Use Machine Learning to Remove Reverberations on Receiver Functions Caused by a Low-velocity Sedimentary Layer

Abstract:

Receiver functions (RFs), which are source-normalized P-to-S converted (Ps) phases (including Pms, PPms, PSms) from velocity discontinuities, have been widely employed to determine layered structures in the Earth’s crust and mantle. For stations underlain by a low-velocity sedimentary layer, strong reverberations on the RFs can mask the much weaker Ps phases and make it difficult to reliably determine the characteristics of the discontinuities. In this study, we utilize a machine learning (ML) approach to remove such reverberations and extract Ps phases from the Moho. We combine convolutional neural network and long short-term memory to set up our neural network (NN), which is trained using synthetic RFs produced using a reflectivity-based method with random physical parameters illustrating Earth structure. Specifically, each group of velocity, density, attenuation, layer thickness, and noise-level parameters generates two pulse functions (PFs) based on the Zoeppritz equations. The first PF contains both the reverberations and the crustal Ps phases. It convolves with a Gaussian function to generate synthetic seismogram, which is used to generate a radial RF. The second PF contains only the direct P and the three Ps phases from the Moho which serves as the target for the NN. Application of the trained NN to the synthetic RFs shows that about 70% of the resulting reverberation-removed RFs contain the direct P and all the three Ps arrivals, while the direct P and at least two of the three Ps arrivals are observed on 76% of the resulting RFs. Promising results were also achieved when the trained NN was applied to data recorded in several sedimentary basins. The ongoing work demonstrates the potential of ML based methods to process seismic waveforms for studying the interior structure of the Earth.