High Resolution Image Download MS PowerPoint Slide Precise lithology identification serves as the cornerstone for characterizing complex metamorphic basement reservoirs and delineating favorable fractured intervals, yet it remains a formidable challenge in offshore environments, where coring data are scarce due to high acquisition costs. This study establishes a novel, machine-learning-driven lithological identification framework for the J Oilfield in the Bohai Bay Basin, targeting a complex reservoir comprising migmatitic gneiss, plagioclase migmatitic granite, monzonitic migmatitic granite, diorite porphyrite, and tectonic breccia. Integrating elemental and conventional logging data with core-derived lithologies as training samples, we propose a hybrid classification approach utilizing linear discriminant analysis (LDA) coupled with a supervised self-organizing map (SSOM). In this workflow, LDA is employed to extract optimal feature vectors that maximize between-class variance. By identifying elements and log responses that exhibit high discriminant weightings while maintaining a low mutual correlation, the dimensionality of the feature space is effectively reduced. These optimized features serve as inputs for the SSOM classifier, utilizing core lithologies to provide supervised constraints. This model further captures the nonlinear mapping between the integrated log responses of elemental and conventional well logging and lithology membership. Validation results indicate that the LDA-assisted SSOM model achieves an overall accuracy of 81.4% and a blind test accuracy of 75.0%, significantly outperforming methods relying solely on ECS or conventional logs. This study demonstrates that the proposed supervised LDA-assisted SSOM workflow provides a robust solution for precise lithology prediction in metamorphic basement reservoirs through the integration of ECS and conventional well log data.
Wang et al. (Wed,) studied this question.