Uncertainty Quantification for Coupled Geothermal Applications Using Physics-Based Machine Learning
Abstract
The numerical characterization of the subsurface is of major importance to provide society with, for instance, clean energy resources. However, in order to characterize the subsurface, we need to consider complex coupled subsurface processes that interact with the highly heterogeneous media. Hence, this task presents us with major challenges such as i) the description of the subsurface processes and their uncertainties, and ii) addressing the computationally expensive nature of the problem. Quantifying rock physics uncertainties and performing other probabilistic inverse methods is, even with current state-of-the-art finite element solver and high-performance infrastructures, computationally not feasible for complex basin- and reservoir-scale geothermal applications due to the large spatial, temporal, and parametric domain of the applications. Therefore, many approaches restrict the physical domain (i.e. lower degree of resolution). In this work, we follow a different approach. We perform our simplification on the mathematical instead of the physical domain. Therefore, we use the non-intrusive reduced basis method, which is a physics-based machine learning technique. We use the method to construct suitable surrogate models that yield an acceleration of several orders of magnitude of the forward problem. This allows, in turn, to perform extensive parameter studies such as uncertainty quantification. In contrast to other techniques, the non-intrusive reduced basis method aims at preserving the physics and explicitly considers the parametric partial differential equations.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMMR55B0021D