New AI method to detect microearthquakes

Scientists from Stanford University and Georgia Institute of Technology in the United States, introduce a new Artificial Intelligence technique to detect small-scale earthquakes.

Despite the radical technological changes of recent years, the methods to identify earthquakes have, more or less, remained the same for about 30-40 years. 

While these techniques are sufficient for the detection of moderate and large seismic shocks, the small-scale events that occur on the same faults, which are critical to identify the behavioral patterns of seismic activity, remain undetected. In fact, out of the half a million (approximately) earthquakes that occur on Earth annually, only 20% are felt by people.

Nevertheless, a new study, published in Nature Communications in August 2020 introduces a deep-learning method for earthquake detection that can be also used for phase picking (the distinction between the arrival times of primary P- and secondary S- seismic waves in an earthquake signal). Phase picking can be utilized to derive the location of a seismic event. 

The developed AI model is known as the Earthquake Transformer and was inspired by two antecedent models, the Phasenet and the CRED. Phasenet was developed in 2018 by 2 of the authors of the study and included sophisticated algorithms to manage successful phase picking. The CRED machine learning technique emerged one year later and was associated with earthquake detection using the same technology that voice-recognition systems utilize.

The Earthquake Transformer was trained via data from seismograms in California and was put to the test using continuous data from earthquake stations in Japan during the M 6.6 Tottori earthquake that occurred in 2000. The developed model proved to be very efficient, identifying more than 21,000 seismic events. Moreover, its contribution was valuable in the detection of small-scale shocks that could not have otherwise been recorded. “Earthquake monitoring using machine learning in near real-time is coming very soon,” Gregory Beroza, Professor of Geophysics at Standford and co-author of the study, stated.

Dr. Mostafa Mousavi, the lead author of the study and currently a researcher at Stanford’s School of Earth, used to manually scan earthquake data to detect true earthquakes by separating them from noise (traffic, sea waves, people walking etc.). This procedure was very time-consuming so a better alternative should emerge. “By improving our ability to detect and locate these very small earthquakes, we can get a clearer view of how earthquakes interact or spread out along the fault, how they get started, even how they stop,” Prof. Beroza, added. The identification process is being conducted as soon as the earthquakes hit.

The new tool will identify and map precisely faults and fault zones so that more accurate seismic hazard maps can be generated. Therefore, maybe a zone that has not still produced a large earthquake in recent history can be identified before a major seismic shock strikes.