Abstract. The Italian earthquake waveform data are collected here in a dataset suited for machine learning analysis (ML) applications. The dataset consists of nearly 1.2 million three-component (3C) waveform traces from about 50 000 earthquakes and more than 130 000 noise 3C waveform traces, for a total of about 43 000 h of data and an average of 21 3C traces provided per event. The earthquake list is based on the Italian Seismic Bulletin (http://terremoti.ingv.it/bsi, last access: 15 February 2020​​​​​​​) of the Istituto Nazionale di Geofisica e Vulcanologia between January 2005 and January 2020, and it includes events in the magnitude range between 0.0 and 6.5. The waveform data have been recorded primarily by the Italian National Seismic Network (network code IV) and include both weak- (HH, EH channels) and strong-motion (HN channels) recordings. All the waveform traces have a length of 120 s, are sampled at 100 Hz, and are provided both in counts and ground motion physical units after deconvolution of the instrument transfer functions. The waveform dataset is accompanied by metadata consisting of more than 100 parameters providing comprehensive information on the earthquake source, the recording stations, the trace features, and other derived quantities. This rich set of metadata allows the users to target the data selection for their own purposes. Much of these metadata can be used as labels in ML analysis or for other studies. The dataset, assembled in HDF5 format, is available at http://doi.org/10.13127/instance (Michelini et al., 2021).

In this work we report the ongoing characterization of the Sos Enattos former mine (Sardinia, Italy), one of the two candidate sites for the Einstein Telescope (ET), the European third-generation underground interferometric detector of Gravitational Waves. The Sos Enattos site lies on a crystalline basement, made of rocks with good geomechanical properties, characterized by negligible groundwater. In addition, the site has a very low seismic background noise due to the absence of active tectonics involving Sardinia. Finally, the area has a low population density, resulting in a reduced anthropic noise even at the ground level. This location was already studied in 2012-2014 as a promising site for an underground detector. More recently, in March 2019, we deployed a new network of surface and underground seismometers at the site, that is currently monitoring the local seismic noise. Most of the energy carried by the seismic waves is due to the microseisms below 1 Hz, showing a significant correlation with the waves of the west Mediterranean sea. Above 1 Hz the seismic noise in the underground levels of the mine approaches the Peterson’s low noise model. Exploiting mine blasting works into the former mine, we were also able to perform active seismic measurements to evaluate the seismic waves propagation across the area. In conclusion we also give a first assessment about the acoustic and magnetic noise in this underground site.

The increase of available seismic data prompts the need for automatic processing procedures to fully exploit them. A good example is aftershock sequences recorded by temporary seismic networks, whose thorough analysis is challenging because of the high seismicity rate and station density. Here, we test the performance of two recent Deep Learning algorithms, the Generalized Phase Detection and Earthquake Transformer, for automatic seismic phases identification. We use data from the December 2019 Mugello basin (Northern Apennines, Italy) swarm, recorded on 13 permanent and nine temporary stations, applying these automatic procedures under different network configurations. As a benchmark, we use a catalog of 279 manually repicked earthquakes reported by the Italian National Seismic Network. Due to the ability of deep learning techniques to identify earthquakes under poor signal‐to‐noise‐ratio (SNR) conditions, we obtain: (a) a factor 3 increase in the number of locations with respect to INGV bulletin and (b) a factor 4 increase when stations from the temporary network are added. Comparison between deep learning and manually picked arrival times shows a mean difference of 0.02–0.04 s and a variance in the range 0.02–0.07 s. The improvement in magnitude completeness is ∼0.5 units. The deep learning algorithms were originally trained using data sets from different regions of the world: our results indicate that these can be successfully applied in our case, without any significant modification. Deep learning algorithms are efficient and accurate tools for data reprocessing in order to better understand the space‐time evolution of earthquake sequences.