Machine Learning for Input Parameter Estimation in Geothermal Reservoir Modeling

Abstract

One of challenges in geothermal development is its uncertainty of estimation of complicated reservoir structures. For instance, permeability varies in several orders, and it is difficult to determine the distributions in a reservoir model. Although there are sophisticated inverse analysis methods (e.g., iTOUGH 2), people often determine the permeability distributions by trial and error according to their experiences. In the early stages of development, the estimation based on people's trial and errors are sometime quicker and more effective. If the process of permeability estimation is automated like the people's intuition, it would be useful to reduce uncertainty of modeling as well as to reduce simulation time and costs. In this study, we proposed a method to estimate permeability distributions by using measurement data based on machine learning. Several permeability distributions were given to numerical simulator, TOUGH2, which generated the temperature and the pressure data as synthetic data. Combinations between the permeability and the temperature/pressure data were studied by support vector machine (SVM). The results shows the feasibility of estimating permeability distributions based on machine learning.