Deep Learning Velocity Model Building using Fourier Neural Operators

We introduce a deep learning workflow that uses Fourier Neural Operators (FNOs) to estimate velocity models from field data with minimal pre-conditioning. Compared to convolutional Neural Networks (CNNs), FNO architectures use global long operators that makes them more powerful in the non-linear mapping from seismic records to earth models. We describe the VMB workflow that includes training on synthetic shot gathers computed from thousands of randomly generated earth models. We demonstrate its performance on a field survey acquired offshore Newfoundland, Canada. Results show that an accurate background velocity model can be inferred directly from the field data after minimal pre-processing and without prior information. The deep learning FNO-based workflow has the potential to significantly reduce the turnaround time of model building projects.