Artificial Intelligence Predicts Slow Earthquakes…

Aloha!

Today we have an entry in Geology. Earthquakes are notoriously difficult to predict. Now researchers have found a way, using AI techniques, to predict the occurrence of "slow" earthquakes. Slow earthquakes can last for a period of weeks or even months but because of their low intensity nature do not cause much damage. They may not even be felt by us. Continuous seismic waves carry the potential signature of an upcoming slow slip failure. These seismic waves can be analysed to find patterns and from the regularity in these patterns predictions can be made as to when an event could occur and its intensity.

Credit: Earthquake Research Institute, University of Tokyo

The researchers created a list of characteristics that precede the appearance of earthquakes. Among them, the researchers noted an exponential increase in seismic energy prior to rupture, as if more and more tiny seismic waves were being emitted from the seismic zone. These crackles were noticeable up to three months before a slow earthquake would usually be detected, which means that the advent of these phenomena could be predicted well in advance. 

A supervised ML approach using gradient boosted trees has been taken. Here the data is in the form of features taken from continuous seismic data over a time-period. A given time window of the continuous seismic data can be used to find signatures of impending failure for the next slow slip event. In the training phase, the algorithm takes as input the seismic features calculated from the first (contiguous) 50% of the seismic data (training set), and attempts to find the best model that maps these features to the time remaining before the next slow slip event (label or target). 

Further details can be found in this Nature article.