A deep-learning-based approach for seismic surface-wave dispersion inversion (SfNet) with application to the Chinese mainland

Abstract

Surface-wave tomography is an important and widely used method for imaging the crust and upper mantle velocity structure of the Earth. In this study, we proposed a deep learning (DL) method based on convolutional neural network (CNN), named SfNet, to derive the vS model from the Rayleigh wave phase and group velocity dispersion curves. Training a network model usually requires large amount of training datasets, which is labor-intensive and expensive to acquire. Here we relied on synthetics generated automatically from various spline-based vS models instead of directly using the existing vS models of an area to build the training dataset, which enhances the generalization of the DL method. In addition, we used a random sampling strategy of the dispersion periods in the training dataset, which alleviates the problem that the real data used must be sampled strictly according to the periods of training dataset. Tests using synthetic data demonstrate that the proposed method is much faster, and the results for the vS model are more accurate and robust than those of conventional methods. We applied our method to a dataset for the Chinese mainland and obtained a new reference velocity model of the Chinese continent (ChinaVs-DL1.0), which has smaller dispersion misfits than those from the traditional method. The high accuracy and efficiency of our DL approach makes it an important method for vS model inversions from large amounts of surface-wave dispersion data.