New method uses diffusion models to improve subsurface reconstruction

A diffusion-regularized approach helps create more geologically consistent models through full waveform inversion.

Researchers introduced a diffusion-regularized Annealed Stein variational gradient descent (ASVGD) framework for full waveform inversion (FWI). This method combines data consistency with structural realism.

Full waveform inversion (FWI) is a tool for high-resolution subsurface model reconstruction. It remains highly nonlinear and ill-posed when initial models are inaccurate or data are noisy.

Annealed Stein variational gradient descent (ASVGD) is a particle-based Bayesian framework for FWI. It approximates the posterior distribution using an ensemble of velocity models to enable uncertainty quantification.

The new method uses a learned diffusion model (DM) as an implicit regularization mechanism. The trained diffusion model acts as a learned geological projection operator to guide particles toward plausible velocity models.

Numerical experiments show the method produces more geologically consistent mean models. It suppresses nonphysical artifacts and yields uncertainty estimates concentrated around complex or poorly constrained regions.

The results suggest that learned diffusion priors can enhance Bayesian FWI. This may improve inversion stability and the interpretability of posterior uncertainty.

Key facts

  • The method uses a learned diffusion model as an implicit regularization mechanism.
  • The approach combines data consistency, posterior diversity, and structural realism.
  • The method produces more geologically consistent mean models and suppresses nonphysical artifacts.

All sections

Edition