A Novel Multiagent Collaborative Learning Architecture for Automatic Recognition of Mudstone Rock Facies

Abstract

Recognizing mud rock lithofacies is essential for mapping the subsurface depositional environments and identifying oil and gas-bearing rock formations. Conventional well logs interpretation techniques are slow, costly and require high domain expertise. Machine learning (ML) techniques have been implemented to automate the recognition of lithofacies from the bulk of well logs generated. However, the reservoir heterogeneity and uneven thickness of rock layers result in imbalanced data conditions that make the ML models biased. This study proposes a novel multiagent collaborative learning architecture (MCLA) to handle the imbalanced data problem during the identification of lithofacies. This research investigates four popular data resampling techniques, i.e. oversampling, SMOTE and ADASYN. Also, resampling techniques are combined with nine different ML classifiers, including Decision tree, ExtraTree, Random Forest, Logistic regression, Support vector machine, K-nearest Neighbour, Naïve Bayes and Ensemble methods. Stacking and voting ensembles combine the outcomes of diverse classifiers working as team members in MCLA. ADASYN, in combination with Stacking, has produced impressive results in terms of accuracy (99.41%) along with MCC (0.98) and G-mean (0.98). The proposed MCLA shows an enhancement of 2% in lithofacies accuracy and an approximately 4% increment in reliability compared with the top-performing Extra Tree classifier considered in this study.

Keywords

• Data imbalance
• resampling techniques
• machine learning
• stacked generalization
• multiagent collaborative learning