Global Sensitivity Analysis to Improve Geomechanical Stress Characterizations Using Physics-Based Machine Learning Models

Robust predictions of in-situ stress states are essential for the safety assessment and long-term stability of nuclear waste disposal sites. However, these predictions are inherently uncertain due to the variability in geological parameters and material properties as well as uncertainties of model calibration data. Thus, a large number of model simulations would be required for a complete investigation of the model uncertainties which is not feasible due to required high numerical resolution with several million discretization points. An alternative to classical full order solutions is to develop surrogate models that run much faster but perform with similar precision.

We propose to use a machine learning-aided methodology to set up and solve these surrogate models. Specifically, we use the non-intrusive reduced basis (NI-RB) method. The resulting surrogate models are 5-6 orders of magnitude faster compared to the initial full-order model which allows an extremely fast computation of many models with different parameters. The initially required full order geomechanical simulations are conducted using GOLEM, based on the MOOSE framework (a multiphysics simulation platform).

For our case study, we use benchmark models and a simplified model inspired by the potential siting area Nördlich Lägern for high-level nuclear waste in northern Switzerland. Preliminary results indicate that our surrogate model accurately replicates the findings of the full order solutions while significantly reducing computational costs. We primarily focus on global sensitivity analyses to identify the most critical parameters impacting the stress field. Our study explores seven scenarios for surrogate modeling, each focused on different model parameters. The first five scenario examine boundary conditions, rock properties (density, Poisson ratio, Young’s modulus), geometrical features and combinations of the three, using a benchmark model to demonstrate general implication for geomechanical studies. For these scenarios, we change between two to thirteen parameters. The sixth scenario uses the simplified study based on the Nördlich Lägern, adjusting 15 parameters (Young’s modulus of each lithological layer) illustrating the potential for future real-case applications.

We show an additional seventh scenario that integrates comprehensive fault considerations, including parameters such as geometry, geographical location, dip angle, and strike direction. These factors are vital in the context of subsurface engineering studies, as they significantly influence the stress fields and the overall stability of the geological formation. A thorough understanding of fault characteristics is paramount for assessing potential risks and ensuring long-term safety and structural integrity.

The results demonstrate that the surrogate models are much faster but keep a similar precision as the full order solution. This shows the potential of surrogate modeling for rapid uncertainty quantification in geomechanics, offering a useful tool for assessing nuclear waste disposal sites, but also different applications like, for example, geothermal exploration.