A Fast Algorithm For Elastic Wave-Mode Separation Using Deep Learning With Generative Adversarial Networks (Gans)

Wave-mode separation and wave-vector decomposition aims to separate a full elastic wavefield into different wavefields corresponding to their respective wave mode. The wave mode separation process allows the handling of different wave modes independently, which is important to ensure the accuracy of migration. Most of the previously developed algorithms require the knowledge of polarization vectors that are calculated using the Christoffel equation in the wave-number domain by using the anisotropic parameters of the mesh nodes. This process can be really expensive, especially for quasi-differential operators. We propose to implement a framework using deep neural networks (DNNs) for elastic wave-mode separation in a heterogeneous anisotropic medium. To avoid the necessity of solving the Christoffel equation at each spatial location of the medium, we develop an efficient algorithm using generative adversarial networks (GANs). In particular, we propose a wave-mode decoupling workflow where we train the neural network using five to six time slices for a particular source location and test on all other time slices for that source location, as well as for all other source locations that are not a part of training. Numerical examples of increasing complexity show that the proposed approach provides an effective wave-mode decomposition method for heterogeneous and strongly anisotropic media.