Truncated Gauss-Newton Implementation for Multi-Parameter Full Waveform Inversion

Thursday, 18 December 2014
Jizhong Yang1, Yuzhu Liu1,2, Liangguo Dong1 and Yi Wang1, (1)Tongji University, Shanghai, China, (2)UC Berkeley Seismological Laboratory, Berkeley, CA, United States
Full waveform inversion (FWI) is a numerical optimization method which aims at minimizing the difference between the synthetic and recorded seismic data to obtain high resolution subsurface images. A practical implementation for FWI is the adjoint-state method (AD), in which the data residuals at receiver locations are simultaneously back-propagated to form the gradient. Scattering-integral method (SI) is an alternative way which is based on the explicit building of the sensitivity kernel (Fréchet derivative matrix). Although it is more memory-consuming, SI is more efficient than AD when the number of the sources is larger than the number of the receivers. To improve the convergence of FWI, the information carried out by the inverse Hessian operator is crucial. Taking account accurately of the effect of this operator in FWI can correct illumination deficits, reserve the amplitude of the subsurface parameters, and remove artifacts generated by multiple reflections. In multi-parameter FWI, the off-diagonal blocks of the Hessian operator reflect the coupling between different parameter classes. Therefore, incorporating its inverse could help to mitigate the trade-off effects.

In this study, we focus on the truncated Gauss-Newton implementation for multi-parameter FWI. The model update is computed through a matrix-free conjugate gradient solution of the Newton linear system. Both the gradient and the Hessian-vector product are calculated using the SI approach instead of the first- and second-order AD. However, the gradient expressed by kernel-vector product is calculated through the accumulation of the decomposed vector-scalar products. Thus, it’s not necessary to store the huge sensitivity matrix beforehand. We call this method the matrix decomposition approach (MD). And the Hessian-vector product is replaced by two kernel-vector products which are then calculated by the above MD. By this way, we don’t need to solve two additional wave propagation problems as in the second-order AD formulae. Therefore, this implementation is more efficient than AD when the number of the sources is just larger than a few percent of the number of the receivers. Finally, we apply our method to Overthrust model to simultaneously update velocity and density. The revealed results prove the effectiveness of this method.