Discovering Signatures of Hidden Geothermal Resources Based on Unsupervised Learning

Abstract

Rapid advancements in the field of machine learning (ML) offer a substantial opportunity to accelerate discovery and reduce the costs associated with geothermal exploration, development, and production lifecycle. Application of new and innovative ML methods to multi-source and multi-physics datasets may lead to the discovery of new signatures or play fairway types for the detection of hidden geothermal resources. In our ML for geothermal exploration research, one of our goals is to discover the signatures (features) characterizing geothermal resources and favorable exploration sites from the regional-scale geothermal datasets. To achieve this goal, we have applied an unsupervised ML method to extract latent/hidden features or signals from the regional-scale geothermal data for geothermal resource exploration. The data describe the known-geothermal resources in southwest New Mexico. The unsupervised ML is based on a non-negative matrix factorization (NMF) method coupled with a custom semi-supervised k-means clustering algorithm. Our methodology, called NMFk, is capable of identifying latent/hidden signals, an optimal number of clusters, and a dominant set of features hidden in the large-scale geothermal datasets. Based on our NMFk analyses and associating the obtained ML results with site information (e.g., regional physiographic provinces), the optimal number of clusters identified is equal to 4. The dominant set of attributes, among a total of 22 geothermal attributes, that were identified through NMFk analysis include air temperature, gravity, depth to water table, elevation, crustal thickness, drainage, and lithium concentration. These dominant attributes in the geothermal data indicate favorable sources of data collection to explore geothermal resources in each province (e.g., The Rio Grande Rift, the Mogollon-Dalit volcanic field). Moreover, the proposed NMFk method is widely applicable to extract features/signals from large-scale geothermal data (including observational and simulation outputs). This broad applicability of our ML tools makes it attractive to discover, quantify, and assess hidden geothermal energy resources (e.g., meeting DOE-EERE GTO’s mission).