Globalizing convergence of full waveform inversion via extended modeling

Abstract:

Full waveform inversion is by now well-established as a useful
geoimaging tool, capable in some instances of producing subsurface
structural information of high resolution and accuracy. Several challenges
to its widespread use remain, however, including the expense of simulation
for accurate seismic physics, and provision of uncertainty quantification allowing for
the strongly nonlinear relation between model and data. Especially
problematic is the requirement that initial model estimates be relatively
accurate, in a kinematic sense at least, predicting travel times of
significant data components to within a half-wavelength. In the context of
exploration seismic imaging of the Earth's upper crust, this requirement is
sometimes difficult or impossible to satisfy. Recent modifications to full
waveform inversion introduce additional model parameters, thus reducing the
"stiffness" of the fit criterion and enabling convergence to useful model
estimates from less accurate initial models than are required for the basic
algorithm. I will review some of these extended modeling adaptations of full
waveform inversion, and explain via some simple examples why one might
expect them to work. In fact, several extended modeling variants are closely
related to travel time data fitting, and inherit its larger domain of
convergence.