10:45 AM - 11:00 AM
[SIT20-19] VP-VS Uncorrelation in the Lowermost Mantle Revealed by Machine Learning Traveltime Catalog
★Invited Papers
Keywords:lower mantle, LLSVP, machine learning, Vp/Vs ratio
The Large Low Shear Velocity Provinces (LLSVPs) represent the largest lateral structures within the Earth, offering a variety of insights into geodynamics and planetary evolution. Despite variations in the distributions of P-wave and S-wave velocities (VP and VS), some studies suggest that these differences can be attributed solely to temperature variations, resolution discrepancies, or the relative effects of wavefront healing for P and S waves, without necessitating diverse compositions or phases. To further investigate this region, we employed deep learning techniques for phase picking of P and S wave onset arrivals, enabling the creation of a global travel time catalog for intermediate to large earthquakes to the latest.
We adapted the methodology from Mousavi et al. (2020) to train a phase-picking model for teleseismic P and S arrivals, utilizing a dataset labeled by Houser et al. (2008). The trained model was then applied to a comprehensive dataset comprising all available long-period (1 Hz sampling) data from IRIS-DMC for earthquakes cataloged by GCMT between 1976 and 2023, encompassing approximately 60,000 events. This process included a travel time catalog containing around 6 million records from seismic stations worldwide, yielding approximately 2 million picks for P and S wave first arrivals, respectively.
Our findings indicate an increase in the traveltime residual ratio (dVS/dVP) with depth from approximately 3 to 6, which aligns with previous studies. The new arrivals provide an opportunity for the first-time comparison with the ISC catalog, which was picked at high frequency. The traveltime residuals suggest a suppression of dVS rather than dVP, countering the hypothesis presented in earlier studies (e.g., Schuberth et al., 2012) and dismissing the possibility of a homogeneous lowermost mantle.
We adapted the methodology from Mousavi et al. (2020) to train a phase-picking model for teleseismic P and S arrivals, utilizing a dataset labeled by Houser et al. (2008). The trained model was then applied to a comprehensive dataset comprising all available long-period (1 Hz sampling) data from IRIS-DMC for earthquakes cataloged by GCMT between 1976 and 2023, encompassing approximately 60,000 events. This process included a travel time catalog containing around 6 million records from seismic stations worldwide, yielding approximately 2 million picks for P and S wave first arrivals, respectively.
Our findings indicate an increase in the traveltime residual ratio (dVS/dVP) with depth from approximately 3 to 6, which aligns with previous studies. The new arrivals provide an opportunity for the first-time comparison with the ISC catalog, which was picked at high frequency. The traveltime residuals suggest a suppression of dVS rather than dVP, countering the hypothesis presented in earlier studies (e.g., Schuberth et al., 2012) and dismissing the possibility of a homogeneous lowermost mantle.