Abstract
The accurate classification of reservoir facies remains a fundamental challenge in petroleum geoscience, with significant implications for resource extraction efficiency and reservoir characterization. Traditional approaches relying on manual interpretation and conventional machine learning methods often struggle with the complexity and heterogeneity of well-log data. This architectural approach, in contrast to traditional single-stream or non-residual designs, significantly enhances the model's ability to concentrate on key geological features while preserving hierarchical representations of the data. Consequently, it more effectively addresses data heterogeneity and contextual dependencies. The framework was trained and evaluated using measurements from eight wells that represent diverse geological settings. Comparative experiments against conventional machine learning models and state-of-the-art deep learning techniques demonstrated the superiority of our method, achieving an area under the receiver operating characteristic curve (AUROC) of 0.883 and an area under the precision-recall curve (AUPRC) of 0.502. These enhancements enable more accurate identification of subtle facies boundaries and lithological variations, particularly in complex geological formations, thereby facilitating improved reservoir delineation and reducing uncertainty in field development planning. Furthermore, reproducibility analyses confirmed consistent performance across independent trials, underscoring the model's robustness and its viability for real-world reservoir characterization workflows.