AE-TabNet: A Hierarchical Class Aggregation Approach for Cross-Dataset Reservoir Classification Using Well Log Data

Abstract

Identifying reservoir fluid types is a critical task in petroleum exploration and georesource management and is commonly performed through the interpretation of well log data. However, this task remains challenging due to the complex, nonlinear relationships among logging measurements and its strong dependence on expert knowledge, which often results in subjective and inconsistent interpretations. To reduce reliance on manual analysis and improve classification robustness, this research proposes a hybrid Deep Learning (DL) framework for reservoir fluid classification. Specifically, an integrated AE-TabNet model is developed by integrating an Autoencoder (AE) with a Tabular Network (TabNet). This integration enhances feature learning by generating more informative feature representations prior to classification. The proposed framework is benchmarked against conventional Machine Learning (ML) models, including Decision Tree (DT), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and the standalone TabNet model. The hyperparameters of all models are optimized using Particle Swarm Optimization (PSO). Performance is evaluated using well log data obtained from the Kansas Geological Survey (KGS). The proposed AE-TabNet model with PSO-optimized hyperparameters achieves an accuracy of 95% in classifying four reservoir fluid types (gas, oil, water, and non-reservoir) in the CLINE oil lease dataset. To mitigate the reliance on additional labeled data and enhance cross-field generalization, a hierarchical class aggregation approach is further incorporated. By consolidating related classes within a hierarchical framework, the proposed approach enables effective classification of previously unseen well log datasets from the SNYDER and STRAHM oil leases without requiring model retraining. The approach achieves classification accuracy of up to 90%, while reducing the need for additional data labeling and, consequently, the cost associated with model deployment.