Improving full-waveform inversion using signal enhancement assessed by adjoint sensitivity kernels

Abstract:

This paper presents a beamforming method that allows us to improve visibility of low-amplitude phases and increase their contribution to adjoint sensitivity kernels before using them in full-waveform inversion (FWI). Signals arriving from large epicentral distances and a range of back-azimuths are considered. This aims at increasing multi-directional contributions to sensitivity kernel calculations. The goal is to accentuate effects from low-amplitude waves. These waves are usually obscured by noise, albeit carrying signicant information about global variations within Earth's mantle.

We focus on underside reflections off the discontinuities at 410 and 670 km depth (SS precursors), which are frequently undetectable in real data. Synthetic waveforms are computed numerically using a spectral-element solver, and for models with and without topographic variations along upper-mantle discontinuities. Realistic noise is added for thorough assessment of the method. After phase alignment and correction for time shifts obtained from 1-D ray tracing, a set of time anomalies is calculated by cross-correlation in time windows around predicted phase traveltimes (underside reflections and main phase). Using this set of time shifts, we sum waveforms creating a main stack for each model.

A least-squares mist measurement is then used to derive an adjoint source determined by the time shift between stacks. Computing traveltime kernels for volumetric and boundary model parameters shows the exact sensitivity of enhanced signals without noisy contributions from destructively interfering phases. This method shows potential for reducing incoherent signals and enhancing weak seismic phases in raw recordings.

The computation of sensitivity kernels in our study helps us realise whether the stacking technique indeed enhances the desired information and whether it is suitable for precursor waves. The computation of gradients establishes the best approach of accounting for finite-frequency effects in an FWI procedure. This, in turn, can speed up convergence towards a physically meaningful model.