Full waveform inversion method based on diffusion model

A research team has published a novel AI-powered method for seismic full-waveform inversion (FWI), a critical technology for creating high-resolution images of the Earth's subsurface used in oil and gas exploration and geophysical research. The paper, "Full waveform inversion method based on diffusion model," addresses a long-standing problem: traditional FWI is highly nonlinear and prone to getting stuck in local minima, leading to inaccurate models. The researchers' innovation is to use a conditional diffusion model—a type of generative AI—as a regularizer, which learns implicit prior distributions of the subsurface to guide the inversion.

Previous attempts used unconditional diffusion processes, which ignored the inherent physical coupling between properties like seismic velocity and rock density. The new method's key advance is modifying the diffusion model's U-Net backbone network to accept two-dimensional density information as a conditional input. This allows the AI to use known density data to constrain and improve the estimation of the velocity model. Experimental results demonstrate that this conditional approach delivers inversion results with superior resolution and structural fidelity compared to older methods.

The proposed technique shows significantly stronger stability and robustness when dealing with complex geological situations, making it less likely to produce erroneous artifacts. By effectively leveraging multi-physics information (density to constrain velocity), the method moves beyond purely data-driven approaches to incorporate domain knowledge. The authors conclude it has strong practical application value for industries reliant on accurate subsurface imaging, potentially leading to more efficient resource discovery and reduced exploration risk.