AI based 3D P- & S-wave velocity model for the Alpine mountain chain from Local Earthquake Tomography

We present a new 3D P & S-wave model based on Local Earthquake Tomography (LET) of the European Alpine mountain chain using data from a total of more than 1100 broadband stations of the AlpArray Seismic Network and additional permanent and temporary stations. We use "SeisBench - A toolbox for machine learning in seismology" to assess the performance of the most commonly used AI pickers and find PhaseNet to be the most suitable. Our final data set comprises 2374 events of Ml >= 1.5 yielding 89,000 Pg-, 64,000 Pn-, 41,000 Sg- & 23,000 Sn-phases. Initially, we include observations from <130km epicentral distance to simultaneously relocate the quakes and invert for upper crustal velocity structure using the SIMUL2017 inversion algorithm. Subsequently, we add the remaining travel times to invert for velocities in the entire crust and upper mantle while fixing the hypocentres from the initial inversion run. 
First order features of our final vp model such as sediment basins and the Alpine orogenic root are in good agreement with previous tomographies and Moho studies of the area. In the Western Alps the well studied Ivrea Geophysical Body (IGB) is imaged as a high velocity anomaly where mantle velocities are present at depths of 15-20km. West of the IGB we find lower crustal velocities reaching depths of ~50km. Both observations are coinciding with the previously imaged Moho jump between deep European and shallow Adriatic Moho.
Similarly, we image the orogenic root in the Northern Apennines as an area of low vp with increased vp/vs-ratio. Beneath the Eastern Po plain we find mantle velocities at shallower depths than published Moho values would suggest.