Detecting Repeating Earthquakes on the San Andreas Fault with Unsupervised Machine Learning of Spectrograms

Abstract

Repeating earthquakes—sequences of colocated, quasi-periodic earthquakes of similar size—are widespread along California’s San Andreas fault (SAF) system. Catalogs of repeating earthquakes are vital for studying earthquake source processes, fault properties, and improving seismic hazard models. Here, we introduce an unsupervised machine learning-based method for detecting repeating earthquake sequences (RES) to expand existing RES catalogs or to perform initial, exploratory searches. We implement the “SpecUFEx” algorithm (Holtzman et al., 2018) to reduce earthquake spectrograms into low-dimensional, characteristic fingerprints, and apply hierarchical clustering to group similar fingerprints together independent of location, allowing for a global search for potential RES throughout the data set. We then relocate the potential RES and subject them to the same detection criteria as Waldhauser and Schaff (2021). We apply our method to ∼4000 small (ML 0–3.5) earthquakes located on a 10 km long segment of the creeping SAF and double the number of detected RES, allowing for greater spatial coverage of slip-rate estimations at seismogenic depths. Our method is novel in its ability to detect RES independent of initial locations and is complimentary to existing cross-correlation-based methods, leading to more complete RES catalogs and a better understanding of slip rates at depth.