LEVERAGING STOCHASTIC SEISMIC INVERSION FOR RESERVOIR PROPERTY ESTIMATION VIA NEURAL NETWORK PREDICTION AND STATISTICAL MODELING IN THE VOLVE FIELD, NORTH SEA.

Abstract:

Stochastic inversion provides a framework for generating high-resolution elastic properties from pre-stack seismic data. These outputs from Geostatistical and Bayesian pre-stack seismic inversion coupled with adequate petrophysical modeling can further be utilized for reservoir characterization studies via neural network prediction and non-parametric statistical modeling.

The mixture density neural network (MDN) provides efficient means for multimodal probabilistic modeling with predictive accuracies close to the standard statistical modeling techniques such as kernel density estimates and Monte Carlo sampling. The mixture density neural network and kernel density estimate is applied to estimate reservoir properties (porosity , volume of shale and water saturation). The predictive accuracies is first ascertained by the petrophysical inversion carried out on a synthetic seismic in the presence of noise. The processes are further applied on angle-stacks in the Volve 3D, North Sea.

The field application revealed that pointwise prediction with MDN requires fewer parameters (mixing coefficients, means, and covariances) to generate equivalent results from Monte Carlo sampling of the posterior distribution. MDN is the preferred method for reservoir characterization over an extensive 3D seismic survey.