A comprehensive high resolution 3D P- and S-wave velocity model for the Alpine mountain chain using local earthquake data

Between 2016 and 2022 the dense and uniformly spaced AlpArray Seismic Network as part of the AlpArray-4DMB project provided an unprecedented seismic data set for the Greater Alpine region (GAR) with an average broad-band station spacing of $\approx$50km. Complementary deployments such as the EASI, CIFALPS2 and SWATH-D networks provided further increased spatial sampling of ground motions in regions of special research interest.
Contemporary, the revolutionary development of machine learning applications in key societal and economic sectors such as finance, medicine and transport was accompanied by ground-breaking progress in AI based seismological signal processing.
This thesis combines the unique seismic data set from the GAR with cutting-edge neural network based seismic picking algorithms to calculate the first comprehensive high resolution crustal 3D P- and S-wave velocity model for the GAR based on Local Earthquake Tomography.
Research questions directly linked to this study cover the benchmarking of the most commonly used neural network based seismic picking algorithms against a manually picked high precision reference catalog and their applicability to waveforms recorded at epicentral distances of up to 1000km. ... mehrI show that the deep neural network PhaseNet is most suitable for this study and that it performs as consistently as human analysts. In terms of number of detected phases it significantly outperforms manual picking approaches especially on seismic traces recorded at larger distances with lower signal-to-noise ratio. In order to remove outliers in the automatically generated arrival time catalog, I developed a data-driven pre-inversion pick selection method which requires only minor manual supervision and is applicable throughout the crustal triplication zone.
I relocate 384 events with local magnitudes $M_L\geq$2.5 while simultaneously inverting for the first comprehensive 1D P- and S-wave velocity model of the GAR including station correction terms. Results from the well established VELEST and the recently developed probabilistic Markov chain Monte Carlo (McMC) algorithms are within their respective uncertainty range and generally agree with previous models of the region. By comparing my event locations to results from two recent studies, I quantify the hypocentre accuracy as $\approx$2.0km in horizontal and $\approx$6.0km in vertical direction when using a 1D velocity model including station corrections. Subsequently, I invert for the 3D P- and S-wave velocity structure of the GAR and northern Apennines using the SIMUL2023 inversion algorithm. This inversion is based on 173,841 P- and 68,967 S-phase arrivals from 2553 events with $M_L\geq$1.5 recorded at 989 seismic broad-band stations.
Main features of the 3D model are in good agreement with previous active as well as passive seismic studies of the region. In the uppermost crust the foreland basins North and South of the Alpine arc are showing up as prominent low velocity anomalies. The crustal root is deepest in the Western and Central Alps while consistently shallowing towards the East which matches the previously proposed model of eastwards lateral extrusion of crustal material.
A striking novel feature is the consistent presence of anomalously low velocities at mid crustal depths throughout the Western and Central Alps which I interpret as stacked European and Adriatic crust.
Furthermore, the seismic signature of the Adriatic indenter indicates a thickened and deformed crust in its northern part, while the southern part appears to be mainly undeformed.
The final 3D velocity model and the arrival time data base including pick probabilities will be made available publicly and thus will contribute to a wide range of future geophysical studies within the region.
The inferred crustal velocity structure will be crucial to refine the resolution of related studies such as teleseismic tomographies of the upper mantle and receiver functions investigating the precise topography of the crust-mantle boundary (Moho) across the orogen. Until now these studies are using a combination of simplified or heterogeneously processed crustal models. This leads to ambiguities in their images of the configuration of subducted lithosphere beneath the mountain chain and thus results in major disagreement in key questions regarding the evolution of the Alpine orogen such as the potential presence of a slab tear or detachment and a subduction polarity switch along the Alpine arc.