Automated velocity model building using Fourier neural operators

We propose a deep-learning strategy based on Fourier neural operators (FNOs) for estimating velocity models from field shot gathers with minimal pre-preprocessing. In contrast with conventional CNN-based architectures based on local operators, FNOs are global convolutional operators efficiently computed in the Fourier domain. We show the advantages of using global FNOs over conventional convolutional neural networks (CNN), to achieve a better non-linear mapping between the recorded data and the subsurface velocity. We show that FNOs can be used to automate velocity model building from field data with minimal preprocessing, as demonstrated by successful inferences on data acquired with multi-sensor technology in offshore Canada.