Initial velocity models for Full Waveform Inversion

Abstract

Full Waveform Inversion of seismic data requires the solution of a strongly non-linear optimization problem. Usually iterative gradient based methods are employed for the solution. These methods is based on the Born approximation and assumes that the derivative of the error function with respect to the velocity can be adequately represented by a linear approximation in each iteration step. This usually requires that the initial velocity model is close to the initial model to avoid cycle-skipping which cannot be described by the Born approximation. A common approach has been to use an initial velocity model based on ray tomography and depth migration which produces smooth models with kinematic properties similar to the true model. We suggest to replace ray tomography with Wave Equation Migration Velocity Analysis (WEMVA) based on a differential semblance error function. This method yields smooth velocity models comparable with models obtained from ray tomography, and in many cases requires only relatively simple one-dimensional models as a starting point. The advantage of this approach is that models obtained with WEMVA can then be used as initial models for Full Waveform Inversion to obtain a potentially automatic two-stage work flow for estimating accurate velocity models.