The IPOC catalog goes deep: preliminary results

Northern Chile is one of the most seismogenic regions on the planet, and has been monitored by a permanent network of seismic stations since 2007. We here present a first step towards a new, more complete seismicity catalog for this region, leveraging modern deep-learning based algorithms for phase picking and association.

We first assessed the performance of EQTransformer, a deep learning based phase picker, in detecting and phase picking seismic data from the Northern Chile Subduction Zone by comparison with a large, meticulously handpicked dataset. We found that the "INSTANCE" model within SeisBench yielded the best performance for our study area. Through systematic threshold variations, we determined the optimal values using Precision-Recall curves (0.4 for event detection, 0.1 for P and S picks). Subsequently, we applied GaMMA, identified as the best performing phase associator in synthetic tests, coupled with NonLinLoc for initial event location. One of GaMMA's key operational criteria is the association threshold, where we required a minimum of five seismic phases to define an event, which yielded a high reliability in the phase association process. Moreover, we refined the catalog by automatically identifying and removing duplicate events. All associated events were consecutively relocated in a 1D and a 2D velocity model, using the VELEST and simul2000 algorithms. Events with disproportionally high RMS residuals as well as single picks with high residuals were removed in the process. In a final step, events were relocated with a double-difference approach.

A first application of this combined approach for the year 2020 yielded 2,838,080 P and S picks in the picking stage, with a total of 83,194 events after association and relocation. This is a nearly tenfold increase in event numbers compared to the IPOC catalog of Sippl et al. (2023), which contains 8,716 events for the same time interval.

In this contribution, we present results from a larger-scale application of our procedure to several years of IPOC data, and compare retrieved geometries as well as event numbers to the previously published IPOC catalog. Our findings demonstrate the potential of modern deep-learning algorithms in the creation of larger and more complete earthquake catalogs. Moving forward, our goal is to extend this preliminary catalog to span the entire 15 years of IPOC operation, facilitating in-depth analysis of regional processes.