The western part of USA has a wide range of geothermal resources (e.g., moderate and high-temperature resources). These resources are characterized by diverse, multi-source, and multi-physics datasets. To explore such diverse datasets, a comprehensive data analyses are required and unsupervised machine learning (ML) offers a viable solution. In this study, we show results on recently developed unsupervised ML tool, NMFk, to characterize low, moderate, and high-temperature resources. NMFk couples non-negative matrix factorization and customized k-means clustering. Using the NMFk method, we can identify (1) latent signals, (2) the optimal number of these signals, (3) dominant sets of attributes to characterize the signals, and (4) spatial signatures. The signals are encapsulated in the data but cannot be observed only from data or using traditional data analysis tools. Also, these distinct signals allow us to make generalized prediction of attributes that characterize new geothermal resource. To discover, quantify, and assess hidden geothermal resource, the NMFk method has been applied to data of seven US regional-scale geothermal sites that are: (1) Tularosa Basin in New Mexico, (2) Tohatchi Hot springs in New Mexico, (3) southwest New Mexico, (4) west Texas area, (5) Brady geothermal site in Nevada, (6) Frontier Observatory for Research in Geothermal Energy (FORGE) in Utah, and (7) the Big Island in Hawaii. All data exhibit unique characteristics about a hidden geothermal resource. Also, few datasets are small and may have missing data points, while some analyzed datasets are big. However, for all datasets, the NMFk efficiently identifies hidden signals and their corresponding dominant attributes to characterize the geothermal system. For all sites, we characterize the geothermal resource better than before, predict attributes that characterize a new resource, and establish new data acquisition criteria to explore the region/basins better.
Machine Learning in Geothermal Development