Machine Learning-based high-resolution dataset for the 2009 L'Aquila earthquake sequence

The seismic catalog consists of 147,251 seismic events recorded during the Mw 6.1 2009 L'Aquila earthquake (central Apennines). The catalog presents events recorded from January 1 to December 31, 2009, with a completeness magnitude (Mc) of 0. This catalog was obtained using deep learning methods for automatic picking (i.e., the PhaseNet convolutional neural network, CNN from Zhu and Beroza, 2019), and the arrival times of the P- and S- waves were associated using the GaMMA (Gaussian Mixture Model Associator, from Zhu et al., 2022) associator, which uses unsupervised clustering techniques and provides also an initial localitions. Seismic waveform were recorded fom 90 seismic stations (70 from the INGV temporary and permanent and 20 from the RESIF French temporary seismic stations, see Margheriti et al., 2011). We included in the dataset new data from 12 stations from EMERSITO working group (https://eida.ingv.it/it/network/4A_2009), enlarging the already existing dataset with better coverage in the epicentral region (Valoroso et al., 2013). The CNN PhaseNet obtained a total of 6,665,989 and 7,238,784 P- and S-picks, respectively, and GaMMA combined 4,451,848 and 4,942,533 of P- and S-picks extracted, obtaining at first, 535,452 seismic events. Due to the huge amount of data extracted, we selected the best-located events, with RMS ≤ 0.4 s, horizontal and vertical hypocenter error (ERH and ERZ) ≤ 1.5 km, GAP (°) ≤ 180° and a minimum time residual for P- and S- wave arrival time ≤ 1 s, obtaining a catalog composed of 190,888 1D single-event locations. We relocated these events with Hypoellipse (Lahr, 1999) to obtain absolute locations and with hypoDD (Waldhauser and Ellsworth, 2000) to obtain relative locations. In addition, since the number of seismic events extracted is very high and it is not possible to invert all the data at the same time, we followed the procedure proposed by Waldhauser et al., (2020). We randomly selected a subsample of 5000 events 1150 times with respect to time origin and location, ensuring a minimum number of 10 and a maximum number of 56 locations for each earthquake. For all the random selections, we computed the differential travel times for each event to its 30 nearest neighbors within a 10 km distance. The double-difference relocations coming from each selection were combined into a single catalog composed of a weighted location average of each event included in more than one inversion, with the weight being a linear function of an event's distance from the centroid of the cluster it belongs to (as proposed by Waldhauser and Schaff, 2008 approach). The new catalog located with the DD-method consists of 147,251 seismic events. In this repository (in .csv format): catalog_190k_1D_mag.csv is the selected absolute location used for the DD-relocation method. catalog_148k_DD.csv is the seismic catalog obtained with DD-relocation method. Each file contains: Origin time; Hypocenter (latitude, longitude and depth) Magnitude (calibrated from GaMMA initial one) Event ID