Seismic Imaging in Highly Scattering Environments using Fourier Neural Operators

Seismic imaging in volcanic settings continues to be an extremely challenging task due to the significant effect of seismic wavefield scattering from sharp, high amplitude seismic impedance changes in the subsurface. The combined effect of these along-path effects with highly rugous surface topography and complex earthquake source mechanisms results in significant codas in recorded seismograms. One of the main challenges in seismic tomography and inversion is harnessing these information-rich codas at the upper end of their frequency content, in order to resolve seismic velocity models on length scales of the smallest significant heterogeneities.

Fourier Neural Operator (FNO) machine learning models have been applied to make predictions of physical systems including flow in porous media but there are only a few examples of their use in seismology. Recent studies have demonstrated that geologically feasible velocity models can be recovered by FNOs from forward-modelled seismograms when trained on generalised model:seismogram populations, in a simulation-to-simulation (sim-to-sim) paradigm. However, an outstanding challenge for FNO research is to progress the successful performance of sim-to-sim FNOs to make robust velocity model predictions from field-gathered seismic data.

Here we generate a large population of velocity models (order 10⁴)  with statistically-generated perturbations designed to represent the scale lengths of heterogeneity observed for volcanic rocks, informed by field measurements such as petrophysical logs. Full waveform modelling is used to produce a seismogram set for each velocity model, accounting for viscoelastic attenuation. We then train an FNO neural network to predict a velocity model from input seismic records. We discuss the resolution limits of the FNO-predicted velocity models, as well as the ability to recover geometric features likely to occur in volcanic settings, from unseen data.