Prediction of Shear-Wave Velocity Using Receiver Functions Based on the Deep Learning Method

The teleseismic receiver function contains a lot of information on converted P-s waves and multiple reflections generated by velocity discontinuities below stations, which is widely used to invert fine crustal and upper mantle velocity structures. Due to the complexity of crustal structures, however, such as the existence of sedimentary or high-velocity layers, the arrival time and amplitude of converted and multiple waves change a lot, resulting in strong non-uniqueness of receiver function inversion. The convolutional neural network, as an efficient feature extraction method, can build the relationship between the receiver function and the shear wave velocity. Therefore, a convolutional neural network is designed to predict the shear wave velocity, whose sample set is built from global model data and high-quality observation receiver function dataset. Tests on dataset shows that the shear wave velocity predicted by the synthetic data is in good agreement with corresponding models. The predicted shear wave velocity from observed data is largely consistent with the global inversion results, and the prediction of the shear wave velocity discontinuities is in accordance with the traditional H-kappa stacking results. Using this method, we image the fine shear wave velocity and crustal structure by inversion of data from an OBS array deployed in the Ryukyu Trench. Test experiments and applications both show this method is not only of high computational efficiency but also of high reliability.