Comparing Facies Prediction Performance of Machine Learning Models Trained on Well and Core Data: A Case Study from Lower Indus Basin

Abstract

Machine learning has excellent potential to predict rock types and depositional trends at a sub-centimeter scale using borehole data in oil and gas wells. The required dataset includes core plug extracted from the wells and well-log data acquired through different tools in run in boreholes from six wells in the lower Indus basin. Core plugs are the only subsurface data that is true to geologic scale and inherent heterogeneity. The research employs a rock-type driven labeling scheme and a rock depositional process focused classification scheme to interpret the training data from core plugs at a sub-centimeter scale. To generate predictions for lithology and facies, an “RGB log” (RGBL) is developed to summarize the core plug image at each depth step. The use of RGBL data has generated even more accurate results and requires far less computing power than core image data. On the other hand, it is anticipated that well-log data will continue to be inadequate in predicting rock types or depositional trends at the sub-centimeter level due to logging speed and step interval. To overcome this challenge, multiple curves are used as inputs for activation functions to predict rock types from well-log data with signatures of encountered rock types. The study demonstrates the potential to transform large quantities of photographed core into a normalized digital format for geologic insights. The methodology involves a machine learning workflow developed in python; employed for the analysis of core image data in a scalable and reproducible manner. This approach can be extended to other geologic basins with similar clastic depositional trends, provided there is an abundance of photographed core plugs. RandomForest and GradientBoosting were used to estimate the facies using well log data; RandomForest was slightly higher in accuracy at 87.1% compared to GradientBoosting's 85.7%. Using RGB log data, MLP-SVM predicted facies with an overall accuracy of 92.31%, with metrics for precision, recall, and F1 scores of 0.96, 0.93, and 0.94, respectively.