On Two Possible Ways to Improve Computational Efficiency in Full Waveform Tomography based on Numerical Solvers

International audienceThe advent of accurate numerical computations of the seismic wavefield in complex 3D earth models, and among them the popular spectral element method (SEM), has opened new perspectives for full waveform mantle tomography, in particular at continental and global scales. Yet, some challenges remain, such as the computational time required to build a high resolution model including teleseismic waveform data at high enough frequencies, given the need to perform many iterations to achieve convergence of this type of non-linear problem.Here I discuss and illustrate two potential avenues for improving the efficiency of the tomographic process based on numerical wavefield computations. The first one is the concept of “box tomography”, in which a target remote region of the mantle is considered and the teleseismic wavefield is computed only once for each external source in a reference, fixed, global 3D model, using any global solver. Any subsequent iterations of the model are performed only within the target region, seamlessly connecting the wavefield outside and inside the box through the introduction of time reversal mirrors (e.g. Masson and Romanowicz, 2017). This approach was already implemented in the case of receivers located within the box (Clouzet et al., 2018) and recently extended to the general case of stations and sources located within or outside the box (e.g. Adourian et al., Fall AGU 2018).The second approach is based on source stacking (e.g. Capdeville et al., 2005). By aligning a set of sources in time and computing the resulting summed wavefield only once, it allows the reduction of computational time by several orders of magnitude. The perceived drawback is the loss of achievable resolution, because the summed wavefield is dominated by large amplitude fundamental mode surface waves that cannot be windowed and weighted as in conventional waveform tomography. However, by applying cross-correlation to the stacked wavefield at available pairs of stations, we have shown that, even while the Green’s function is incompletely reconstructed, the large amplitude surface wave energy can be isolated and weighted down to bring out the contribution of body waves and overtones (e.g. Romanowicz, et al., 2013). I illustrate the first application of this method to real data and discuss possible ways to handle missing records