Seismic imaging on volcanoes using Machine Learning

Despite advances in seismic instrumentation and seismic network densities, the ability to obtain detailed images of subsurface volcanic structure is still compromised. This leaves large uncertainties in the time evolution and nature of shallow magma emplacement, for example. Ideally it is desirable to see objects at the scale of individual sills, but strong wave scattering in volcanic settings makes this difficult to achieve and tomographic images smooth out objects at this scale. Multiple scattering creates a ‘fog’ through which it is difficult to pick singly scattered (reflected) events of interest. We use a Deep Learning approach to try capture information from this full wavefield and use that to build detailed images. Specifically we employ a Fourier Neural Operator (FNO) to model and invert seismic signals in heterogeneous synthetic volcano models. The FNO is trained using 40,000+ simulations of full wavefield elastic waves propagating through these 2D models. Once trained, the forward FNO network is used to predict elastic wave propagation and is shown to accurately reproduce the seismic wavefield. That is, the FNO can act as a fast and highly efficient forward full wavefield simulator. The FNO is also trained to predict highly heterogeneous velocity models given a set of seismograms. We show that this Deep Learning approach accurately predicts known synthetic velocity models based on surprisingly small sets of input seismograms, capturing details of the velocity structure that would lie outside the ability of current seismic methods in volcano imagery. This offers a potential new approach to imaging in volcanic environments. Although the upfront training cost of 40k simulations is very large, once trained the run times for the FNO are negligible.