An Alternative Realization of Gauss-Newton for Frequency-Domain Acoustic Waveform Inversion

Abstract:

Since FWI was studied under the least-square misfit optimization proposed by Tarantola (1984) in time domain, it has been greatly improved by many researchers. Pratt (1998) developed FWI in frequency domain using a Gauss-Newton optimization. In recent years, FWI has been widely studied under the framework of adjoint-state methods, as summarized by Plessix (2006).

Preconditioning and high order gradients are important for FWI. Many researches have focused on the Newton optimization, in which the calculation of inverse Hessian is the key problem. Pseudo Hessian such as the diagonal Hessian was firstly used to approximate inverse Hessian (Choi & Shin, 2007). Then Gauss-Newton or l-BFGS method was widely studied to iteratively calculate the inverse approximate Hessian H_aor full Hessian (Sheen et al., 2006). Full Hessian is the base of the exact Newton optimization. Fichtner and Trampert (2011) presented an extension of the adjoint-state method to directly compute the full Hessian; Métivier et al. (2012) proposed a general second-order adjoint-state formula for Hessian-vector product to tackle Gauss-Newton and exact Newton.

Liu et al. (2014) proposed a matrix-decomposition FWI (MDFWI) based on Born kernel. They used the Born Fréchet kernel to explicitly calculate the gradient of the objective function through matrix decomposition, no full Fréchet kernel being stored in memory beforehand. However, they didn’t give a method to calculate the Gauss-Newton. In this paper, We propose a method based on Born Fréchet kernel to calculate the Gauss-Newton for acoustic full waveform inversion (FWI). The Gauss-Newton is iteratively constructed without needing to store the huge approximate Hessian (H_a) or Fréchet kernel beforehand, and the inverse of H_a is not need to be calculated either. This procedure can be efficiently accomplished through matrix decomposition. More resolved result and faster convergence are obtained when this Gauss-Newton is applied in FWI based on the Born Fréchet kernel in 2D numerical experiments. This Gauss-Newton FWI need much less computation than adjoint-state FWI when source-station number is more than only a few percent of the receiver-station number. Such condition is common in controlled source exploration, OBS exploration, natural seismology, and even some certain seismic exploration.