Constructing shear velocity models from surface wave dispersion curves using deep learning

Surface wave tomography has been widely used to determine shear wave velocities by inverting surface wave dispersion curves. Conventional least-squares inversions strongly depend on an initial model and Monte Carlo inversion algorithms are usually time-consuming. In this study, we apply a deep neural network (DNN) to surface wave dispersion curves to investigate whether the initial model can be relaxed and whether reliable shear velocity models can be constructed. By applying our method to synthetic and field data, our results show that: (1) by constructing a well-trained DNN model from the global continental CRUST1.0 data, the DNN approach is effective and efficient to determine shear velocity structures using Rayleigh wave dispersion curves; (2) using the well-trained DNN model, no prior model is required, relaxing the requirement of an initial model; (3) the welltrained DNN model can be used to construct pseudo 3D seismic models across different continental areas.