Full-Waveform Inversion with Dynamic Time Alignment

We present a new method, full-waveform inversion (FWI) with dynamic time adjustment, for mitigating cycle skipping in FWI. The method has two salient features. Firstly, time shifts are computed between the predicted and recorded data. Secondly, these shifts are applied to the predicted data before computing the L2 misfit. A convergent FWI can be obtained by minimizing both the L2 data misfit (post time shift) and the time shift itself. This combines robust features of traveltime FWI (WTI) with the wavefield dynamics of FWI without invoking a Born approximation. Local crosscorrelations can fail to detect the time shift for every event in a shot record, and this can lead to dislocations in the shifted data. However, the unshifted events will have a small amplitude in the crosscorrelation seismogram. Thus, an additional feature of our scheme is that we threshold the amplitude of the correlation seismogram to provide an automatic mask against unshifted events and, thereby, avoid artificial dislocations. Comparisons are made with conventional FWI and WTI for a range of 2D and 3D models. We show that the addition of wavefield dynamics (data misfit) provides improved resolution over WTI.