One of the controlling elements for starting and spreading hydraulic fractures in shale reservoirs is the presence of natural fractures. Thus, understanding the spatial distribution of natural fractures in these unconventional reservoirs is critical for hydraulic simulation and stimulation testing, and reservoir development. This paper develops a workflow for modeling the spatial distribution of natural fractures in a shale reservoir, using machine learning and geostatistical methods. This study focuses on the Hydraulic Fracturing Test Site 1 (HFTS1), located in the Midland Basin.
Machine learning was utilized in the first part of this study to create a model that can be used to predict the presence of natural fractures along wells. The model inputs are well logs, and the model output is the natural fracture presence or absence as derived from core. For the Upper Wolfcamp and Middle Wolfcamp formations, five machine learning algorithms were explored, and two model cases were created and tested. In the second part of this study, a natural fracture presence volume was developed using a property modeling workflow. The predicted natural fracture values were upscaled to meet the grid resolution, variogram analysis was used to identify the spatial variability in the dataset, and geostatistical simulation was used to populate the fracture presence or absence across the entire 3D grid.
The results show that the "best" models were developed using the random forest algorithm, and have an F1-Score range of 0.8 to 0.9 in Case 1 and 0.7 to 0.84 in Case 2. These F1-Scores indicate good model performance as the highest attainable value is 1, which indicates perfect performance. Additionally, using geostatistical modeling, the natural fracture predictions along the wells were incorporated with existing well data to develop a 3D representation of natural fracture presence within this unconventional reservoir.
The findings of this study show that we can estimate the spatial distribution of natural fractures across the reservoir, using data from a limited interval such as core. This project's findings can be utilized to constrain a discrete fracture network model for hydraulic modeling and stimulation tool testing. The methodology applied in this research can be replicated or modified for modeling natural fractures in other shale reservoirs.
