Reproduce “Feature-based Inversion of 2.5D Controlled Source Electromagnetic Data using Generative Priors” on OpenResearch

About this paper

Hongyu Zhou, Haoran Sun, Rui Guo, Maokun Li, Fan Yang, Shenheng Xu

A feature-based inversion framework for 2.5D marine Controlled Source Electromagnetic (mCSEM) data uses a Variational Autoencoder (VAE) to learn geological priors within a Projected Gauss-Newton (PGN) optimization scheme. This approach demonstrated superior accuracy in reconstructing subsurface resistivity models, reducing artifacts, and showing robust performance under noise and complex geological conditions in both synthetic and real-world CSEM datasets.

Key insights

• Combining a higher-order optimization method (Gauss-Newton) with periodic projections from a deep generative prior (VAE) effectively mitigates the local minima issue while maintaining computational efficiency for nonlinear geophysical inversion.
• Feature-level generative priors, learned from synthetic geological models, significantly enhance the geological plausibility and accuracy of inverted models by suppressing artifacts and guiding the solution towards realistic subsurface structures.
• The "plug-and-play" architecture allows for flexible integration of learned priors into physics-driven inversion, preserving explicit control over data misfit and forward modeling, enabling robust performance across diverse survey conditions.

Summary by alphaXiv.