Feature-based magnetotelluric inversion by variational autoencoder using a subdomain encoding scheme

Magnetotelluric (MT) data inversion aims to reconstruct a subsurface resistivity model that minimizes the discrepancy between inverted and measured electromagnetic data. Conventional pixel-based minimum-structure inversion often yields a smoothed-out reconstruction with a relatively low resolution. A priori geophysical knowledge can be embedded into inversion and improve the reconstruction resolution through proper reparameterization. However, existing reparameterization approaches, such as model-based and parametric transform-based inversion, have limited ability to incorporate various a priori information. The effectiveness of existing deep generative model-based inversion algorithms is still debatable when applied to scenarios with complex backgrounds. We develop a feature-based MT data inversion method based on a variational autoencoder (VAE) with a subdomain encoding scheme. Instead of encoding the entire domain of an investigation, we adopt a 1D subdomain encoding scheme to encode the 1D resistivity-depth models using a single VAE. The latent variables for the 2D model are a combination of the latent variables for 1D models, and the encoded region of interest (ROI) can be flexibly determined. The latent variables of ROI and the pixels outside the ROI are simultaneously inverted using the gradient-descent method. Our 1D subdomain encoding scheme reduces the complexity and diversity of the data set, and it can flexibly embed a priori knowledge with various uncertainties. Synthetic data inversion and inversion of the Southern African Magnetotelluric Experiment field data validate our method’s ability to effectively improve inversion accuracy and resolution.