Magnetotelluric Data Inversion Using Subdomain Encoding Scheme with Variational Autoencoder

TL;DR: A novel 2D magnetotelluric data inversion scheme based on variational autoencoder with subdomain encoding is proposed. The method improves the accuracy and consistency of the reconstructed model with prior knowledge.

Abstract: A 2D inversion scheme based on variational autoen-coder (VAE) is applied to magnetotelluric (MT) data inversion. Trained with the carefully designed data set, the VAE can reparameterize the subsurface model with latent variables in an adaptive manner. To improve the computational efficiency and the generalization ability of VAE, the subsurface model is divided into subdomains, all encoded with the same VAE. In the inversion, the latent variables of each sub domain are inverted with the Gauss-Newton method. This method can flexibly incorporate prior knowledge into inversion. Numerical experiments show that the reconstructed models are more accurate and better consistent with prior knowledge than traditional inversion in the spatial domain.