Inversion of Three-Dimensional Discrete Fracture Networks Using Hydraulic Tomography

Abstract:

Fractures have a great impact on flow and transport especially in rocks with a low-permeable matrix. Thus, proper knowledge of fracture characteristics is essential for instance in geothermal applications or nuclear waste repositories. A common approach is representing multiple interacting fractures via discrete fracture network (DFN) models. Although established DFN simulation techniques exist, there is still a lack of concepts to calibrate such models and to integrate as much as possible information from field sites in models. In our contribution, we develop a stochastic inversion method to infer the properties of a three-dimensional (3D) DFN based on the data from hydraulic tomography experiments. The unknown parameters to be adjusted by the inversion algorithm are the coordinates of the center of the fractures and the fracture length. They are interpreted as random variables characterized by the posterior probability distribution according to a Bayesian framework. The posterior distribution is analyzed by drawing samples with Markov chain Monte Carlo (MCMC) methods.

For the demonstration of the DFN inversion procedure, a synthetic case study is set up. It is based on the recent hydraulic characterization at the Grimsel test site in Switzerland. Here, tomographic configurations have been used for cross-borehole hydraulic tomography to investigate fracture flow in crystalline rock. For our case study, prior information about transmissivity and storativity are obtained from hydraulic packer tests, and the position and orientation of the fractures intersecting the boreholes stem from optical televiewer measurements. A hydraulic overpressure is applied to different injection points and the resulting pressure signals or flow rates are measured at the various receiver points as the basis for the inversion.

Our results reveal that the main properties and fluid flow paths of the DFN are correctly identified by the inversion algorithm. This is evaluated further with respect to different theoretical tomographic and inversion setups. To examine the robustness of the inversion procedure, the performance of multiple MCMC chains starting from different initial configurations is presented.