Accurate and Robust Full Waveform Inversion Driven by Poisson Flow Generative Model Priors

Full-waveform inversion (FWI) is an effective approach for constructing accurate subsurface models by minimizing the misfit between observed and simulated seismograms. However, its strong nonlinearity and ill-posedness often lead to trapping in local minima, especially when the initial model is far from the truth. A prior-guided inversion framework is proposed, which integrates prior information generated by the Poisson flow generative model into FWI. The learned prior is incorporated as a regularization term to guide the updates towards the true model. The proposed method has three key advantages: (1) reduced dependence on the initial model, (2) computational flexibility in sampling strategies and (3) accurate recovery of velocity models. Numerical experiments on a simple synthetic model, the Marmousi model and the three-dimensional Overthrust model demonstrate the advantages of the novel framework, which outperforms conventional and alternative regularized FWI approaches and maintains robustness in the presence of noise and complex geological features. Notably, the proposed method reduces model errors by more than (Formula presented.) compared with conventional FWI. The results highlight the potential of integrating generative priors to enhance both the stability and accuracy of FWI in challenging scenarios.