Hao Zhang, and Tariq Alkhalifah

 GitHub: https://github.com/DeepWave-KAUST/PostFWI-pub

A post-FWI refinement framework based on a Flow Matching (FM) generative model to enhance subsurface velocity models using learned geological priors. Instead of incorporating generative priors directly into the FWI optimization process, which often requires additional computationally expensive data-fitting iterations, our method performs posterior-guided sampling after FWI is completed. The proposed framework guides the deterministic generative process using the FWI result together with optional well-log constraints to generate geologically consistent high-resolution velocity models without re-running waveform simulation or additional inversion updates. By combining observational guidance with learned prior information, the method improves structural continuity, enhances resolution, and enables local well-tied corrections that better honor subsurface measurements, including correcting depth mismatches around well locations. Synthetic and field data experiments demonstrate that the proposed approach can effectively refine FWI results while maintaining low additional computational cost.