Final Technical Report - Applications of Machine Learning Techniques to Geothermal Play Fairway Analysis in the Great Basin Region, Nevada
Geothermal play fairway analysis (PFA) is a methodology that integrates geological, geophysical, and geochemical parameters to produce geothermal potential maps and identify promising areas for hidden systems. Although a previous PFA was successful in identifying new geothermal systems in Nevada, key challenges affected the PFA, including the estimation of parameter weights, a limited number of training sites, and limitations on some datasets. In this project, we applied machine learning (ML) techniques, including both supervised and unsupervised approaches, to ~1/3 of Nevada to mitigate those limitations and better characterize signatures for subsurface permeability related to hidden systems. Major project activities and tasks included: 1) review and compilation of regional datasets; 2) selection of positive and negative training sites; 3) transformations of datasets for ML applications; 4) development and testing of ML techniques; 5) data dissemination; and 6) identifying future data needs. Fourteen datasets were incorporated into this study. Favorable structural settings; location, age, and slip rates on Quaternary faults; geodetic strain rates; and earthquake density were adapted directly from the PFA project. Datasets utilized in the PFA study but revised and/or enhanced for this project included slip and dilation tendency on Quaternary faults, gravity, heat flow, fluid geochemistry, and well-spring temperatures. New features incorporated into this study were magnetic data, paleo-geothermal deposits, and a digital elevation model. To optimize the application of ML techniques, the threshold for positive training sites was reduced from known temperatures of ≥130°C used in the PFA project to sites with known temperatures ≥37°C that have potential for hosting moderate to high temperature (≥130°C) reservoirs, which netted 83 positive sites. In addition, 62 negative training sites were defined from relatively deep (>1 km in the west and >2 km in the east due to a regional carbonate aquifer) and cool wells that have temperature gradients less than the regional average. Transformation of the datasets to formats enabling ML applications involved: 1) converting all datasets to a common grid (e.g., defined geospatial reference system and raster cell size) and machine-readable tabular format (.csv); 2) engineering of geographically continuous feature sets via weights-of-evidence analysis, distance-to-feature gridding, and statistical methods for interpolation; 3) conversion to and from high-resolution xyz files to ensure portability between map-based platforms (e.g., ArcGIS) and non-map-based platforms (e.g., R language statistical programs, Python-based ML tool kits); 4) quality-assurance assessments, and 5) conversion of data to Python Pandas data frames to allow for efficient numerical transformation, filtering, and selection. The ML algorithms and code were formatted as Python-based Jupyter Notebooks and made available via the Geothermal Data Repository, then linked to the DOE CODE platform. The development and testing of ML techniques was the focus of this project. Initially, ML was applied to the original PFA datasets to test and verify the previous results, as well as execute a baseline analysis of performance of prediction errors in the training set. ML was then applied utilizing the TensorFlow model, enhanced and new datasets, and training sites. Two approaches produced promising results: 1) supervised Bayesian probabilistic neural networks (BNN) that generated geothermal potential maps with confidence intervals, and 2) unsupervised principal component analysis paired with k-means clustering (PCAk) that generated both cluster maps to help identify spatial patterns, as well as new combined feature inputs. Two primary feature sets were analyzed. The first comprises a set of regional- and intermediate-scale permeability and heat data and illustrates a promising design that augments the original permeability modeling in the PFA. The second feature set includes the same regional and intermediate feature layers as in feature set one but with the addition of local permeability features, as defined by favorable structural settings. The second feature set illustrates how a model design may find a balance between disparate data types to produce predictive favorability maps that yield a dynamic analog to the original PFA model. Through comparative analysis of both feature sets, the BNN and PCAk techniques were combined to guide exploration. For example, in both the feature set one map and previous regional permeability model, prominent fairway belts stand out in the western parts of the Walker Lane and central Nevada seismic belt, reflecting high strain rates, active faults, abundant earthquakes, and steep gravity gradients. Many bands of high favorability in the PFA regional permeability model lie along main segments of major normal faults, owing to steep gravity gradients. In the BNN analysis, many of the same segments exist at high percentiles (e.g., 50th percentile), but the predictions at a higher degree of confidence (i.e., 5th-percentile map at the 95th percentile of model scores) form more localized zones of favorability, particularly at the ends and discontinuities of major Quaternary normal faults. These models are consistent with the observation that hydrothermal systems are generally located with discontinuities of major normal faults (e.g., fault tips or step-overs), as opposed to the mid-segments of such faults. Notably, however, the BNN model for feature-set two generally performed better with precision and identifying true positives relative to feature-set one. Complementary PCAk helped to constrain previously unrecognized feature controls on geothermal favorability. For example, geodetic strain rate and Quaternary fault distance were most important for feature-set one. For feature-set two, geodetic strain rate and favorable structural settings had the greatest influence. Conversely, in both feature sets, fault recency, fault slip and dilation tendency, and elevation appear less relevant to the classification of positive and negative sites. Overall, the ML analyses indicate that highly prospective areas are most common in the western part of the study area, commonly corresponding to favorable sites identified in the Nevada PFA. However, the importance of certain datasets as a proxy for geothermal activity may vary between regions. Major differences in strain rates, active faulting, and regional aquifers may warrant modification of the groupings and weightings of various features in the PFA approach, possibly requiring customizing the approach for each region. More detailed analyses that reduce exploration risks in the future will be partly dependent upon greater coverage of high-resolution LiDAR and detailed geophysical data (e.g., gravity, magnetics, and magnetotellurics). In addition, more work is needed in characterizing the signature of favorable structural settings in the potential field data.
- Research Organization:
- Nevada Bureau of Mines and Geology, University of Nevada, Reno
- Sponsoring Organization:
- USDOE Office of Energy Efficiency and Renewable Energy (EERE), Renewable Power Office. Geothermal Technologies Office
- Contributing Organization:
- Ormat Nevada, Inc.
- DOE Contract Number:
- EE0008762
- OSTI ID:
- 2335471
- Report Number(s):
- DOE-UNR-8762-FTR
- Country of Publication:
- United States
- Language:
- English
Similar Records
Integrated Geologic and Geophysical Approach for Establishing Geothermal Play Fairways and Discovering Blind Geothermal Systems in the Great Basin Region, Western USA: A Progress Report
Integrated Geologic and Geophysical Approach for Establishing Geothermal Play Fairways and Discovering Blind Geothermal Systems in the Great Basin Region, Western USA: A Progress Report