Lift and Relax for PDE-constrained inverse problems in seismic imaging

TL;DR

The paper introduces LRWI, a novel method combining convexification and soft wave-equation constraints to improve seismic waveform inversion, enabling successful inversion with poorer initial models and higher starting frequencies.

Contribution

LRWI integrates lifting and relaxation techniques to reformulate seismic inversion as a convex problem, enhancing robustness against local minima and poor initial models.

Findings

LRWI outperforms FWI and WRI in challenging inversion scenarios.

LRWI allows higher starting frequencies for successful inversion.

LRWI reduces the numerical rank of the moment matrix to one.

Abstract

We present Lift and Relax for Waveform Inversion (LRWI), an approach that mitigates the local minima issue in seismic full waveform inversion (FWI) via a combination of two convexification techniques. The first technique (Lift) extends the set of variables in the optimization problem to products of those variables, arranged as a moment matrix. This algebraic idea is a celebrated way to replace a hard polynomial optimization problem by a semidefinite programming approximation. Concretely, both the model and the wavefield are lifted from vectors to rank-2 matrices. The second technique (Relax) invites to consider the wave equation, not as a hard constraint, but as a soft constraint to be satisfied only approximately - a technique known as wavefield reconstruction inversion (WRI). WRI weakens wave-equation constraints by introducing wave-equation misfits as a weighted penalty term in the objective function. The relaxed penalty formulation enables balancing the data and wave-equation misfits by tuning a penalty parameter. Together, Lift and Relax help reformulate the inverse problem as a set of constraints on a rank-2 moment matrix in a higher dimensional space. Such a lifting strategy permits a good data and wave-equation fit throughout the inversion process, while leaving the numerical rank of the rank-2 moment matrix to be minimized down to one. Numerical examples indicate that compared to FWI and WRI, LRWI can conduct successful inversions using an initial model that would be considered too poor, and data with a starting frequency that would be considered too high, for either method in isolation. Specifically, LRWI increases the acceptable starting frequency from 1.0 Hz and 0.5 Hz to 2.0 Hz and 2.5 for the Marmousi model and the Overthrust model, respectively, in the cases of a linear gradient starting model.