Application of supervised machine learning classification models to identify borehole breakouts in carbonate reservoirs based on conventional log data

Abstract

Breakouts provide crucial information regarding evaluating in situ stresses and verifying the geomechanical model. One of the most common methods for identifying borehole deformation, such as breakouts and drilling-induced fractures, is a combination of image logs and multiple pad calipers to identify borehole elongation. However, image logs are suitable for geomechanical studies but are usually unavailable in most drilled wells due to technical and financial reasons. Conventional wireline logs, including gamma-ray, neutron porosity, density, resistivity, and single pad caliper, on the other hand, are widely used in drilled wells. Moreover, in recent years, machine learning has been widely applied to classification and optimization problems in the geology and petroleum industry. This research investigated the possibility of predicting the occurrence of borehole breakouts. The models employed encompass K-Nearest Neighbors (KNN), Decision Tree (DT), and Random Forest (RF). The input values used in this study are 4099 subintervals from conventional logs and their corresponding identified breakouts from real data as classification labeling parameters for the same interval. The performance capacity of the classification models was evaluated based on Accuracy, Precision, Recall, and F1 Score. The results are promising, and RF classification performance is more reliable than the other models. Furthermore, the RF predictive model had Accuracy, Precision, Recall, and F1 Scores equal to 92.41%, 86.36%, 87.08%, and 86.7%, respectively, superseding the performance of the other models. Overall, this study shows that the application of machine learning classification models demonstrated a reasonable and fast performance in identifying borehole breakouts from conventional logs.