From a mountain of vague signals, AI can deduce that the Earth is becoming dangerous

Earthquakes and volcanic eruptions come without warning – or so it seems. Subtle announcements of such natural disasters are often hidden in complex patterns of many different signals, which we humans do not easily recognize. The intelligence of computers can help to discover these patterns and thus make better predictions.

According to British volcanologist Andrew Hooper, affiliated with Leeds University, it starts with “data, data and more data”. Hooper leads the project Deepvolc that with artificial intelligence (artificial intelligenceor AI) monitors and learns to recognize the deformation of the earth before volcanoes erupt.

That sounds vague, Hooper admits, but he explains: “Magma is molten rock that moves in the mantle or lower crust due to pressure differences in the mantle. When the pressure of the magma becomes too great, it seeks a way up, resulting in a volcanic eruption at the Earth’s surface. Rising magma causes the earth’s soil to move. These movements are a harbinger of an eruption and we try to predict them with AI.”

The model learns from historical patterns so that it can predict the course of current soil deformations

Andrew Hooper volcanologist

Locally, according to Hooper, it is already possible to predict eruptions a few days in advance by other omens: gases released from a steaming crater, or by sounding sound waves from the soil. “However, to measure those clues for each individual volcano, you need an awful lot of equipment,” he says. “Unfortunately, more than half of the 1,400 active volcanoes have no measuring instruments and, in addition, gases and strong seismic waves can only be measured on Earth shortly before an eruption.” Predicting earlier helps to warn people earlier about a possible eruption.

“Predicting the deformation of the soil can play a key role in this,” says Hooper. “It’s just not that easy.” A telling example is the spectacular eruption of Iceland’s Fagradalsfjall in 2021. A study published in Nature, on which Hooper collaborated, accurately imaged the soil deformation prior to the eruption. One of the results: based on one volcano, it is actually impossible to recognize patterns in that deformation, so the precise moment of eruption cannot be predicted.

Accurate to millimeters

For the research in Iceland, Hooper and his team used data from the European satellite Sentinel-1, which visits every volcano on Earth twice every twelve days. With the radar of this satellite, soil deformation is measured with millimeter precision. Hooper: “If we see a vertical deformation of the soil over time, that is an omen for a possible eruption.” Every year about sixty volcanoes erupt, preceded by deformation of the soil. To find patterns in this, Hooper uses thousands of historical soil deformations since the 1990s for his AI model. The satellite is adding more and more current data to this. “1,500 active volcanoes, visiting twice in twelve days and processing all the data…”, Hooper sighs. “That is really impossible for a person to do, but it is for computers.”

With all this data, Hooper trained it AI model Icasar that learns to recognize patterns in how soil deformations develop prior to an eruption. The model looks at so-called interferograms, an image consisting of a series of satellite images that show the difference in height of the earth’s surface. “To us, these images look like simple photos with colored contours, but when we feed the AI ​​model with thousands of those images, it recognizes all kinds of patterns in the frequency and intensity of the bottom deformations,” says Hooper. “The model learns from historical patterns, allowing it to predict the course of current soil deformations.”

Hidden seismic symphony

Hooper tested the model in 2019 for the Sierra Negra volcano in the Galapagos Islands. The model simulated the monitored soil deformation during 3.5 years prior to an eruption in 2018. Based on historical data from other volcanoes, the model produced almost identical deformations at Sierra Negra, Hooper concludes in a publication in Advancing Earth and Space Science. The AI ​​model is not yet finished, but according to Hooper it can therefore be tested and eventually applied to all active volcanoes. For example, an eruption can be predicted earlier because people are no longer dependent on instruments around individual volcanoes. He cites an example of the Mount Edgecumbe volcano in Alaska. Here in 2022, seismic waves will be measured with measuring instruments measured from the ground. “When we checked our model, it predicted soil deformation since 2018 based on historical patterns in other places,” says Hooper. “The current deformation of the soil observed by the satellite matched the predicted deformation of our model in intensity and timing.”

We have always seen that as empty noise, but the computer thought otherwise

Paul Johnson seismic geophysicist

AI also comes to the rescue when predicting earthquakes, says American Paul Johnson, seismic geophysicist at the Los Alamos National Laboratory in the United States.

For Johnson, his career in the earthquake lab started with an installation that simulates earthquakes. “Earthquakes occur on fault lines when plates in the Earth’s crust collide, rub, or move apart,” explains Johnson. “The force first builds up and when it is large enough, tectonic plates abruptly overshoot into the slip zone.” Johnson calls the resulting quake the “time of failure”, the end of a so-called quake cycle. Predictions are now made the day or minutes before an earthquake. However, during the earthquake cycle, which can last thousands of years, there are already all kinds of hidden signals in sound waves from the earth’s surface, the research team discovered. machine learningmodel by Johnson about which he published in 2019 Nature.

Simulated earthquakes in the lab

You can compare this process to sounds you hear on the street outside a football stadium. Without following the game, you can hear what is happening on the field through cheers and jeers. It Johnson’s model registers such signals in the soil. When the model has registered a lot of bubbling soils, the model learns to recognize and eventually predict patterns.

“There is always noise around fault lines,” says Johnson. “We have invariably seen that as empty noise, but the computer thought otherwise.” The applied machine learning recognized signals in that seismic symphony during one earthquake cycle that Johnson could physically translate as indicators of force build-up between tectonic plates, energy distribution along the fault line, and small differences in frequencies of sound waves from the soil. Because these signals were discovered based on patterns in thousands of simulated quakes in the lab, Johnson says they would never have been discovered with a purely human observation.

In the real world, silent earthquakes only lead to a major earthquake after many cycles

Paul Johnson seismic geophysicist

Outside the lab, Johnson used the model to examine historical noise from New Zealand, Chile and the Cascadia subduction zone stretching from Canada to northern California. In the latter region, two tectonic plates slide under each other where, according to Johnson, silent, imperceptible earthquakes occur every thirteen to fourteen months. Johnson: “Here we discovered the same signals in the noise as during the simulated quake cycle in the lab. The computer also found predictable patterns in all those signals prior to an earthquake.”

According to Johnson, this discovery shows that the timing of earthquakes is not random. According to him, this would mean that patterns can also be predicted with a trained model. Johnson used one for this deep learningmodel that works with so-called artificial neural networks consisting of a large number of neurons that receive and transmit signals to each other. “Based on thousands of signals found, we wanted the computer to figure out how they would evaluate and eventually lead to an earthquake.” He discovered that the model can provide an estimate of when the quake will actually occur throughout the entire quake cycle is going to take place. The closer you get to the moment of the quake, the more accurate the model becomes. Johnson’s team retroactively tested the model to predict more than 60 silent earthquakes in Hawaii in 2018. The monitored data was compared to the model’s results. More than 90 percent of quakes were predicted correctly.

Vegetation and water vapor

As promising as this sounds, according to Johnson there are still snags to AI and earthquake prediction. “Silent earthquakes in the real world only lead to a large quake with a tangible impact after many cycles,” he says. “We don’t yet know if the model can predict beyond a single earthquake cycle.” In addition, except on historical data, the model has not yet been tested on current active fault lines.

There are also bumps in the road for Andrew Hooper, the volcanologist. The Sentinel-1 satellite’s radars emit only short-wave electromagnetic radiation. When a volcano is covered with vegetation, it reduces the accuracy in measuring soil deformation because those waves are reflected by plants. In January 2024 launched a new satellite who will work with long-wave radars for Hooper’s project: “That can help us to better understand volcanoes in these areas as well.”