Comparison of multifrequency selection strategies for simultaneous-source full-waveform inversion

Abstract

The computational cost of full-waveform inversion (FWI) is a major obstacle to estimating the velocity model for large-scale problems. One way to reduce the overall cost of waveform inversion is by adopting a simultaneous-source strategy. In other words, multiple sources are simultaneously fired to simulate supershot gathers and thereby reduce the number of seismic modeling simulations that are performed during the inversion. However, the use of simultaneous sources introduces crosstalk artifacts that arise from the interference among the sources that constitute a supershot. We analyzed the influence of different simultaneous multifrequency selection strategies on crosstalk artifacts. Our analysis focused on a frequency-domain FWI algorithm that is implemented with simultaneous sources that are randomly encoded with random time shifts. In the multiscale conventional FWI strategies, a finite set of discrete frequencies was selected and the inversion was carried out sequentially from low- to high-frequency data components. First, the long wavelength components of model parameters were recovered from the low-frequency data, and then more details and features were extracted as the inversion proceeded to the higher frequency data. We examined six frequency selection strategies and tested the performance of the algorithm with encoded data sets. Numerical tests showed that high-fidelity results could be attained by inverting partially overlapped groups of temporal frequencies. Our FWI algorithm is based on a matrix-free Gauss-Newton method. To mitigate crosstalk artifacts during the numerical inversion, a new encoding was generated at every iteration. We also found that high-resolution images can be obtained by resampling new source positions and new encoding functions at every iteration of the FWI algorithm.