This article is more than 1 year old

Boffins want machine learning to predict earthquakes

Shocking! Lab experiments prove promising, but predicting real destruction is a lot harder

Earthquakes are, by their nature, unpredictable. Although geologists understand why and how the tremors occur, forecasting them more than a few minutes ahead is very difficult.

A team of scientists believes that machine learning could help solve this problem one day. A paper published Wednesday in the Geophysical Research Letters describes a method that relies on listening for acoustic signals from a laboratory simulation of failing fault lines.

Stress is applied to two heavy steel blocks, causing them to slip and slide over one another like tectonic plates during an earthquake. The movement releases energy in the form of seismic waves that can be studied acoustically.

Bertrand Rouet-LeDuc, co-author of the paper and a researcher at Los Alamos National Laboratory in New Mexico, US, said the prediction system relies on spotting a specific wave pattern that is emitted whenever the fault fails.

“There is lots of controversy regarding whether the prediction of earthquakes is actually possible, or if they are random in nature and will never be predicted. The fact that we find a new signal that makes it possible to predict failure for a miniature fault in the lab gives us hope that it is in fact possible.”

The system relies on a series of decision-tree algorithms. The experiment is broken down into two 150-second intervals. Seismic data generated from the first half is used for training and information about the sound emitted by the laboratory fault, and the remaining time before the failure is given.

For the second half, the system is only given the sound waves coming from the fault and has to estimate the time before the fault fails. The researchers repeated the training and testing stage for fifteen cycles.

“During the training stage, the machine learning algorithm learns to recognize a pattern in the sound coming from the fault that tells how much stress it is under, and how close it is to failing. In the testing stage, it turns out that even when the fault seems to behave differently (for example the recurrence time between events is different), its sound emission still follows the same predictable pattern,” Rouet-LeDuc explained to The Register.

The predictions become more accurate until failure. After a simulated quake, the system can guess the time for when the next tremblor will happen and is accurate to within 10 per cent. When the earthquake simulator is just about to fail, the algorithm is correct to within 2.5 per cent, the researchers claim.

“For example, for cycles that last 10 seconds on average, 2 seconds before failure the algorithm makes predictions of between [approximately] 1.9 seconds and 2.1 seconds,” Rouet-Leduc said.

Although the laboratory earthquakes provide good simulations, it’s a very simplistic model. Real seismic data is a lot messier. There is ambient noise produced by human activity and the environment. Sounds from nearby faults also interfere with the signal pattern the researchers are looking for, making predictions more difficult.

The researchers have started to work with real data, and are trying to figure out how to isolate the correct signal to pinpoint when an earthquake is just about to happen.

“The novelty of our work is the use of machine learning to discover and understand new physics of failure, through examination of the recorded auditory signal from the experimental setup,” Paul Johnson, co-author of the paper and a researcher at Los Alamos National Laboratory, told The Reg.

“I think the future of earthquake physics will rely heavily on machine learning to process massive amounts of raw seismic data.” ®

More about

TIP US OFF

Send us news


Other stories you might like