Haleh Azizi

Date of Award

January 2021

Document Type


Degree Name

Master of Science (MS)


Computer Science

First Advisor

Hassan Reza


Oil reservoirs are divided into three categories: carbonate (fractured), sandstone and unconventional reservoirs. Identification and modeling of fractures in fractured reservoirs are so important due to geomechanical issues, fluid flood simulation and enhanced oil recovery.Image and petrophysical logs are individual tools, run inside oil wells, to achieve physical characteristics of reservoirs, e.g. geological rock types, porosity, and permeability. Fractures could be distinguished using image logs because of their higher resolution. Image logs are an expensive and newly developed tool, so they have run in limited wells, whereas petrophysical logs are usually run inside the wells. Lack of image logs makes huge difficulties in fracture detection, as well as fracture studies. In the last decade, a few studies were done to distinguish fractured zones in oil wells, by applying data mining methods over petrophysical logs. The goal of this study was also discrimination of fractured/non-fractured zones by using machine learning techniques and petrophysical logs. To do that, interpretation of image logs was utilized to label reservoir depth of studied wells as 0 (non-fractured zone) and 1 (fractured zone). We developed four classifiers (Deep Learning, Support Vector Machine, Decision Tree, and Random Forest) and applied them to petrophysics logs to discriminate fractured/non-fractured zones. Ordered Weighted Averaging was the data fusion method that we utilized to integrate outputs of classifiers in order to achieve unique and more reliable results. Overall, the frequency of non-fractured zones is about two times of fractured zones. This leads to an imbalanced condition between two classes. Therefore, the aforementioned procedure relied on the balance/imbalance data to investigate the influence of creating a balanced situation between classes. Results showed that Random Forest and Support Vector Machines are better classifiers with above 95 percent accuracy in discrimination of fractured/non-fractured zones. Meanwhile, making a balanced situation in the wells by a higher imbalance index helps to distinguish either non-fractured or fractured zones. Through imbalance data, non-fractured zones (dominant class) could be perfectly distinguished, while a significant percentage of fractured zones were also labeled as non-fractured ones.