Frequency Selection Strategies to Overcome Cycle Skipping in FWI

Abstract

Full-waveform inversion (FWI) has become an effective tool to develop high-resolution velocity models of the subsurface. Do to the numerous challenges of its non-linear data-fitting technique, it has become one of the most extensively studied procedures in modern day seismology. With a lack of quality low-frequency information obtained in seismic data acquisition, traditional methods of waveform inversion have required initial velocity models close to the true model. Laplace-Fourier-domain waveform inversion (WI) was developed to help overcome some of the traditional limitations of FWI. While it allows for successful waveform inversion from simple linearly increasing or even homogenous starting velocity models, it can still experience some cycle skipping related artifacts when inverting real data in these cases.

This work attempts to address the cycle skipping issue in Laplace-Fourier-domain WI in two ways. First, extrapolated low frequency information obtained from synthetic seismograms will be used to promote global convergence and avoid local minima. One of the primary limitations in seismic acquisition data is a lack of low-frequencies due to a poor signal-to-noise ratio. This study presents a new inversion workflow incorporating extrapolated low frequency information from synthetic seismograms. Synthetic seismograms are generated from inverted velocity models using reliable high frequency data. Low frequency seismic information is extrapolated from these synthetic seismograms and used to improve inversion results starting from a 2-layer homogeneous starting velocity model.

This work goes on to suggest using previously disregarded high frequency information to improve both model resolution and in turn reduce the size and impact of cycle skipping related artifacts in the inverted velocity model. Three real data examples are shown with comparisons of velocity models inverted using frequencies up to 15 Hz juxtaposed to those using frequencies up to 30 Hz. By expanding the frequency lists used in Laplace-Fourier-domain WI, improved inversion results can be seen with mitigated cycle skipping related artifacts.