Exploring Physics-Based Machine Learning for Geothermal Applications

Abstract

Geothermal energy has a crucial role to assist the transition towards sustainable energy sources. To ensure its efficient and safe use, it is mandatory to have a thorough understanding of the subsurface and relevant pyhsico-chemical processes, along with the capabilities of addressing and quantifying related uncertainties of the material properties. However, to carry out such an assessment is computationally challenging because of the need to resolve models with higher resolutions in space and time and the desire to consider nonlinear processes described by coupled partial differential equations. Machine learning methods have gained popularity for the construction of surrogate models, which facilitate to address these computational challenges. Nevertheless, machine learning also encounters major challenges in producing explainable and rigorous models as required in the field of geosciences, especially in areas where we need to provide predictions. In this work we present how the non-intrusive reduced basis method can effectively address the aforementioned challenges when applied to complex coupled nonlinear multi-physics applications. In a nutshell, the non-intrusive reduced basis method is a hybrid approach that combines elements of physics-based and data-driven methods, thereby mitigating the limitations of each individual approach. Throughout the paper, we rely on a designated geothermal case study in Northeast Germany and compare our approach against more classical data-based approaches. We further discuss how the obtained surrogate model can be used for intensive parameter investigations in the form of global sensitivity analyses and uncertainty quantification.