Diffusion Models Enhance Subsurface Imaging in Geophysics

Full Waveform Inversion (FWI) is a critical technique for recovering subsurface velocity from seismic data, but it faces challenges due to its ill-posed and non-convex nature. While classical regularizers stabilize the inversion, they often fail to reproduce realistic geological structures. Recent diffusion-prior methods have aimed to improve realism but struggle with a delicate balance between data fidelity and prior consistency. A new approach, Decoupled Latent Optimization (DLO), has been introduced to address these limitations. DLO redefines the standard latent-optimization formulation into a quadratic-penalty objective, involving both an auxiliary physical variable and a latent variable. This decoupling allows the data-fidelity gradient to operate directly in physical space, while the diffusion sampler contributes solely through a decoded prior sample. Crucially, DLO preserves the standard smoothed-velocity initialization used in classical FWI. Evaluations on the OpenFWI benchmark demonstrate that DLO outperforms both classical regularizers and existing diffusion-based methods across clean, noisy, and missing-trace data acquisitions. The method successfully recovers intricate fault structures on the Marmousi and Overthrust benchmarks, even when the prior was trained on smaller models. DLO also exhibits robustness to initialization smoothing and measurement noise, marking a significant advancement in geophysical imaging.