking 40 years’ worth of data,” Dempsey says.

You’d need hundreds of scientists to go through it if you were going to do it by hand.

Will we have a new rapid alert system for volcanoes soon?

A trial system looking for these patterns is already up and running at Whakaari, the volcano that tragically erupted in 2019.

“You want to have some confidence in these systems before you put them into use, but I’d say the ability to use these is already here,” Dempsey says.

Volcano
On 3 June 2018, Guatemala’s Fuego Volcano erupted. Credit: iStock / Getty

How this method would be folded into existing warning systems is yet to be determined.

Ardid is now setting his sights on volcanoes around the world, to see if detecting for this pattern could be used to provide warning of wet volcano eruptions globally.

“We are trying to locate if this pattern is happening before eruptions overseas, we have been finding that they have happened in volcanoes in Indonesia, Iceland and Guatemala,” he says. “We just recently started researching a volcano on the border of Chile and Argentina in South America where we also found some pretty clear signal.”

Please login to favourite this article.