Upscaling Uncertain Permeability Using Small Cell Renormalization

Abstract

Sedimentary rocks have structures on all length scales from the millimeter to the kilometer. These structures are generally associated with variations in rock permeability. These need to be modeled if we are to make predictions about fluid flow through the rock. However, existing computers are not powerful enough for us to be able to represent all scales of heterogeneity explicitly in our fluid flow models—hence, we need to upscale. Small cell renormalization is a fast method for upscaling permeability, derived from an analogue circuit of resistors. However, it assumes that the small scale permeability distribution is known. In practice, this is unlikely. The only information available about small scale properties is either qualitative, derived from the depositional setting of the reservoir, or local to the wells as a result of coring or logging. The influence of small scale uncertainty on large scale properties is usually modelled by the Monte Carlo method. This is time-consuming and inaccurate if not enough realisations are used. This paper describes a new implementation of renormalization, which enables the direct upscaling of uncertain small-scale permeabilities to produce the statistical properties of the equivalent coarse grid. This is achieved by using a perturbation expansion of the resistor-derived equation. The method is verified by comparison with numerical simulations using the Monte Carlo method. The prediction of expected large-scale permeability and its standard deviation are shown to be accurate for small cell standard deviations of up to 40% of the mean cell value, using just the first nonzero term of the perturbation expansion. Inclusion of higher order terms allows larger standard deviations to be modeled accurately. Evaluation of cross-terms allows correlations of actual cell values, over and above the background structure of mean cell values. The perturbation method is significantly faster than conventional Monte Carlo simulation. It needs just two calculations whereas the Monte Carlo method needs many thousands of realisations to be generated and renormalized to converge. This results in significant savings in computer time.