Obtaining reliable estimates of the subsurface state distributions is essential to determine the location of e.g. potential nuclear waste disposal sites. However, providing these is challenging since it requires solving the problem numerous times yielding high computational cost. To overcome this, we use a physics-based machine learning method to construct surrogate models. We demonstrate how it produces physics-preserving predictions, which differentiates it from purely data-driven approaches.

The rotation of the principal stress axes in a fault structure because of a rock stiffness contrast has been investigated for the impact of the ratio of principal stresses, the angle between principal stress axes and fault strike, and the ratio of the rock stiffness contrast. A generic 2D geomechanical model is employed for the systematic investigation of the parameter space.

It is generally assumed that faults have an influence on the stress state of the Earth’s crust. It is questionable whether this influence is still present far away from a fault. Simple numerical models were used to investigate the extent of the influence of faults on the stress state. Several models with different fault representations were investigated. The stress fluctuations further away from the fault (> 1 km) are very small.

The subsurface is subject to constant stress. With increasing depth, more rock overlies an area, thereby increasing the stress. There is also constant stress from the sides. Knowledge of this stress is fundamental to build lasting and safe underground structures. Very few data on the stress state are available; thus, computer models are used to predict this parameter. We present a method to improve the quality of the computer models, even if no direct data on the stress state are available.

Numerical methods can be used to estimate the stress state in the Earth’s upper crust. Measured stress data are needed for model calibration. High-quality stress data are available for the calibration of models for possible radioactive waste repositories in Switzerland. A best-fit model predicts the stress state for each point within the model volume. In this study, variable rock properties are used to predict the potential stress variations due to inhomogeneous rock properties.