Paper Summary
We have trained a deep neural network based on Fourier Neural Operators (FNOs) to replace the
traditional tomographic inversion engine and successfully performed velocity predictions on field datasets. Our implementation does not require residual move-out picking, masking, or interpretation work. We demonstrate how FNO network predictions can condition the initial Full Waveform Inversion (FWI) model by reducing cycle-skipping, building a structurally consistent model directly from a 1-D function, and effectively replacing passes of reflection tomography. The results suggest that this could be a powerful tool in accelerating the velocity model building (VMB) sequence.
The key elements of this work include the operator training, iterative image domain implementation, and the considerably large dataset seen by the network. We also discuss why this approach could potentially converge faster, utilize more information than traditional tomography engines, and mitigate limitations of ray-based methods in high-contrast environments.