Plug-and-Play Stein variational gradient descent for Bayesian post-stack seismic inversion

Sampling from a target posterior distribution provides a fundamental approach to quantifying uncertainties in geophysical inverse problems. However, selecting appropriate prior information in probabilistic inversion is crucial (yet non-trivial) as it influences the ability of a sampling-based inference algorithm to provide geological realism in the posterior samples. To tackle such a challenge, we present a novel algorithm, called the Plug-and-Play Stein Variational Gradient Descent (PnP-SVGD), which allows sampling from a regularized target posterior distribution, where the target posterior distribution is regularized by a convolutional neural network (CNN) based denoiser. By applying the proposed methodology to a post-stack seismic inversion problem on the Volve field data, we showcase its ability to produce high-resolution, geologically trustworthy samples representative of the subsurface structures, which we argue could be used for post-inference tasks such as reservoir modelling and history matching.

Izzatullah, M., Alkhalifah, T., Romero, J., Corrales, M., Luiken, N., & Ravasi, M. (2023). Plug-and-Play Stein variational gradient descent for Bayesian post-stack seismic inversion. 84th EAGE Annual Conference & Exhibition.

The authors thank King Abdullah University of Science and Technology (KAUST) and the DeepWave sponsors for supporting this research. We are grateful to Equinor and the Volve license partners for providing access to the data used in one of the examples.

European Association of Geoscientists & Engineers

Conference/Event Name
84th EAGE Annual Conference & Exhibition


Additional Links

Permanent link to this record