A Comparative Study of Three Supervised Machine-Learning Algorithms for Classifying Carbonate Vuggy Facies in the Kansas Arbuckle Formation

Diagenetic features, such as vugs, fractures and dolomite bodies can have significant impacts on carbonate reservoir quality. Challenges remain in characterizing these diagenetic features from well logs, as they are often mixed with changes in mineral and fluid concentrations. In this paper, a data-driven approach is developed to classify vuggy facics bascd on core and well logs from a key well penetrating the Arbuckle formation in Kansas. Three supervised machine-learning methods, namely artificial neural networks (ANN), support vector machines (SVM), and random forests (RF), are compared for their accuracy, stability, and computational efficiency. Hyperparameters are tuned using cross-validation and Bayesian optimization. Different feature selection methods and data labeling schemes arc also evaluated to optimize the prediction. Results indicate predicting a binary classification (vuggy/nonvuggy) presents an-80% accuracy, compared to a 65% accuracy using a five-class vug-size-based classification label. A direct input of well logs as training features is recommended instead of using derived petrophysical properties. Among the three machine-learning algorithms, ANN outperforms the other two methods for vug/nonvug detection, whereas for vug-size classification, RF is the best algorithm to apply. This work also suggests RF shows the least sensitivity to hyperparameters (i.e., maximum number of splits and minimum leaf sizes) according to the response surfaccs constructed via Bayesian optimization. For the dataset used in this study. SVM is the most computationally efficient algorithm.