DoE-based history matching for probabilistic uncertainty quantification of thermo-hydro-mechanical processes around heat sources in clay rocks

In the context of geotechnical and geological barriers, a thorough analysis of uncertainty and sensitivity is a crucial aspect of any physics-based performance assessment. While experimental data are scarce in actual waste repositories, large-scale experiments in underground research laboratories (URLs) provide such data that can be used to not only qualify THMC process models but also uncertainty assessment methodologies. In this paper, we adopt a Design of Experiments (DoE)-based history matching workflow - an approach popular in the oil and gas industry - and scrutinize its applicability for multiphysical analyses of nuclear waste disposal-related processes using synthetic experimental data. Based on an analytical solution of a coupled thermo-hydro-mechanical (THM) problem of a heat source embedded in a fluid-saturated porous medium mimicking a disposal cell in an argillaceous host formation, we discuss the adaptability of the workflow as a way to address parameter and model uncertainties for barrier integrity assessment. We thereby put particular focus on the relative importance of providing defined input parameter distributions for quantities generally afflicted with epistemic uncertainty and the constraints imposed by experimental (URL) or monitoring (repository) data. We found that once constraining data is available, the particular a priori distribution plays only a minor role for the outcome, such that we can conclude that the often unknown distributions can be substituted by uniform priors under such conditions. However, detailed knowledge of parameter distributions can increase the efficiency of the workflow significantly. We conclude that the presented workflow is particularly suitable for performing uncertainty quantification and sensitivity analysis for geotechnical applications where monitoring or other experimental data are available, as it allows us to deal with models of great complexity, epistemic uncertainty and it incorporates canonically to use of measured data in order to reduce uncertainty.