Velocity model building from raw shot gathers using machine learning

Abstract

Summary We present a machine learning setup that can estimate a velocity model from raw seismic shot gathers without the need for an initial velocity model. Our setup is based on a convolutional neural network (CNN) trained on pairs of random generated synthetic velocity models and corresponding forward modelled synthetic shot gathers. The network is trained to predict the correct velocity model for a given input shot gather. We evaluate the performance of the trained network on both synthetic and real seismic data, and observe that the system is able to estimate background velocity trends directly from the raw shot gathers without need for preprocessing or preconditioning. Once trained, the network is very fast to run, and can deliver a velocity model in seconds running on a single GPU. The preciscion and resolution of the estimated velocity models is not on par with state of the art velocity model building techniques such as FWI and/or reflection tomography, but shows that machine learning can robustly extract meaningful velocity information from raw shot gathers, and that there might be potential in using such methods for velocity model building.