Automated and Reproducible Full Waveform Inversion with Multiple Data Sets

Abstract:

We present a series of methodological developments intended to (1) accelerate and automise full seismic waveform inversion from local to global scales, and (2) improve tomographic resolution and its quantification.

Our developments include an open-source framework for the management of seismic data and iterative non-linear inversions. This ensures that information on provenance, processing, modelling and inversion is systematically archived, thus facilitating reproducibility. Furthermore, tools for automised window selection, misfit measurements and input file generation for various forward solvers are provided.

To enhance resolution in regions poorly covered by earthquake data, we incorporate ambient noise correlations in the inversion. Since correlations are affected by the distribution of noise sources, we only measure the more robust traveltime differences of narrow-band surface waves; disregarding waveform details that would be exploitable in the case of earthquake data.

To quantify resolution of full waveform inversion models at minimal computational cost, we employ a newly developed stochastic sampling technique that extracts various resolution proxies from the Hessian through the application of quasi-random test models.

The Western Mediterranean serves as the real-data testing ground for our developments. Data from the IberArray project combined with noise and earthquake recordings from nearly 1000 stations throughout Europe provide exceptional coverage. Embedded within a multi-scale model of the Globe, our tomographic images provide a detailed snapshot of Western Mediterranean geodynamics, including, for instance, the lateral extent and fine-scale details of subducted lithospheric slabs in the region.