Geophysical visualization of water content distribution in bentonite by joint seismic and radar tomography

SUMMARY Bentonite is often considered as buffer material for deep geological radioactive waste repositories. To support decision making and safety assessment of radioactive waste repositories, international agencies and research institutions proposed the implementation of monitoring programmes. While the overall concepts of such monitoring programmes have been largely developed, the selection of key observations parameters, such as temperature, pressure and water content, and the technical implementation are still under development. The direct measurement of such parameters requires the placement of sensors inside a repository, which can significantly affect its safety functions and only provides information at the typically sparse sensor locations. Geophysical tomography can help gaining valuable insights into the state of the repository non-invasively by providing images of the distribution of geophysical parameters from measurements that are purely taken from the outside. However, the extracted geophysical parameters are often difficult to interpret and the geophysical tomography problem is non-unique, meaning that there exist multiple models that explain the data equally well. Here, we demonstrate that this non-uniqueness can be significantly reduced by simultaneously employing multiple geophysical methods in a joint tomography scheme. We simultaneously invert seismic and ground penetrating radar (GPR) traveltimes and amplitudes by imposing structural similarity constraints on the tomographic velocity and attenuation images. The resulting, estimated geophysical parameter maps show a strongly improved correlation when compared to results obtained from individual inversions, which in turn facilitates the establishment of constitutive relationships between the geophysical parameters (seismic and GPR velocity and attenuation) with the water content, as key parameter for the evaluation of the state of a radioactive waste repository. Using data from the full-scale emplacement (FE) experiment, we employ a supervised machine-learning model that enables the translation of the tomographic velocity and attenuation images obtained in bentonite to an image of the distribution of the water content inside the repository, where the machine learning model is trained using direct point measurements of the water content at sparse locations inside the tomographic plane. Due to the lack of direct water content sensors in the FE experiment, we use neutron log data (which are directly linked to water content) to train the machine learning model. Ultimately, this enables us to extrapolate the sparse neutron log data to a spatially cohesive distribution inside the repository corresponding to a visualization of the spatial distribution of water content.