Register for an account

X

Enter your name and email address below.

Your email address is used to log in and will not be shared or sold. Read our privacy policy.

X

Website access code

Enter your access code into the form field below.

If you are a Zinio, Nook, Kindle, Apple, or Google Play subscriber, you can enter your website access code to gain subscriber access. Your website access code is located in the upper right corner of the Table of Contents page of your digital edition.

Planet Earth

Is Machine Learning Getting Us Closer to Predicting Eruptions?

The 2019 tragedy at New Zealand White Island might have been prevented if there was more warning. Could machine learning allow for better, true prediction of volcanic eruptions?

Rocky Planet iconRocky PlanetBy Erik KlemettiJuly 21, 2020 2:55 PM
White Island
White Island/Whakaari in New Zealand. Credit: David Broad, Wikimedia Commons.

Newsletter

Sign up for our email newsletter for the latest science news

When Whakaari (White Island) in New Zealand unexpectedly erupted in December 2019, more than 40 tourists found themselves trapped on a small island that was exploding. The hot gases and water, flying rocks and ash killed 21 people during that eruption. This tragedy was a wake-up call for tour operators who would regularly bring people to this restless volcano in the Bay of Plenty. It is a volcano that produces steam-driven explosions that come with little warning, and it is these types of blasts that have killed dozens of people on volcanoes around the world over the past decade.

Part of the problem is how we think about volcanic danger. What people want is a prediction — when exactly will the volcano erupt? The volcanological community isn't in the game of prediction because we just don't know enough about what exactly triggers an eruption to be that precise.

Instead, volcano monitoring relies on forecasts that capture a probability of eruption. Seeing many earthquakes and increasing gas emissions from a volcano? We can say that the likelihood of an eruption has increased. Maybe we can even say it could happen in the next month. However, it is not going to be "tomorrow at noon," just like a weather forecast.

These forecasts and probability models lead to volcanic alert levels. The scientists monitoring volcanoes and emergency managers need to communicate the risk, so most countries use some volcanic alert system to show how potentially dangerous a volcano is at a given moment. Alert Level 1? The chances of an eruption are low. Alert Level 3? The chances are very high for an eruption.

Both forecasts and alert levels are defined by people looking at all the data coming in from a volcano — earthquakes, gas emissions, deformation, thermal and more — and interpreting what all the signs might mean. Usually, a probability tree (or decision tree) is used. These branching flow charts allow for taking the signs at the volcano and its past activity to produce a probability for specific events.

Unrest at a volcano? Based on its past activity and the type of unrest, it might be 85 percent chance that nothing happens, 15 percent chance of an eruption. In that 15 percent chance, it might be 95 percent chance of a small eruption and 5 percent chance of something larger.

With people involved, these decisions of alert level can take time. A new study in Nature Communications by Dempsey and others has developed a machine learning approach to forecasting volcanic eruptions and setting alert levels at volcanoes. They examined nine years' worth of data from Whakaari to train a computer to look for the precursory signs of volcanic blasts — especially the steam-driven eruptions whose precursors can be subtle — and then developed criteria for the computer to decide if an alert is needed.

Dempsey and his colleagues set limits on when the computer would call for an increased alert. If 80 out of 100 outcomes in the probability tree led to an eruption, then an alert was needed. With that threshold, almost all the eruptions over the past nine years were caught. The December 2019 eruption produced an alert four hours before the blast, which could have been enough time to get tourists away from danger. Only a purely hydrothermal eruption in 2016 was missed and this might have been because it was unlike other blasts at Whakaari (more on this in a bit).

Using this method, alert status can be changed faster as signs change. The data can be reinterpreted on a minute-by-minute scale. As the authors point out, this also takes out the potential bias caused by personal, political or economic influences on setting alerts.

Yet, there are problems. The missed eruption in 2013 is an example of how the computer can't identify potential eruptions if the precursors have not been seen before — remember, it "learns" based on the data it is fed. Dempsey and colleagues point out that it still needs a review by people to watch for such unique events.

The biggest question, though, ends up outside the computations. Who has access to this information? When do emergency managers or scientists publicize the changes? Without clear guides to how the computer forecasts can be integrated with effective management plans, the value is diminished.

However, this is all a big step forward. With these unexpected blasts like what happened at Whakaari in 2019 or Ontake in 2014, warning even a few hours beforehand could save many lives. We aren't yet "predicting" an eruption, but we are getting more refined in interpreting the signs of impending explosions.

Correction: An earlier edition of this article mistaking listed the October 2013 eruption as one that was missed by the computer analysis. It was the eruption in 2016.

2 Free Articles Left

Want it all? Get unlimited access when you subscribe.

Subscribe

Already a subscriber? Register or Log In

Want unlimited access?

Subscribe today and save 70%

Subscribe

Already a subscriber? Register or Log In