Joint long-period P and S velocity inversion for Earth's mantle based on deep learning

Many large-scale structures in the mantle have been proposed to explain seismic observations and constrain geodynamic models. While the geophysical community cannot agree on the morphology and nature(s) of large low shear velocity provinces (LLSVPs) due to the difference in approaches, decorrelated P and S velocity anomaly (dV_P no longer proportional to dV_S), inherently associated with changes in composition and/or phase, can help examine geodynamic models and imply the thermal/chemical evolution of the mantle. To further apply the inference to finer structures and to improve the precision for quantitative mineral physical implications, it is necessary to build a new seismic dataset for P and S waves measured in a self-consistent manner.

In this study, we trained a phase-picking model using code modified from EQTransformer (Mousavi et al., 2020). Our training dataset includes 65,298 traces, where teleseismic P and S arrivals are manually picked at the long-period (~20 sec) onset. Based on the machine-learning architecture proven useful for seismicity at local to regional distances, we managed to reproduce the manual picking results by machine and extend the picking catalog for seismic data to the present. We also conduct tomographic inversion for the global mantle to obtain a three-dimensional velocity model for both P and S waves. The new model has a higher resolution, allowing interpretations to understand geodynamics better.