Teaching machines to listen for earthquakes

New approach could help control seismic activity linked to geothermal production. Vhairi Mackintosh reports.

The Geysers geothermal field in California, site of novel earthquake research.
The Geysers geothermal field in California, site of novel earthquake research.
Jerry Dodrill/Getty Images

Seismic waves are often studied to determine the magnitude and epicentre of an earthquake. Sound waves have been studied by machine learning algorithms to detect patterns in music and human speech.

In a study published in the journal Science Advances, researchers from Columbia University in New York, US, have shown that machine learning algorithms also have the capability to differentiate earthquake types using their sound wave patterns.

“It's a totally new way of studying earthquakes,” says lead author Benjamin Holtzman.

Most previous studies that have used machine learning algorithms to understand earthquakes have focused on detection. In this novel approach, the scientists used the algorithms to characterise 46,000 earthquakes that occurred over a three year period within the Geysers geothermal field in California – the oldest and one of the most seismically active fields in the world.

Geothermal energy production involves extracting vapour from hot water injected deep into the crust from the surface through wells or fractures. The steam is then used to drive turbine generators that produce electricity. However, this carbon-dioxide-free power source also induces tremors.

Holtzman, a geophysicist, has previously worked on projects converting the seismic waves of earthquakes to sound. Encouraged by the findings of one of his previous studies that showed the human ear could differentiate between certain types of earthquakes, he wondered if machine learning algorithms could reveal new information from the audible quake data.

After converting the seismic waves to sound, the researchers analysed the spectral fingerprint of each earthquake using three algorithms. The first detected the most common frequencies, the second picked out the most common frequency combinations and calculated their unique acoustic fingerprints, and the third grouped similar earthquake fingerprints.

The researchers found that the earthquake clusters had a repeating temporal pattern that matched the seasonal pattern of the injected volumes of water in the Geysers reservoir, suggesting a causal link.

In winter, when a lot of water is injected underground, there are many earthquakes producing one type of signal. Conversely, when rates of water-injection are low in summer, there are fewer earthquakes and another distinct signal.

“The work now is to examine these clusters with traditional [seismic analysis] methods and see if we can understand the physics behind them,” says co-author Felix Waldhauser.

“Usually you have a hypothesis and test it. Here you're building a hypothesis from a pattern the machine has found.”

The researchers now plan to apply this approach to other natural and laboratory-simulated earthquakes to see if the signal types can be attributed to different faulting processes.

If so, combining the machine learning approach with standard seismic analysis methods could help improve the efficiency and safety of geothermal energy production.

If engineers can understand what is happening in real-time, they could regulate the amount of water in the reservoir to both reduce the number of associated large earthquakes as well as create small-scale seismicity to enhance fracture networks and, in turn, the productivity of the geothermal field.

Vhairi Mackintosh is a scientific educator, writer and researcher based in Melbourne who holds a PhD in earth sciences.
  1. http://advances.sciencemag.org/content/4/5/eaao2929
  2. https://www-sciencedirect-com.ezp.lib.unimelb.edu.au/science/article/pii/S1071581915001263
Latest Stories
MoreMore Articles