In this talk, we present some recent developments of using Frozen Gaussian approximation (FGA) in seismic tomography. The FGA is rigorously derived for scalar/elastic wave equation with analysis of its accuracy determined by the ratio of short wavelength over large domain size. We develop the FGA as an efficient parallel asymptotic solver for high-frequency seismic wave propagation and apply it in seismic inversion. In order to overcome the computational difficulty in summing up a large number of Gaussians targeted at a 3-D mesh, we use stochastic sampling techniques to reduce the number of Gaussians and reconstruct “low-resolution” wavefields, but the resulted “stochastic gradient” still preserves necessary information and leads the iteration process converge to the correct velocity model. This is the joint work with James Hateley (Vanderbilt U), Yixiao Hu (Tsinghua), Zhongyi Huang (Tsinghua), Ping Tong (NTU), Xu Yang (UCSB)