Rock Facies Characterization Using Machine Learning Algorithms

In the upstream field of exploration and production of hydrocarbons, the characterization of rock facies is critical for estimating rock physical properties, such as porosity and permeability, and for reservoir detection and simulation. The precise identification of rock properties is closely related to the net pay thickness determination of reservoirs, and is thus a definitive factor in the drilling decision-making process. In this dissertation, I applied five different machine learning algorithms to characterize rock facies with various techniques and strategies using a field dataset. The input dataset is acquired from the Panoma gas field in southwest Kansas. It contains five wireline logs and two geological indicators with a corresponding rock facies type at a sampling interval of half a foot. The total dataset has nine different rock facies. Besides exact facies there are also adjacent facies (facies with similar features that can be considered as the same one) used in classification tasks. The machine learning algorithms I used are Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Decision Trees, Artificial Neural Network (ANN), and Convolutional Neural Network (CNN). Under each methodology, I also tested the method with various techniques including feature engineering, regularization, and padding strategy, in order to improve the final results. I created two benchmark datasets for training and predicting, respectively, and all the algorithms are trained and predicted on the same dataset. The results were evaluated by the F-1 score, where F-1 is a metric used to quantify the average prediction accuracy. The results show that, for classifying exact rock facies, the Decision Trees had the best performance with a 0.66 F-1 score. For classifying adjacent facies, KNN with feature engineering technique achieved the highest F-1 score, 0.94. For each specific facies, facies #9 (Phylloidalgal bafflestone) has the highest averaged F-1 score. Facies #5 (mudstone-limestone) has the lowest averaged F-1 score which indicates it is the hardest facies to classify. My results indicate that machine learning algorithms have great application potential in automatic rock facies characterization with high accuracy and efficiency. It could significantly enhance the rock physical property estimation process and meanwhile reduce manual effort.Earth and Atmospheric Sciences, Department o