Feature-based Inversion of 2.5D Controlled Source Electromagnetic Data using Generative Priors

TL;DR

The paper tackles ill-posed 2.5D marine CSEM inversion by marrying a physics-driven forward/inverse scheme with a learned generative prior. It introduces a plug-and-play framework where a variational autoencoder constrains conductivity realizations to the decoder range, while a Gauss-Newton update operates on the nonparametric model with periodic projections onto the VAE manifold. This approach preserves the forward operator and data misfit while embedding feature-level geological priors, improving boundary resolution and robustness across survey configurations. Numerical and field experiments demonstrate consistent gains in reconstruction quality and generalization, suggesting practical benefits for hydrocarbon reservoir imaging in CSEM surveys.

Abstract

In this study, we investigate feature-based 2.5D controlled source marine electromagnetic (mCSEM) data inversion using generative priors. Two-and-half dimensional modeling using finite difference method (FDM) is adopted to compute the response of horizontal electric dipole (HED) excitation. Rather than using a neural network to approximate the entire inverse mapping in a black-box manner, we adopt a plug-andplay strategy in which a variational autoencoder (VAE) is used solely to learn prior information on conductivity distributions. During the inversion process, the conductivity model is iteratively updated using the Gauss Newton method, while the model space is constrained by projections onto the learned VAE decoder. This framework preserves explicit control over data misfit and enables flexible adaptation to different survey configurations. Numerical and field experiments demonstrate that the proposed approach effectively incorporates prior information, improves reconstruction accuracy, and exhibits good generalization performance.

Figures (20)

• Figure 1: Schematic of the finite difference mesh for 2.5D mCSEM forward modeling. "Ext." denotes the extended region.
• Figure 2: Schematic of the variational autoencoder used in this work.
• Figure 3: The workflow of the proposed method consists of two stages: training and inversion. In the training stage, a variational autoencoder (VAE) is used to learn prior patterns in conductivity models. After training, the decoder is extracted and incorporated into the inversion stage, where iteratively updated conductivity models are projected onto the output space of the decoder. This approach differs from supervised end-to-end schemes, which typically require paired (data, model) examples for training.
• Figure 4: Visualization of some example images in the training dataset and corresponding reconstructions.
• Figure 5: Visualization of intermediate models transferring between two ground truth models in Fig. \ref{['datasets']}.