Retrieving stress and elastic properties distributions from stress orientation and satellite data using physics-informed machine learning

Abstract

Traditional geomechanical modelling involves a lot of a priori knowledge regarding, among others, the distribution of material properties and boundary conditions to apply to a model. Setting those parameters is usually time-consuming, starting with extensive literature reviews to get initial estimates, which are then refined by trial and error by calibrating the numerical simulations against observations. In particular, the boundary conditions typically remain poorly constrained since precise data is rarely available. In this contribution, we present a physics-based machine learning approach to infer the current displacements and full stress tensor distribution of an effective 2D linear elastic model, based on stress orientation and Global Navigation Satellite System (GNSS) data. This allows for automatic retrieval of a reasonable approximation of the model's elastic material properties, consistent displacement, and stress values over the whole physical domain, including at the boundaries, which could be used for instance in following forward simulations. We show an application to the Australian continent, for which a rich dataset of stress orientation is available from the World Stress Map project and the GNSS measurements are particularly steady. This allows us to compare various options to account for stress orientation and displacement information as input data. Interestingly, we recover the smoothest stress field compatible with the (very accurate) GNSS observations and consequently identify areas where the resulting stress orientation differs from current estimates.