Multi-parameter full waveform inversion for velocity and density based on Born sensitivity kernels

Yang Ji-Zhong,Liu Yu-Zhu,Dong Liang-Guo

DOI: https://doi.org/10.6038/cjg20160328

2016-01-01

Abstract:Density is difficult to reconstruct in multi-parameter full waveform inversion (FWI). This difficulty mainly comes from the coupling effects between velocity and density. In this study, we apply the truncated Gauss-Newton method to multi-parameter FWI for simultaneous estimation of velocity and density in acoustic media. The inverse Hessian whose off-diagonal blocks reflect the coupling effects between different parameters is incorporated to decouple velocity and density during the inversion procedure. In conventional truncated Newton algorithms, the gradients of the objective function with respect to model parameters are calculated with the first-order adjoint-state method, where the forward propagated source wavefields are zero-lag cross-correlated with the back-propagated residual wavefields. The model update is computed through a matrix-free conjugate gradient (CG) solution of the Newton linear system. This operation requires the computation of the Hessian-vector products, which can be acquired by the second-order adjoint-state formulas or the finite difference method. In such a way, two additional forward wave-propagation problems have to be solved in each internal CG loop, which increase the computational cost dramatically. In this study, we calculate the gradient based on the Born sensitivity kernels, where the kernel-vector products are transformed to the accumulation of vector-scalar products. Meanwhile, the Hessian vector products are implemented by means of two nested kernel-vector products, thus the forward simulations are no longer required. To demonstrate the effectiveness of the proposed method, we apply it to a canonical inclusion model and the SEG/EAGE Marmousi-2 model to simultaneously determine velocity and density. The results show that this method can obtain reasonable results for both velocity and density without using any inversion strategy. The vertical profiles extracted from the reconstructed models depict good consistency with the true ones. Accurate consideration of the inverse Hessian is of great importance in the context of multi parameter FWI, which helps to mitigate the coupling effects and rescale the magnitudes of different parameters. The Hessian-vector products can be efficiently computed using kernel-vector products without solving any additional forward simulation problems, thus the computational cost of the truncated Gauss-Newton FWI is reduced. Numerical experiments prove the capability of the proposed method on recovering multiple parameters simultaneously.