The study focuses on the use of data analytics to determine optimal locations to place fractures along the wellbore. The petrophysical data of the formation, such as mineralogy, total organic carbon (TOC), ultrasonic velocity, Young's modulus, Poisson's ratio, and creep displacements, are used to define different clusters. These clusters were then ranked in order of increasing fracturing potential to characterize the formation. The data available to this study included petrophysical measurements at 660 sampled locations in 20 wells in the Eagle Ford, Wolfcamp, and Woodford formations.
In the first step, different clusters are identified based on four key variables, namely TOC, clays, Young's modulus and Poisson's ratio using unsupervised clustering algorithms like K-means and hierarchical clustering. Four different clusters were identified and ranked in order of their fracturing potential. The other measurements, such anisotropy parameters (ε and λ) and creep parameters, not directly used in the clustering process were also incorporated in ranking the clusters. Cluster 1 was rich in clays (51 wt%), had a high Poisson's ratio (0.29), and anisotropy (ε = 0.32). This cluster was deemed ductile and unsuitable for fracturing. Cluster 2 was rich in TOC (7.2 wt%), had a high Poisson's ratio (0.26) and anisotropy (ε = 0.25). This cluster was also deemed ductile and unsuitable for fracturing. Cluster 3 had low TOC (1.6 wt%) and clays (11 wt%) but high Poisson's ratio (0.26) and a very high Young's modulus (72 GPa). This cluster was also rich in carbonate minerals. This cluster would require high energy to break and therefore deemed unsuitable for fracturing. Cluster 4 had high quartz content (36 wt%), but low TOC (2.2 wt%), Young's modulus (54 GPa), Poisson's ratio (0.19), anisotropy (ε = 0.1) and creep. This cluster was deemed most suitable for fracturing.
Next, since core data are generally available in a select few wells while log data are available in a much larger set of wells, the clusters were upscaled using the available logs (gamma ray, neutron, and density). Three different classification techniques, namely decision trees, gradient boosting and support vector machines (SVM), were applied and compared. The gradient-boosting technique gave the minimum error and was finally selected to predict the clusters at the log level. Ibis study also shows examples from both vertical and horizontal wells where this workflow was applied to identify optimal fracture locations.
