From data to model: microstructure aware models for uncertainty estimation of reactive transport in granitoide rocks

Safety of nuclear waste repositories in crystalline host rocks depends on realistic predictions of radionuclide migration in undisturbed geologies beyond the geotechnical barrier. There, the fluid migration paths will be - in absence of large scale connectivities like fissures, fault systems and joints - along weakzones like microcracks, alterations and grain boundaries. The retention potential of crystalline rocks is thus not only controlled by its modal mineralogy but also by the (heterogeneous) distribution of mineral grains, e.g. by contact area of different mineral surfaces to migration paths. Until now, reactive transport models assume homogeneous and isotropic distribution of minerals in the host rock. Including the spatial correlation of transport and mineralogy, especially the modal mineralogy along fluid migration paths, would significantly improve the estimation of radionuclide retention potential.
Modelling the microstructures is subject to uncertainties. Such uncertainties can be derived by estimating this spatial co-occurances from measured microstructures. The workflow requires spatially distributed data from analytical methods, which a) provide information about mineral composition including voids and b) allow to detect migrations paths and the mineral surface types with related areas. Here, information from gneissic samples is used to train a structure simulation model. The resulting variability of "accessible" mineral surfaces then allows to derive (by geochemical speciation codes) the variability of contaminant distribution coefficients based on sorption data and pore water composition. Finally, the spatial geological predictions are applied in reactive transport models to calculate the uncertainty of the radionuclide retention within a representative rock volume. In this contribution, we present a workflow from (real) samples to microstructure aware retention models, and discuss challenges of input data uncertainties, how they affect the total uncertainty of the model, and whether these models can be used for upscaling.