73 Comparing Fracture Development From Multi-Year Operations at the Utah Frontier Observatory for Research in Geothermal Energy

Faculty Mentor: Kristine Pankow (Geology and Geophysics, University of Utah)

The Utah Frontier Observatory for Research in Geothermal Energy (Utah FORGE) is a field scale laboratory near Milford, Utah designed to help de-risk enhanced geothermal systems (EGS). During EGS operations, pressurized fluids are injected to create a permeable reservoir. This process generates microseismicity. In 2024, Stages 3R through 10 of the injection well were stimulated using both proppants and slick water. Microseismicity maps into two main fracture zones (Niemz et. al., 2025). This study aims to characterize the internal nature of these fracture zones in order to better understand the risk potential of the fractures and to better inform future activities at the Utah FORGE site. In this study, microseismic events are clustered using an unsupervised machine learning approach called a Bayesian Gaussian Mixture Model (BGMM) and fracture planes are estimated from these clusters by performing Principal Component Analysis (PCA) (Karvounis et. al., 2023). Bayesian Gaussian Mixture Models and the more popular K-Means clustering method are both unsupervised clustering algorithms which adjust the defined clusters through numerous iterations until the results converge to a set of output clusters. However, BGMMs diverge from K-Means by inferring the ideal number of clusters to describe the input data, describing such clusters using Gaussians, and probabilistically assigning points to clusters, thus providing generally more robust clustering for data clouds where the clusters may be close together or overlapping, as is the case within the two main fracture zones being studied here. One issue with BGMM clustering and other machine learning clustering methods is a dependence on input order of the data. To address this, random sorting of the event catalog is implemented, and the BGMM is run multiple times. After clusters have been identified, PCA is used to fit planes through the identified clusters, yielding approximate fracture planes. Building off the approach developed by Karvounis et. al. (2023), eigenvalues are then computed, allowing fracture orientations and degrees of planarity to be quantified to determine whether these planes fit the data well. Computations are also run to determine the strike, dip, and area of the planes described. The results from the multiple runs are assessed to determine the variability in clustering and to characterize the range of potential fracture planes defined by the microseismicity. To develop the above workflow, this analysis has been performed on a subset of the 2024 stimulation microseismic catalog, focusing on Stage 3R and running 100 computations of the clustering algorithm. By comparing the most commonly obtained fracture planes and their quantifiable parameters, greater confidence can be had in the results of this process (Van Fleet et. al., 2026). The next step in the process is to apply this workflow to the full 2024 stimulation microseismic catalog (Niemz et. al., 2025) as well as to the microseismic catalogs from the 2022 stimulation (Dyer et. al., 2023) and the 2023 circulation (Niemz et. al., 2024). By comparing the results of this analysis across the multi-year operations at Utah FORGE to each other as well as to the results of other studies such as Rutledge et. al. (2025), the fracture system at Utah FORGE can hopefully be better understood, aiding in both reservoir engineering and earthquake hazard monitoring as operations move forward.

Bibliography