Bayesian Gibbs Markov chain: MRF-based Stochastic Joint Inversion of Hydrological and Geophysical Datasets for Improved Characterization of Aquifer Heterogeneities.

Abstract:

Modeling aquifer heterogeneities (AH) is a complex, multidimensional problem that mostly requires stochastic imaging strategies for tractability. While the traditional Bayesian Markov chain Monte Carlo (McMC) provides a powerful framework to model AH, the generic McMC is computationally prohibitive and, thus, unappealing for large-scale problems. An innovative variant of the McMC scheme that imposes priori spatial statistical constraints on model parameter updates, for improved characterization in a computationally efficient manner is proposed.

The proposed algorithm (PA) is based on Markov random field (MRF) modeling, which is an image processing technique that infers the global behavior of a random field from its local properties, making the MRF approach well suited for imaging AH. MRF-based modeling leverages the equivalence of Gibbs (or Boltzmann) distribution (GD) and MRF to identify the local properties of an MRF in terms of the easily quantifiable Gibbs energy. The PA employs the two-step approach to model the lithological structure of the aquifer and the hydraulic properties within the identified lithologies simultaneously. It performs local Gibbs energy minimizations along a random path, which requires parameters of the GD (spatial statistics) to be specified. A PA that implicitly infers site-specific GD parameters within a Bayesian framework is also presented.

The PA is illustrated with a synthetic binary facies aquifer with a lognormal heterogeneity simulated within each facies. GD parameters of 2.6, 1.2, -0.4, and -0.2 were estimated for the horizontal, vertical, NESW, and NWSE directions, respectively. Most of the high hydraulic conductivity zones (facies 2) were fairly resolved (see results below) with facies identification accuracy rate of 81%, 89%, and 90% for the inversions conditioned on concentration (R1), resistivity (R2), and joint (R3), respectively. The incorporation of the conditioning datasets improved on the root mean square error (RMSE) of an unconditionally randomly generated starting model by 65%, 77%, and 78% for R1, R2, and R3, respectively. The PA was also found to converge rapidly, requiring 1.3 hrs, 4.2 hrs, and 2.5 hrs of runtime, to invert R1, R2, and R3, respectively. The PA provides a viable computationally efficient approach for improved modeling of AH.