A comparative study of heterogeneous ensemble methods for the identification of geological lithofacies

Mudstone reservoirs demand accurate information about subsurface lithofacies for field development and production. Normally, quantitative lithofacies modeling is performed using well logs data to identify subsurface lithofacies. Well logs data, recorded from these unconventional mudstone formations, are complex in nature. Therefore, identification of lithofacies, using conventional interpretation techniques, is a challenging task. Several data-driven machine learning models have been proposed in the literature to recognize mudstone lithofacies. Recently, heterogeneous ensemble methods (HEMs) have emerged as robust, more reliable and accurate intelligent techniques for solving pattern recognition problems. In this paper, two HEMs, namely voting and stacking, ensembles have been applied for the quantitative modeling of mudstone lithofacies using Kansas oil-field data. The prediction performance of HEMs is also compared with four state-of-the-art classifiers, namely support vector machine, multilayer perceptron, gradient boosting, and random forest. Moreover, the contribution of each well logs on the prediction performance of classifiers has been analyzed using the Relief algorithm. Further, validation curve and grid search techniques have also been applied to obtain valid search ranges and optimum values for HEM parameters. The comparison of the test results confirms the superiority of stacking ensemble over all the above-mentioned paradigms applied in the paper for lithofacies modeling. This research work is specially designed to evaluate worst- to best-case scenarios in lithofacies modeling. Prediction accuracy of individual facies has also been determined, and maximum overall prediction accuracy is obtained using stacking ensemble.