Well Log Analysis and Comparison of Supervised Machine Learning Algorithms for Lithofacies Identification in Pab Formation, Lower Indus Basin

Abstract

Machine learning uses higher-order statistical computation for model prediction by combining statistics with advanced algorithms. These predictions can be used to infer the characteristics of potential hydrocarbon reservoirs using well log data. The objective of this study is to predict and segregate the potential reservoir and non-reservoir facies in the Cretaceous Pab Formation based on well log analysis and machine learning techniques using the borehole data of three wells drilled in the Lower Indus Basin. The Petrophysical interpretation has been performed to mark the potential reservoir facies, namely, gas sand, wet sand, and shale using cutoff values for shale volume (Vshale) and water saturation (SW). Before any further analysis, the quality of the well log data for the wells was checked by using heat maps and box plots to identify the outliers in the available dataset. The important available featured logs were used in the Random Forest (RF) and the Decision Tree (DT) algorithms to predict the facies in each well by training the other two wells. The prediction accuracy score between the petrophysical and the predicted potential reservoir facies using RF classifier shows 92 to 94% accuracy, whereas prediction accuracy score of DT classifier is 80 to 85%. The comparison of the two techniques indicates that the predictions using RF algorithm are relatively more accurate therefore making it more efficient for predicting the potential reservoir facies in the wells where the sophisticated well logs are unavailable.