USTC: Machine Learning Helps to Reveal Global Earthquake Rupture Pattern

Earthquake is one of the natural disasters that human society suffers. In the past two decades, medium to large scale earthquakes have caused millions of deaths globally for the rub in sensitive observation and precise prediction. Despite varied rupture processes, analyzing their similarities and differences has intrigued scientists to understand the seismological process and early detection. However, previous research could scarcely evaluate the global earthquake differences by merely stacking rupture processes of multiple earthquakes, nor can they systematically study the whole rupture process through the catalog of selected rupture characteristics alone.

Recently, researcher LI Zefeng from University of Science and Technology of China (USTC), utilized machine learning algorithm to study global earthquakes, involving more than 3,000 earthquakes with magnitude above 5.5. And he presented a panoptic view of the similarity and diversity of global earthquake rupture processes. The research was published in Geophysical Research Letters on April 11th.

The challenge for overall process comparison results from complexity comparison of STFs. In his study, Prof. Li used source time functions (STFs) to describe earthquake source processes, which contains multiple information with high dimensionality, as well as great variations of amplitude and duration.

However, by applying the machine learning algorithm of variational autoencoder (VAE), researcher LI successfully condensed and reconstructed the information contained in STFs. Hereafter, he further presented a panoptical picture on the similarity and diversity in the rupture processes.

The standard STF model not only suggested a variety of characteristics for global earthquakes, but also highlighted less-noticed special classes of earthquake. The study revealed two unique types of earthquakes: runaway earthquakes with energy release at the late stage, and complex earthquakes with multiple energy releases. In addition, it showed that large earthquakes were mainly caused by simple ruptures and relatively complex ruptures. Furthermore, the energy release mode of large earthquakes has weak magnitude dependence.

This study presents more views on earthquake rupture patterns, which has a profound implication for earthquake early warning and prediction.