A recurrent neural network for ??1 anisotropic viscoelastic full-waveform inversion with high-order total variation regularization Available to Purchase

A theory-guided deep learning approach to full waveform inversion has features which make it increasingly relevant as we treat problems of greater and greater complexity (e.g., when number of parameter classes increases). Here, we design a recurrent neural network (RNN) for anisotropic viscoelastic full waveform inversion. Eight parameters are simultaneously inverted for, the elastic parameters C₁₁, C₁₃, C₃₃, C₄₄ and their corresponding attenuation parameters QC₁₁, QC₁₃, QC₃₃, QC₄₄. The RNN is built around the velocity-stress anisotropic viscoelastic wave equation. It is a convenient environment in which to compute medium property FWI updates, given a non self-adjoint wave operator, and when the objective function is not analytically differentiable, but some experimentation is required to determine best choices of optimization and regularization in this multi-parameter setting. An objective function based on the l₁ norm with high-order total variation (TV) regularization is selected, and the numerical response of the network, given simulated data for a range of acquisition geometries, is examined. Given accurate starting models, we observe that the high-order TV regularization is adept at recovering sharp edges in model parameters. We also observe a significant improvement in parameter recovery given a combination of cross-well and surface data in comparison with surface acquisition.