Accelerating full-waveform inversion using source stacking followed by cross-correlation: the first application to real long period data at the global scale

Abstract:

Source stacking (Capdeville et al., 2005) has been proposed to effectively reduce the numerical cost in wavefield computations for global tomography. However high-amplitude fundamental mode surface wave always dominate the summed waveforms, where the contribution from overtone energy is hidden. Cross-correlating the summed waveforms between station-pairs can boost up the contribution of overtone energy and body waves (Romanowicz et al., 2019). Since we can apply the same processing to the 3D synthetics and observations, and we have the good knowledge of location and mechanism, the quality of reconstruction of green’s function is less important than in the case of Ambient Noise Tomography. To evaluate the capability of resolving the shear wave velocity and its anisotropic structure in the upper mantle, we conduct a synthetic tests based on a realistic synthetic dataset derived from SEMUCB-WM1 (French and Romanowicz, 2014), which includes a 3D crustal model and mantle structure down to scales of ~800 km. The synthetic 3-component dataset, computed using the Spectral Element Method, is built for 273 globally distributed events and 515 stations. We performed the inversion in two steps: first using filtered "data" at periods >60 s and later extending the bandwidth down to 38 s period. The results show structure can be well recovered down to at least 1,500 km depth both by straight source stacking and by stacking followed by cross-correlation. Moving on to real data, we collected a new global dataset including 355 events and 340 stations for a time interval that includes first and second orbit surface waves. We then compare the resulting long-wavelength global 3D radially anisotropic shear velocity model to other existing models constructed by using conventional approaches and evaluate misfits to waveforms on individual source-station records. We deal with the issue of missing data by combining two different strategies: (1) using a cluster analysis for grouping events and stations in order to minimize missing waveforms in each group and (2) replacing missing waveforms by synthetics computed in the current 3D model using an efficient approximate method such as normal mode perturbation theory.