Probabilistic full waveform inversion based on tectonic regionalization - development and application to the Australian upper mantle

Abstract

We present a first study to investigate the feasibility of a probabilistic 3-D full waveform inversion based on spectral-element simulations of seismic wave propagation and Monte Carlo exploration of the model space. Through a tectonic regionalization we reduce the dimension of the model space to 12, and we incorporate complete seismograms in order to exploit all available information in the period range from 60 to 200 s. S-wave velocity variations in the Australian Archean and Proterozoic lithospheres are generally well-constrained and strongly positive, in agreement with previous inferences from deterministic tomography. The maximum likelihood model reveals significantly elevated P velocities. While consistent with body wave studies, they are, however, not well constrained by our data, thereby providing an interesting example of a comparatively insignificant maximumlikelihood model. Our data are notably affected by 3-D density variations. The effect, however, appears to be misleading. Both the maximum-likelihood model and the posterior probability densities strongly prefer unrealistically large positive density variations that are inconsistent with independent information from geodynamics and mineral physics. This suggests that highly probable and less extreme density models may be hidden in small and hardly detectable subvolumes of the 12-D model space. It follows that deterministic full waveform inversions for density may require particularly accurate initial models. From a methodological perspective we must conclude that a transition to significantly higher dimensions would currently be difficult. Available computing power clearly imposes restrictions. However, even when Moore’s law continues to hold, the largest obstacle appears to be our inability to efficiently map small high-dimensional subvolumes with high probability.