Joint inversion of Surface-wave Dispersions and Receiver Functions based on Deep Learning

Joint inversion of surface-waves and receiver functions has been widely used to image Earth structures to reduce the ambiguity of inversion results. We propose a deep learning method (DL) based on multi-label Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) with a spatial attention module, named SrfNet, for deriving the Vs models from Rayleigh-wave phase and group velocity dispersions and receiver functions (RFs). We use a spline-based parameterization to generate velocity models instead of directly using the existing models from real data to build the training dataset, which improves the generalization of the method. Unlike the traditional methods, which usually set a fixed Vp/Vs ratio, our new method takes advantage of the powerful data mining ability of CNN to simultaneously constrain the Vp model. A loss function is specially designed that focuses on key features of the model space (such as the Moho and the surface sedimentary layer). Tests using synthetic data demonstrate that our proposed method is accurate and fast. Application to southeast of Tibet shows a consistent result and comparable misfits to observation data with the previous study, indicating the proposed method is reliable and robust.