Reservoir characterization must be based on information from various sources. Well observations, seismic reflection times, and seismic amplitude versus offset (AVO) attributes are integrated in this study to predict the distribution of the reservoir variables, i.e., facies and fluid filling. The prediction problem is cast in a Bayesian setting. The a priori model includes spatial coupling through Markov random field assumptions and intervariable dependencies through nonlinear relations based on rock physics theory, including Gassmann's relation. The likelihood model relating observations to reservoir variables (including lithology facies and pore fluids) is based on approximations to Zoeppritz equations. The model assumptions are summarized in a Bayesian network illustrating the dependencies between the reservoir variables. The posterior model for the reservoir variables conditioned on the available observations is defined by the a priori and likelihood models. This posterior model is not analytically tractable but can be explored by Markov chain Monte Carlo (MCMC) sampling. Realizations of reservoir variables from the posterior model are used to predict the facies and fluid-filling distribution in the reservoir. A maximum a posteriori (MAP) criterion is used in this study to predict facies and pore-fluid distributions. The realizations are also used to present probability maps for the favorable (sand, oil) occurrence in the reservoir. Finally, the impact of seismic AVO attributes—AVO gradient, in particular—is studied.
The approach is demonstrated on real data from a turbidite sedimentary system in the North Sea. AVO attributes on the interface between reservoir and cap rock are extracted from 3D seismic AVO data. The AVO gradient is shown to be valuable in reducing the ambiguity between facies and fluids in the prediction.