Abstract
Diffusion models offer stable training and state-of-the-art performance fordeep generative modeling tasks. Here, we consider their use in the context ofmultivariate subsurface modeling and probabilistic inversion. We firstdemonstrate that diffusion models enhance multivariate modeling capabilitiescompared to variational autoencoders and generative adversarial networks. Indiffusion modeling, the generative process involves a comparatively largenumber of time steps with update rules that can be modified to account forconditioning data. We propose different corrections to the popular DiffusionPosterior Sampling approach by Chung et al. (2023). In particular, we introducea likelihood approximation accounting for the noise-contamination that isinherent in diffusion modeling. We assess performance in a multivariategeological scenario involving facies and correlated acoustic impedance.Conditional modeling is demonstrated using both local hard data (well logs) andnonlinear geophysics (fullstack seismic data). Our tests show significantlyimproved statistical robustness, enhanced sampling of the posterior probabilitydensity function and reduced computational costs, compared to the originalapproach. The method can be used with both hard and indirect conditioning data,individually or simultaneously. As the inversion is included within thediffusion process, it is faster than other methods requiring an outer-looparound the generative model, such as Markov chain Monte Carlo.