On accounting for the effects of crust and uppermost mantle structure in global scale full-waveform inversion

Fundamental mode surface wave data have often been used to construct global shear velocity models of the upper mantle under the so-called 'path average approximation', an efficient approach from the computational point of view. With the advent of full-waveform inversion and numerical wavefield computations, such as afforded by the spectral element method, accounting for the effects of the crust accurately becomes challenging. Here, we assess the merits of accounting for crustal and uppermost mantle effects on surface and body waveforms using fundamental mode dispersion data and a smooth representation of the shallow structure. For this we take as reference a model obtained by full-waveform inversion and wavefield computations using the spectral element method, model SEMUCB-WM1 and compare the waveform fits of synthetics to different parts of three component observed teleseismic records, in the period band 32-300 s for body waves and 40-300 s for surface waves and their overtones for three different models. The latter are: a dispersion-only based model (model Disp_20s_iter5), and two models modified from SEMUCB-WM1 by successively replacing the top 200 km (model Merged _200 km) and top 80 km (model Merged _80 km), respectively, by a model constrained solely by fundamental mode surface wave dispersion data between periods of 20 and 150 s. The crustal part of these three models (resp. SEMUCB-WM1) is constrained from dispersion data in the period range 20-60 s (resp. 25-60 s), using the concept of homogenization which is tailored to simplify complex geological features, enhancing the computational efficiency of our seismic modelling. We evaluate the fits to observed waveforms provided by these three models compared to those of SEMUCB-WM1 by computing three component synthetics using the spectral element method for five globally distributed events recorded at 200+stations, using several measures of misfit. While fits to waveforms for model 3 are similar to those for SEMUCB-WM1, the other two models provide increasingly poorer fits as the distance travelled by the corresponding seismic wave increases and/or as it samples deeper in the mantle. In particular, models 1 and 2 are biased towards fast shear velocities, on average. Our results suggest that, given a comparable frequency band, models constructed using fundamental mode surface wave data alone and the path average approximation, fail to provide acceptable fits to the corresponding waveforms. However, the shallow part of such a 3-D radially anisotropic model can be a good starting model for further full waveform inversion using numerical wavefield computations. Moreover, the shallow part of such a model, including its smooth crustal model, and down to a maximum depth that depends on the frequency band considered, can be fixed in full-waveform inversion iterations for deeper structure. This can save significant computational time when higher resolution is sought in the deeper mantle. In the future, additional constraints for the construction of the homogenized model of the crust can be implemented from independent short period studies, either globally or regionally.