A fast method for searching for repeating earthquakes, applied to the northern San Francisco Bay area

Repeating earthquakes (REs) are the regular or semi-regular failures of the same patch on a fault, producing near-identical waveforms at a given station. Sequences of REs are commonly interpreted as slip on small locked patches surrounded by large areas of fault that are creeping (Nadeau and McEvilly, 1999). Detecting them, therefore, places important constraints on the extent of fault creep at depth. In addition, the magnitude and recurrence interval of these RE sequences can be related to the creep rate and used as constraints on slip models. In this study we search for REs in northern California fault systems upon which creep is suspected, but not well constrained, including the Rodgers Creek, Maacama, Bartlett Springs, Concord-Green Valley, West Napa and Greenville faults, targeting events recorded at stations where the instrument was not changed for 10 years or more. A pair of events can be identified as REs based on a high cross-correlation coefficient (CCC) between their waveforms. Thus a fundamental step in RE searches is calculating the CCC for all event waveform pairs recorded at common stations. This becomes computationally expensive for large data sets. To expedite our search, we use a fast and accurate similarity search algorithm developed by the computer science community (Mueen et al., 2015; Zhu et al., 2016). Our initial tests on a data set including 1500 waveforms suggest it is around 40 times faster than the algorithm that we used previously (Shakibay Senobari and Funning, AGU Fall Meeting 2014). We search for event pairs with CCC>0.85 and cluster them based on their similarity. A second, location based filter, based on the differential S-P times for each event pair at 5 or more stations, is used as an independent check. We consider a cluster of events a RE sequence if the source location separation distance for each pair is less than the estimated circular size of the source (e.g. Chen et al., 2008); these are gathered into an RE catalogue. In future, we plan to use this information in combination with geodetic data to produce a robust creep distribution model for all of the faults in this region.