Machine Learning Could Help Scientists Predict and Control Seismic Activities

A new study has proven the effectiveness of listening algorithms in evaluating earthquake patterns over time.

The Geysers geothermal reservoir, from which the team analyzed three years of seismic data using machine listening. (Image courtesy of Columbia University.)

The Geysers geothermal reservoir, from which the team analyzed three years of seismic data using machine listening. (Image courtesy of Columbia University.)

Researchers at Columbia University developed a new technique that allows them to apply machine learning in studying vast earthquake data sets. This method will enable seismologists to better model the present and predict the future. Their work should result in a clearer understanding of the role human activities might play in triggering seismic events.

The Process

Published in Science Advances, the Columbia team’s study outlines a unique process for analyzing earthquake data. The team compiled a set of 46,000 earthquake seismograms, for which they then used frequency mapping to convert to a spectrogram. The data set represented three years of seismic recordings from The Geysers, a highly active geothermal reservoir in California. By using sound metrics that showed the frequencies of the waves associated with each earthquake, they were able to employ algorithms to group sets of earthquakes by sound profile.

The grouping of these clusters was a three-step algorithmic process. The first algorithm selected large blocks of the most common frequency patterns in the entire data set. The next computer model pinpointed the frequency patterns that occurred most often in each 10-second spectrogram to give each a unique fingerprint. To complete the sorting process, a clustering algorithm categorized each individual earthquake frequency with no additional context in order to guarantee the results’ integrity. 

The Implications: Tightly Controlled Human-induced Earthquakes?

The scientists found that the clusters of similar sounding earthquakes might have similar spectrograms due to the volume of water injected below ground at different times of the year. When the team compared their quake groupings against average monthly water-injection volumes at the sample site, a seasonal pattern emerged. High injection rates in winter, which result from increased runoff, are correlated with an increase in total earthquakes of a given frequency profile. In comparison, low injection rates in the warmer parts of the year seemed to result in fewer earthquakes with an entirely different sound profile.

This finding is important for future energy production. Optimizing injection levels could bring major gains in total energy outputs from geothermal sites like The Geysers. Scientists could use the information to change and increase injection levels to the safe maximum, which would result in more cracks in the crust below a geothermal reservoir—and produce more energy. The team also thinks their research could be applied to better model naturally occurring earthquakes. By applying machine listening to historical sound profiles of major seismic events, seismologists might gain insight that helps them predict the next catastrophic quake.

For more on the latest applications engineers are tackling with machine learning tools, check out Autonomous Vehicles Armed with Machine Learning Algorithms.