Diffusion model priors in ensemble smoother seismic inversion

Diffusion deep learning models provide an alternative to conventional Gaussian-based priors for generating prior realizations in probabilistic prestack seismic inversion. The goal is to enhance the inversion of elastic properties (P-wave velocity, S-wave velocity, and density) using priors generated by diffusion models to guide an iterative seismic inversion process, which combines prior information with recorded seismic responses to improve predictions. While the conventional geostatistical-based priors are often sufficient and straightforward to apply, they usually struggle to capture complex data distributions in heterogeneous areas and rely on manually defined parameters. In contrast, diffusion-based priors learn directly from the data, offering more realistic and geologically accurate representations of subsurface properties. As a result, inversion outcomes align more closely with the true model and exhibit reduced uncertainty, making them particularly beneficial in complex geologic settings.