Application of Machine Learning to Create a Discrete Fracture Network Model for Utah FORGE Fracture Injections

Abstract

Abstract A method to process microseismic event locations from three injections into the Utah FORGE 16A(78)-32 geothermal well has been developed as part of the 2023 SPE Geothermal Datathon. One objective of the datathon was to develop methods using a few tunable parameters that are capable of multiple realizations of the Discrete Fracture Network (DFN). The method uses open-source software tools and comprises seven steps. The first step is to calculate the square-root of elapsed time from the first event of each stage. The next step is to use DBSCAN (Density Based Spatial Clustering of Applications with Noise) on this RootTime variable, followed by the application of DBSCAN to the spatial variables in each time slice. Each of the resulting clusters is analyzed by principal component analysis to generate fracture planes. DBSCAN leaves multiple outliers that are then harvested using two methods. Criteria are provided to fuse fractures together that are close spatially. The final step is to consider if connective fractures are required to ensure communication of the fracture network with the perforated interval. The Utah FORGE dataset comprises 2798 event locations from three injections. The analysis in time yielded 54 clusters of data, and the spatial analysis then provided 73 distinct fractures, with a residue of 25% outliers. Outliers were harvested in two steps: first, capturing outliers that were adjacent to mapped fractures, and then evaluating the remaining outliers for individual fracture planes using relaxed DBSCAN parameters. After these two steps, the outlier population was reduced to less than 4%, and the total number of mapped fractures grew to 87. It was recognized that fractures can propagate across time slices, so a fracture fusion step was conceived to combine subparallel fractures that were indistinguishable from each other based on error analysis. This was particularly necessary for Stage 3 that had mostly vertical fractures. In this step, 24 fractures were combined, resulting in a total of 63 fractures in the DFN. In the final step, it was recognized that there were no fracture intersections with the perforated interval for Stage 2, and thus an aseismic flow path was inferred. A vertical and a horizontal fracture were inserted to represent this flow. Each DBSCAN application has two input parameters, resulting in possibly many clusters and multiple outliers. The development of steps to harvest outliers and fuse adjacent fractures were conceived to utilize as much data as possible and to recognize the relative errors in event locations. With regards to the Datathon goal of achieving an automated processing sequence, the algorithm runs without manual intervention once the user has chosen four parameters for each stage: the minimum number of points in a cluster and the accepted percentage of outliers for each of the time and spatial clustering steps. The calculated dominant fracture azimuth of N-20-E compares favorably with data from the field, providing some indication of the quality of the results.