Identifying regions Of hydrothermal upflow In the Great Basin using supervised machine learning

Abstract

The U.S. Geological Survey (USGS) is updating the conventional hydrothermal resource assessment for the Great Basin. As part of this updated assessment, a hydrothermal favorability map is produced to show where geologic characteristics are favorable for hydrothermal resources, thereby allowing the use of conditional statistics (i.e., the likelihood of a resource being present in an area based on its predicted favorability and extent of exploration) to estimate the total (i.e., discovered and undiscovered) geothermal resources. If geologic characteristics vary across the Great Basin, then the factors controlling hydrothermal systems likely vary as well. Therefore, dividing the region based on the geologic characteristics that influence hydrothermal circulation could lead to a more accurate estimate of total geothermal resources. Geoscientists have previously developed an understanding of the regions within the Great Basin with marked differences in behavior (e.g., structurally, hydrologically, and geothermally) using expert knowledge and/or unsupervised machine learning analyses. The unsupervised analyses (e.g., principal component analysis) have been used to identify groups of data with similar behavior and describe characteristics of the regions. Use of only unsupervised methods fundamentally answers the question which areas are different?, after which expert input can identify if these differences are meaningful for predicting a condition of interest. To reduce the need for post-analysis expert input, the use of a supervised method explicitly incorporates what one is trying to predict (i.e., the known resources) and answers the question which areas have differences that are important for predicting a condition of interest? Herein, hydrothermal upflow is the condition of interest, and a supervised method is tested to delineate regions within the Great Basin based on geologic characteristics that promote or impede hydrothermal upflow. We group hydrothermal systems and their surrounding areas in the Great Basin based on similar decision tree paths using Extremely Randomized Trees (ExtraTrees), a supervised machine learning algorithm. Tree-based machine learning methods are a series of branches through which choices are made (e.g., heat flow above or below a specified magnitude) to categorize training sites. Sites that follow the same paths through the decision trees (down the same branches) are similar in a quantifiable way. This approach identifies nine distinct regions with varying resource characteristics (e.g., frequency of occurrence and reservoir properties). This workflow delineating regions by the characteristics that control hydrothermal upflow allows for a region-by-region estimation of undiscovered hydrothermal resources through conditional statistics and thus a more precise estimation of the total geothermal resources within the Great Basin.