Reproduce “Decoupled Latent Optimization of Diffusion Models for Full Waveform Inversion” on OpenResearch

About this paper

Chen Min, Zheng Ma

Researchers at Shanghai Jiao Tong University introduce Decoupled Latent Optimization (DLO), a framework for integrating diffusion model priors into Full Waveform Inversion (FWI) by reformulating strict latent constraints into a soft quadratic penalty. This approach enables the use of physically informed initializations and effectively reconstructs complex subsurface velocity models, consistently showing superior accuracy and robustness to noise and missing data compared to existing methods.

Key insights

• Decoupling the latent variable `z` from the physical velocity `v` through a soft penalty term stabilizes optimization and allows separate gradient computations, avoiding complex backpropagation through the combined PDE solver and sampler.
• Maintaining a fixed penalty parameter `λ` throughout optimization avoids the ill-conditioning often associated with increasing `λ` in classical penalty methods.
• The ability to incorporate physically meaningful initializations for the velocity model significantly guides the FWI optimization towards favorable basins of attraction, enhancing robustness.

Summary by alphaXiv.