Consistent integration of geo-information

Thursday, 18 December 2014: 2:25 PM
Thomas Mejer Hansen and Knud Skou Cordua, Niels Bohr Institute - University of Copenhagen, Copenhagen, Denmark
Probabilistically formulated inverse problems can be seen as an application of data integration. Two types of information are (almost) always available: 1) geophysical data, and 2) information about geology and geologically plausible structures. The inverse problem consists of integrating the information available from geophysical data and geological information. In recent years inversion algorithms have emerged that allow integration of such different information. However such methods only provides useful results if the geological and geophysical information provided are consistent. Using weakly informed prior models and/or sparse uncertain geophysical data typically no problems with consistency arise. However, as data coverage and quality increase and still more complex and detailed prior information can be quantified (using e.g multiple point based statistics) then the risk of problems with consistency increases. Inconsistency between two independent sources of information about the same subsurface model, means that either one or both sources of information must be wrong.We will demonstrate that using cross hole GPR tomographic data, that such consistency problems exist, and that they can dramatically affect inversion results. The problem is two folded: 1) One will typically underestimate the error associated with geophysical data, and 2) Multiple-point based prior models often provide such detailed a priori information that it will not be possible to find a priori acceptable models that lead to a data fit within measurement uncertainties. We demonstrate that if inversion is forced on inconsistent information, then the solution to the inverse problem may be earth models that neither fit the data within their uncertainty, nor represent realistic geologically features. In the worst case such models will show artefacts that appear well resolved, and that can have severe effect on subsequent flow modeling. We will demonstrate how such inconsistencies can be identified during inversion. Finally, we will demonstrate that reducing the certainty of the two types of information (to a realistic level), can lead to a consistent data integration problem with a trustworthy solution.