Velocity model building from one-shot VSP data via a convolutional neural network

Abstract

Velocity model building is an essential step in seismic exploration, which runs through the whole process of seismic data acquisition, processing, and interpretation. The velocity information is conventionally obtained by iterative optimization methods such as full-waveform inversion or tomography. These traditional methods are computationally expensive, and they require an initial velocity model and human interactions. To simplify the model building problem, we develop a supervised end-to-end conventional neural network to reconstruct the P-wave velocity models directly from raw seismic data. The network takes in one-shot seismic traces simulated with acoustic wave equations in VSP geometry. To train the network, we create 870 2D synthetic seismic images and corresponding labeled images, which are shown to be sufficient for the network to learn the nonlinear relationship between one-shot seismic data and the corresponding velocity model. The numerical examples show that the trained network is capable of predicting accurate layered velocity models from only one-shot seismic data.