Building long-wavelength velocity for salt structure using stochastic full waveform inversion with deep autoencoder based model reduction

Building velocity model for salt structure remains challenging because the strong heterogeneity of medium. Full waveform inversion (FWI) is such a technique that enables us to build a high-resolution velocity model for salt structure. However, it needs an accurate enough initial model to prevent cycle-skipping during data fitting. In this work, we propose a stochastic FWI method based on global optimization to build such an initial model. To mitigate the “curse of dimensionality” problem of global optimization, we embed a deep learning technique called deep autoencoder into the proposed method. Benefiting from the dimensionality reduction characteristic of deep autoencoder, the proposed method can transform the original large model dimensional FWI problem into a low model dimensional one that can be effectively optimized by global optimization. Numerical results verify that the proposed method can build a reasonable long-wavelength velocity model for salt structure that can then be used as an initial model for FWI.