Bayesian reconstruction of binary media with unresolved fine-scale spatial structures

We present a Bayesian technique to estimate the fine-scale properties of a binary medium from multiscale observations. The binary medium of interest consists of spatially varying proportions of low and high permeability material with an isotropic structure. Inclusions of one material within the other are far smaller than the domain sizes of interest, and thus are never explicitly resolved. We consider the problem of estimating the spatial distribution of the inclusion proportion, F(x), and a characteristic length-scale of the inclusions, δ, from sparse multiscale measurements. The observations consist of coarse-scale (of the order of the domain size) measurements of the effective permeability of the medium (i.e., static data) and tracer breakthrough times (i.e., dynamic data), which interrogate the fine scale, at a sparsely distributed set of locations. This ill-posed problem is regularized by specifying a Gaussian process model for the unknown field F(x) and expressing it as a superposition of KarhunenLoève modes. The effect of the fine-scale structures on the coarse-scale effective permeability i.e., upscaling, is performed using a subgrid-model which includes δ as one of its parameters. A statistical inverse problem is posed to infer the weights of the Karhunen-Loève modes and δ, which is then solved using an adaptive Markov Chain Monte Carlo method. The solution yields non-parametric distributions for the objects of interest, thus providing most probable estimates and uncertainty bounds on latent structures at coarse and fine scales. The technique is tested using synthetic data. The individual contributions of the static and dynamic data to the inference are also analyzed.