Nonlocal and localized analyses of conditional mean steady state flow in bounded, randomly nonuniform domains: 2. Computational examples

Abstract

In paper 1 [Guadagnini and Neuman, this issue] of this two‐part series we presented exact nonlocal equations for first and second conditional ensemble moments of hydraulic head and flux under steady state groundwater flow in bounded, randomly nonuniform domains in the presence of uncertain source and boundary terms. We derived recursive approximations for these equations based on expansion in powers of a small parameter σY, which represents the standard estimation error of log hydraulic conductivity; developed a two‐dimensional finite element scheme for their solution, under the action of deterministic forcing terms, to first order in σY2 showed how to localize the conditional mean flow equations so as to yield familiar‐looking, Darcian differential equations in which, however, all quantities are conditional on data and therefore nonunique; and proposed that traditional deterministic flow models be reinterpreted in light of these localized equations. We stressed that all quantities which enter into our nonlocal and localized conditional moment equations are defined on a consistent support scale thereby obviating the need for upscaling (from support to grid scale) or the a priori definition of a computational grid; that such a grid can therefore be defined a posteriori based on the degree of smoothness one expects the conditional moments to exhibit; and that conditioning points are often spaced widely enough to render these moments relatively smooth and hence resolvable on a coarser grid than is necessary for Monte Carlo simulations. This notwithstanding, we use in this paper a high‐resolution grid to explore superimposed mean‐uniform and convergent flows in unconditional and conditional log hydraulic conductivity domains and to compare the relative accuracies of corresponding nonlocal, localized, and Monte Carlo finite element solutions. Whereas our nonlocal solution is nominally restricted to mildly nonuniform media with σY2 ≪ 1, we find that it actually yields remarkably accurate results for strongly nonuniform media with σY2 at least as large as 4. Localized solutions deteriorate relative to nonlocal results as one approaches some points of conditioning and singularity or as σY2 increases. Though they require much less computer time than do nonlocal solutions, localized solutions have the added disadvantage that they yield no information about predictive uncertainty.