Data-adaptive global full-waveform inversion

SUMMARY We present a novel approach to global-scale full-waveform inversion (FWI) that can reduce computational cost by over an order of magnitude, compared previously published methods, without sacrificing physical and mathematical rigour. This is based on data-adaptation, thereby application-oriented specialization, two complementary levels. On the simulation level, we exploit approximate azimuthal symmetry seismic wavefields implementing wavefield-adapted meshes discrete adjoints, lowering numerical cost. measurement use quasi-stochastic where variable mini-batches data are used during iterative misfit minimization in promote parsimonious exploitation data. In addition methodological developments, long-period (100–200 s) waveforms from 1179 earthquakes for 3-D whole-mantle structure. The 72 iterations approximately equals one third single iteration using FWI with widely cubed-sphere-based non-stochastic gradient optimization. resulting LOng-Wavelength earth model (LOWE) constitutes first global constructed entirely spherically symmetric initial mantle While mostly serving as showcase method, LOWE contains wealth regional-scale structures compare well earlier tomographic images. Being conservatively smooth minimal assumptions, it may therefore serve starting future inversions at shorter period or smaller scales.