Gaussian Processes enabled model calibration in the context of deep geological disposal

Abstract

This work introduces a surrogate modeling approach for an emplacement drift of a deep geological repository based on Gaussian Processes (GPs). The surrogate model is used as a substitute for the high-fidelity mechanical model in many-query scenarios, such as time-dependent sensitivity analysis and calibration. Our GP-based approach emulates the behavior of an emplacement drift of a deep geological repository with significantly reduced computational time, enabling faster design iterations and effective incorporation as well as interpretation of monitoring data. Our findings show that only a few key parameters are essential for accurately reflecting in-situ conditions in complex rock salt models, which is critical for ensuring safety in deep geological disposal.