Skip to main content
Earth sciences

Earth sciences

Artificial intelligence spots unusual feature at Earth’s core-mantle boundary

19 Jun 2020
Seismic discovery
Inside out: illustration showing the locations of earthquakes (orange stars); seismometers (blue triangles); and hotspots (yellow triangles). (Courtesy: Doyeon Kim)

An artificial intelligence algorithm originally developed for astrophysics has revealed a previously unknown feature on the core-mantle boundary deep inside the Earth. The algorithm allowed researchers in the US to find patterns in seemingly unconnected seismic data from different sources and the team believes that the new technique could glean insights from other types of geophysical datasets in future.

The motion of tectonic plates continues to shape the surface of the Earth and this activity is monitored constantly by a worldwide network of seismometers. Studying how seismic signals travel through the different layers of the Earth has informed much of our knowledge of the structure of Earth’s interior. In the early 20th century, for example, researchers inferred that the Earth has a liquid outer core because shear waves from an earthquake were not detected by seismometers on the opposite side of the globe – because a liquid cannot shear.

As seismology has grown more sophisticated and more monitoring stations have been installed, researchers have discovered much more about Earth’s interior, such as physical stratification and plumes of rising hot material within the mantle and mysterious “ultra-low velocity zones” on the core-mantle boundary, where waves travel much more slowly than expected. “They’re inferred from the data but we don’t really know what these things are,” says seismologist Doyeon Kim of University of Maryland, College Park, who led the latest research.

Ever-growing volume of data

Researchers normally make such deductions from patterns of seismological signals that clearly stand out to a human observer. However, the ever-growing volume of seismic data from across the world may contain patterns that humans cannot perceive. Kim and colleagues therefore worked with astrophysicists at Johns Hopkins University in nearby Baltimore to apply a computer algorithm called the Sequencer developed by astrophysicists. Last year, the algorithm revealed a previously unknown relationship between the masses of supermassive black holes and the properties of their host galaxies.

In a paper published in Science, the team describes how the Sequencer systematically sifted through thousands of seismic signals of diffracted waves, measuring the value of a specific quantity called the Wasserstein metric. Then it played a mathematical game of join the dots – finding the shortest path between all the data points. When there was clearly one optimal path, this showed a trend in the data.

When the researchers focused on signals beneath the northern Pacific Ocean, two regions stood out as strong generators of diffracted, delayed waves called postcursors, which are produced when seismic waves interact with anomalous structures. The first is beneath Hawaii. This was known to host a seismic wave anomaly, but data from the Sequencer — as well as additional analysis — shored up the hypothesis that it may result from a mantle plume and helped to localize it more precisely.

“Mega ultralow-velocity zone”

The second notable anomaly is, they believe, a previously unknown “mega ultralow-velocity zone” beneath the remote Marquesas Islands in French Polynesia. Scientists already know of some similar zones — beneath Iceland and Samoa, for example – that have been associated with areas of highly unusual geochemical compositions. This has led to suggestions these features may harbour material predating the giant impact on Earth that is thought to have formed the Moon. The new discovery beneath Marquesas, suggests Kim, could potentially test this hypothesis.

Kim explains the work is part of a trend in many sciences towards the use of artificial intelligence to optimize search strategies and find patterns in data that elude humans. “Broadly speaking in seismology and geophysics in general, we use supervised machine learning to search for ‘labelled’ objects such as earthquake signals from seismic recordings,” he explains, “However, where it gets interesting is in unsupervised learning, where we don’t know what we’re looking for in our datasets. That’s where this algorithm fits – you’re looking at datasets as points in a high-dimensional space and looking for patterns and clusters.” He says a forthcoming paper by the Johns Hopkins researchers describes how the Sequencer was used to sort samples of another type of earthquake wave called surface waves.

“This is very innovative, and represents a direction we need to go in seismology,” says earth scientist Edward Garnero of Arizona State University in the US. “Methods like this may help us to find things we may not even have been looking for – and apparently they did.” He is cautious, however, about the assumption that the delay in the postcursor signals came solely from the core-mantle boundary, wondering whether mantle heterogeneity elsewhere might have played a role. Ultimately, he says: “In seismic modelling, we are always up against the issue of knowing if our favoured solution model is [both] unique and the real earth.”

Copyright © 2024 by IOP Publishing Ltd and individual contributors