In the lab, machine learning improves quake forecasts

by Adityarup "Rup" Chakravorty
Thursday, December 14, 2017

Researchers used machine learning to analyze acoustic signals (bottom signal) from a laboratory fault model subjected to shear stress (top signal) and found the early "groaning, creaking and chattering" emitted by the fault model helped predict when it would fail. Credit: Rouet-Leduc et al., Geophysical Research Letters, September 2017.

The U.S. Geological Survey estimates that about 500,000 detectable earthquakes occur worldwide every year. But accurate forecasts of when quakes will occur have long been out of reach, in large part because of the complexities of fault behavior.

In a new study in Geophysical Research Letters, researchers used machine learning to analyze the “groaning, creaking and chattering” coming from a laboratory fault model to forecast when it would fail and result in a “lab quake.” The findings could help researchers better understand fault mechanics and perhaps eventually improve earthquake forecasts, says Paul Johnson, a scientist at Los Alamos National Laboratory and lead author of the study.

The apparatus at the heart of the work consists of three steel plates stacked vertically with glass beads sandwiched between the plates. The plates mimic a portion of a fault, while the glass beads represent fault gouge — the ground-up material that develops when rock surfaces rub against each other. Johnson and his colleagues manipulated the steel plates to create shear stresses similar to those that occur on real faults, and then recorded the acoustic signals emanating from the glass beads.

They analyzed several aspects of these acoustic signals, such as their amplitude and kurtosis, the shapes of the signal peaks. “We found that these acoustic signals behave consistently across experiments and evolve over time,” Johnson says. “Analyzing the signals tells us … when the next failure — and lab quake — is going to be.”

The key component for forecasting when the fault model would fail was a machine learning algorithm. (Machine learning is a form of artificial intelligence.) “Previously, we didn’t have the ability to store or analyze the seismic data we were accumulating,” Johnson says. But data storage capacities have expanded greatly in recent years, and “now, we can turn a machine learning algorithm loose on the data and see what it finds.”

To apply the machine learning algorithm, the researchers first used their laboratory fault model to generate detailed datasets of simulated earthquakes, including the acoustic signals emanating from the fault during the buildup to the lab quakes. In each case, the acoustic signals could be associated, in hindsight, with the timing of the fault failure — that is, when the steel plates slipped.

The researchers then split these datasets into two groups, using one group to “train” the algorithm to search for patterns in the data and build a model to forecast fault failure. The second group of datasets was used to assess how well the forecasts from the machine learning model matched with what had actually happened on the model fault.

Johnson and his colleagues found that forecasts of when the laboratory fault would fail were significantly more accurate with the machine learning method compared to forecasts made using a simpler statistical method. To evaluate the performance of both, the team applied a measure called the coefficient of determination, or simply R2; R2 ranges from zero to one, with zero indicating a total lack of predictive ability, and one indicating a perfect match between forecast and reality. The team found that the simpler statistical method had an R2 value of about 0.3, while the machine learning algorithm yielded an R2 value of about 0.9.

Additionally, the machine learning model found a signal in the acoustic signals coming from the fault “that we didn’t even know was there,” Johnson says. “It seems to be a fingerprint of the stress in this laboratory model that tells you when the next failure is going to be.”

It’s unclear whether — or when — the team’s lab-based technique might contribute to improved forecasts of real earthquakes, but Johnson says he expects that the work will help advance understanding of frictional physics and material failure on faults.

“In the future, [this work] could dovetail with other earthquake forecasting technologies, like early warning [systems],” says Chris Marone, a geoscientist at Penn State. “The techniques used here could also help provide physical explanations for post-seismic deformations, such as damaging aftershocks.”

However, Marone is cautious about extrapolating the laboratory findings to real-world fault systems. He says he would like to see the team’s experiments repeated with actual fault gouge material instead of glass beads “to make sure that the results don’t just apply to simple granular systems, such as the beads.”

Another potential issue is the attenuation of seismic waves, which lose energy as they propagate from their source. “Seismic attenuation means it is unclear how close to a fault one will have to be to measure and analyze the acoustic signals” needed for the new machine learning technique to work, Marone says.

“We hope that what we learn will scale up and help us understand what’s happening in the world,” Johnson says.

© 2008-2021. All rights reserved. Any copying, redistribution or retransmission of any of the contents of this service without the expressed written permission of the American Geosciences Institute is expressly prohibited. Click here for all copyright requests.