Lithological Classification from Well Logs using Machine Learning Algorithms

Abstract

Well logging is a widely used geophysical method to gather information from subsurface rocks and establish lithological classification. However, the criteria of lithological classification are loosely defined and human error can significantly contribute to the uncertainty of the interpretation. This study uses machine learning approaches to classify rock types from well logs of the Snake River Plain (SRP) in Idaho. To achieve the comprehensive results, three machine learning algorithms, K-nearest neighbour (K-nn), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGB) are employed on fifteen types of well logs from four geothermal wells in the SRP under three experimental conditions. In the first experiment, the classifiers are trained and tested with data from the same well in the train:test ratio of 7:3. The second scenario assigns data from three wells as a training subset and the remaining well as test subset. The third experiment uses the largest amount of data as a training subset, which combines data from three wells and 70% of the data from the remaining well. Hyperparameters in all classifiers are optimized to enhance model performance. Results suggest that SVM and K-nn exhibit comparable performance in all experiments, resulting in 89.68% (s = 10.40) and 88.84% (s = 9.92) of average accuracy, respectively. XGB shows the highest prediction accuracy in this study with average prediction accuracy at 90.67% (s = 8.21). This is largely because XGB partitions data into subgroups based on available features iteratively until every class is clearly separated from each other. In addition, XGB can recognize missing values in well logs and does not use these values for classification. XGB further indicates that gamma ray, neutron, and temperature are the top three important features that are used to improve the prediction accuracy.