33 Geochemistry, Hydrology, and Microbiology in the Yellowstone and Other Western North American Hydrothermal Systems

Friday, 17 May 2024
8:00 AM - 12:00 PM
Meeting Room 2

INTEGRATING UNCERTAINTY INTO THE HYDROTHERMAL RESOURCE FAVORABILITY MAPS FOR THE GREAT BASIN, USA

Paper number: 33-5
Program: Theme Sessions
Time: 9:20 AM – 9:40 AM

Authors

Abstract

We detail a machine learning workflow that predicts hydrothermal upflow and estimates uncertainty imparted by bias resulting from different train-test splits by using XGBoost with Monte Carlo Cross Validation (n = 120). Feature data are extracted from datasets (e.g., strain, gravity) compiled under the INnovative Geothermal Exploration through Novel Investigations Of Undiscovered Systems (INGENIOUS) project. Label data are defined as the difference between heat flow measured at each well in the study area and the background conductive estimate. That difference is estimated to be correlated to nearby convective heat flow, so a difference of zero implies no nearby convection, and larger differences being correlated to large amounts of convective heat flow. Our previous work found that the convective signals contain large outliers (3,083 signals between -91 and 1,000 mW/m2 and 192 signals in the much larger range of 1,000 and 11,105 mW/m2) that bias results for traditional regression; however, these outliers may hold valuable information for large, power-producing hydrothermal systems. Therefore, in order to minimize outlier effects, we discretize the continuous label values into bins for very low, high, and very high signals using optimal bin boundaries identified by preliminary modeling to allow for use with ordinal regression.

Because hydrothermal systems are sparse (approximately less than 0.1% of the total area), an additional criterion on the resulting models is that total area with high convective upflow should be minimized; in this respect, XGBoost with ordinal regression and the binned convective signals better predicts hydrothermal systems than XGBoost with regression and the unbinned convective signals. Well sites with low convective signals have the greatest variance in predictions. We hypothesize that the greater variance of predictions at well sites with low convective signals is a result of the biased sampling of well sites with low convective signals since many of the wells with low convective signals were initially expected to host a hydrothermal system prior to exploration drilling. Presently, we are testing this hypothesis by randomly subsampling unlabeled locations distant from convective wells as having low convective signals to augment the known locations with low convective signals.

Geological Society of America Abstracts with Programs. Vol. 56, No. 4

doi: 10.1130/abs/2024CD-399310

© Copyright 2024 The Geological Society of America (GSA), all rights reserved. Permission is hereby granted to the author(s) of this abstract to reproduce and distribute it freely, for noncommercial purposes. Permission is hereby granted to any individual scientist to download a single copy of this electronic file and reproduce up to 20 paper copies for noncommercial purposes advancing science and education, including classroom use, providing all reproductions include the complete content shown here, including the author information. All other forms of reproduction and/or transmittal are prohibited without written permission from GSA Copyright Permissions.