A Second-order Adjoint Truncated Newton Approach to Time-domain Multiparameter Full Waveform Inversion in Viscoacoustic Medium
Résumé
The truncated Newton (TRN) method has been demonstrated as an important optimization method for multi-parameter full waveform inversion (FWI). In this study, we propose an implementation of this strategy for time-domain FWI based on the 1st-order VTI viscoacoustic wave system. Compared with previous frequency-domain FWI studies with 2nd-order wave equation using TRN, the time-domain counterpart allowing for anisotropy and attenation is appealing due to widely available well-developed preprocessing tools, straightforward data windowing and the simultaneous inversion within broad frequency bands. To overcome the computational challenge for efficiently building the gradient and Hessian-vector product in the presence of seismic attenuation, we use wavefield reconstruction technique with checkpointing-assisted reverse-forward simulation (CARFS) algorithm. The TRN method equipped with the pseudo Hessian preconditioner and unity-based normalization, allows fast convergence when updating multiple parameters. The 2D Valhall synthetic study demonstrates that preconditioned TRN (PTRN) significantly improves the multiparameter reconstruction under (Vp,rho, 1/Q) parameterization, while mitigating the crosstalk issue from parameters of different classes. Moving to field data application using PTRN is an attractive direction in the future.