Evaluation of Two Numerical Optimization Schemes for Elastic Full Waveform Inversion

The capability to deliver high resolution subsurface models has made full waveform inversion (FWI) a powerful process for building models of seismic properties governing wave propagation. We apply elastic FWI to recover subsurface models and evaluate the sensitivity of the inversion to the choice of optimization schemes, step length parameter, starting model and acquisition geometry. We perform the inversion using two optimization schemes: non-linear conjugate gradient (NCG) and limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS). Synthetic examples showcase trade-offs typical for multiparameter problems, and highly non-linear nature of FWI, as evidenced by many local minima. We conclude that although inversion is sensitive to all of the tested parameters, L-BFGS is more robust to the choice of step length. Furthermore, the role of the initial model is much more significant than the role of acquisition geometry.