Key points are not available for this paper at this time.
Between 2020 and 2022 the ANILLO+DEEPtrigger (Y6+XZ) Seismic Network, comprising 108 seismic stations, operated for eighteen months in Northern-Central Chile (24.5S - 29S). Employing Deep Learning (EQTransformer, Mousavi et al., (2022)) and Phase Association (GaMMA, Zhu et al. (2021)) algorithms, we identified over 30,000 seismic events in an area with a notable absence of moderate-to-large events in the past century, since the 1922 M8.5 Atacama earthquake. From the initial catalog, we selected a well-distributed subcatalog of 1000 earthquakes, consisting of 26,570 P- and 22,109 S-wave arrival times, by selecting for events with an optimal spatial distribution, small residuals, and abundant P- and S-arrivals. These selected events served as input for VELEST (Kissling et al., 1994) to compute a new 1-D velocity model representative of this region by minimizing the subset residuals. To reduce both residuals and location errors associated with the seismicity, we relocated the entire catalog using staggered tomographic inversions based on SIMUL2000 (Thurber Eberhart-Phillips, 1999), simultaneously inverting for seismic velocity models and hypocentral parameters within the iterative damped least squares method. Following the proposed method, we gradually increased model complexity, transitioning from 2-D Vp and Vp/Vs to ultimately a 3-D fine Vp and Vp/Vs solution with low node separation.Next, synthetic resolution tests were conducted to assess the reliability of the spatial limits and boundaries within the solutions. In this context, distinctive patterns were identified for each profile of the three-dimensional model, revealing enhanced horizontal and vertical resolution in the central region beneath the network. Conversely, a decline in resolution was noted at the peripheries, primarily attributable to reduced station coverage causing poorer seismic event relocations.Our results reveal both regional and local patterns. We observed a mantle wedge with vertical thicknesses ranging from ~35km in the southernmost profiles less than 25 km in the northern region, consistent with previous seismic tomography observations in northern Chile (Pastn-Araya et al., 2021). The Vp/Vs ratio and Vp values allow us to discern the distribution of the hydrated slab, which, spatial correlated with seismicity, provides evidence of irregular dehydratation processes along both dip and strike directions.Relocated seismicity exhibits some noteworthy features. Shallower crustal sesmicity is predominantly related to high rates of mining activity. In the subduction areas, the most prominent cluster is located at depths of 20-50 km, delineating the seismogenic zone. At greater depths, double and even triple seismic bands add structural complexities to the observations.From 26.5S to 29.5S, between 20 km and ~75 km depth, seismicity predominantly aligns with the interplate contact defined by SLAB2 (Hayes et al., 2018). In contrast, northward from 26.5S, our deepest seismicity, situated between 75 and 125 km depth, diverges from SLAB2, depicting a steeper dip angle.Lastly, we recommend integrating OBS and back-arc stations, whose data would improve off-shore and back-arc resolution, contributing to a more comprehensive understanding of seismotectonic environments. Non-supervised Deep Learning results can provide exceptional databases for tomographic studies, yielding residuals similar to human-picked databases but within shorter timeframes.
Hernádez-Soto et al. (Mon,) studied this question.