Multi-Parameter Full-Waveform Inversion for Vti Media Based on Born Sensitivity Kernels
Liu Yu-Zhu,Wang Guang-Yin,Yang Ji-Zhong,Dong Liang-Guo
DOI: https://doi.org/10.6038/cjg20150418
2015-01-01
Abstract:Anisotropic full waveform inversion (FWI) is being widely studied. However, it remains a problem for FWI to reveal all the anisotropic parameters simultaneously because of their coupling. Parameterization and inversion strategy are important solutions to such a problem. In this study, we propose a new VTI FWI method based on a pseudo-acoustic wave equation. This method uses a matrix decomposition algorithm to explicitly construct the gradient of the objective function instead of the traditional adjoint-state method. Meanwhile, a triple-round strategy is tested effective to implement the joint inversion of all the three parameters.In FWI, the least-squares objective function is usually used, and the increment Delta m = (Delta nu, Delta epsilon,Delta delta) around a starting model m(0) = (nu(0), epsilon(0), delta(0)) can be expressed as the product of a direction vector p and a step length t under local optimization theory. The direction vector p equals the product of the transposed kernels K and the wavefield residual. Based on frequency-domain pseudo-acoustic wave equations in VTI media, we derive the corresponding sensitivity kernels for the three parameters nu, epsilon and delta. However, the kernels are too huge to store in memory. Therefore, we use a matrix decomposition algorithm, instead of the adjoint-state method, to calculate p and t by accumulation of single vector-scalar products. By this way, the huge kernel matrix K is not necessarily stored in memory beforehand. In order to obtain VTI parameters simultaneously, we analyze the sensitivities and find that nu has the most dominant influence on the wavefield, delta is the weakest, and epsilon is in the middle. Therefore, we use a triple-round strategy to reveal the three parameters. In the first round, we invert simultaneously for (nu(1), epsilon(1), delta(1)) according to an initial model (nu(0), epsilon(0), delta(0)). In the second round, we use (nu(1), epsilon(0), delta(0)) as the initial model and invert simultaneously for (nu(2), epsilon(2), delta(2)). In the third round, we use (nu(2), epsilon(2), delta(0)) as the initial model and invert simultaneously for (nu(3), epsilon(3), delta(3)). By this way, both the strong and weak parameters can be successfully constructed.In order to test the inversion capability of this method, we design a 2D VTI homogenous model with three anomalies inside. In this experiment we only invert for nu and epsilon simultaneously, with delta being fixed as the true model because delta has a very weak influence on data. The dimension of this model is 4 km x 4 km, with the discretized interval 10 m x 10 m. In total 52 shots are uniformly distributed along the shot side, and 60 receivers are uniformly laid along the other three sides away from the shot side. Inversion is accomplished in the frequency domain. The starting frequency is 2 Hz, and the end frequency is 11.5 Hz, with frequency interval 0.5 Hz. The starting model of nu is homogeneous without anomalies. The starting model of epsilon is a highly smoothed version of the true epsilon model. Finally, both nu and epsilon are well revealed, while the resolution of obtained epsilon is not as high as nu because of its natural weaker influence on the wavefield.We propose a new FWI scheme for VTI media based on Born kernels derived from a pseudoacoustic wave equation. A matrix decomposition algorithm is employed to calculate the directions and step length through accumulation, while without need to store the huge Frechet kernel or the approximate Hessian beforehand. To reveal all the anisotropic parameters, we use a triple-round strategy to overcome the different influences of different parameters on the wavefield. Numerical experiment proves the effectiveness of this method and shows its potential to construct all the parameters in complex anisotropic media.