Elastic wave full-waveform inversion in the time domain by the trust region method

Abstract

In this paper, we investigate the elastic wave full-waveform inversion (FWI) by the trust region method. FWI is an optimization problem of minimizing the misfit between the observed data and simulated data. Usually the gradient-based search methods such as the steepest descent method or quasi-Newton's methods are applied to update the model parameters iteratively. And the L-BFGS method with approximated Hessian information is preferred for its high computational efficiency and accuracy. At each iteration, a line search method computes a search direction and then finds a suitable step length or how far to move along the direction. The trust region method is an effective alternative to solve nonlinear optimization problems. In this paper, we investigate the performance of the trust region method in elastic time-domain FWI. In the trust region method, a trial step length called the trust-region radius is forced within a certain neighborhood of the current iterate point and a trust-region subproblem requires to solve. In order to solve the trust-region subproblem, the dogleg method and the two dimensional subspace method are adopted in this paper. In the trust region method, as long as the trust-region radius is well updated, the model updating can be performed for every iteration with the fixed trial step and there is no more extra computation for forward problem like the line search method. Numerical computations for the benchmark Marmousi model are given. The inversion results and comparisons show that the trust region method is very efficient and behaves better than the line search method such as the well-known L-BFGS method.