Enhancing Reservoir Characterization With High-Resolution Lithofacies Prediction Using Advanced Feature Engineering and Ensemble Classifiers

Lithofacies identification is essential for understanding sedimentary basins, especially in the Cretaceous Lumshiwal Formation of the Kohat Sub-basin, where data scarcity and complex geological structures pose significant challenges. The Kohat Sub-basin, which is part of the tectonically active western Kohat Fold and Thrust Belt, exhibits steep folds and faulting. Therefore, integrated geological and geophysical approaches are essential for effective reservoir characterization. In this study, machine-learning (ML) algorithms are used to enhance the resolution of a sparse data set obtained within a complex geophysical environment. As a result, high-resolution techniques are applied to effectively tackle the challenges faced in highly complex regimes. This study utilizes advanced ML algorithms, specifically the Extra Tree (ET) and Extreme Gradient Boosting (XGB) classifiers, to improve lithofacies prediction within the Kohat Sub-basin. Utilizing well-log data from six wells, including parameters such as gamma ray (GR), neutron porosity (NPHI), resistivity, caliper (CALI), density (RHOB), spontaneous potential, and microspherically focused log (MSFL), the ML models were trained for accurate lithofacies prediction. Red dots denote values outside of the interquartile range (IQR), while heat maps and box plots were utilized to visualize the data and find outliers. The Lumshiwal Formation’s facies were divided into three categories: gas sand, wet sand, and shale, using cutoff values for the Vshale and Sw logs for each well. The agreement between measured and projected lithofacies is shown by confusion matrices. The precision with which lithofacies might be predicted by ML methods was proven through numerical experiments. With an accuracy of almost 94%, the ET classifier beat the XGB classifier. Particularly in heterogeneous reservoirs with different lithofacies, the ET classifier proved to be highly reliable, with lithofacies prediction accuracy ranging from 93 to 95%. The findings demonstrate how advanced ML techniques can enhance reservoir characterization in brownfields with little data, such as the Kohat Sub-basin. This provides insightful information for enhancing exploration tactics and efficient reservoir management.