An Igneous Rock Lithology Identification Method Based on Data-Algorithm Bidirectional-Driven Framework

Abstract

Current igneous rock lithology identification technology faces compounded technical bottlenecks arising from significant disparities between conventional logging responses and lithological sensitivity, challenges in constructing continuous stratigraphic models using discrete mineral experimental data, and class imbalance in core-logging calibration samples. To address these challenges, this study proposes a Data-Algorithm Bidirectional-Driven Framework (DABDF) for igneous lithology identification. First, a feature space optimization algorithm constrained by discrete minerals is developed, integrating mineral chemical components with log data to establish a petrophysically meaningful feature characterization framework. Subsequently, a hierarchical resampling mechanism based on Mahalanobis distance is designed to mitigate recognition bias in minority sample classes. Finally, a probabilistically interpretable Bayesian deep forest model is constructed to achieve high-precision identification of complex lithologies. Validation experiments employing a nested verification strategy were conducted using 8 356 logging data from 20 wells in the eastern sag of the Liaohe basin. The proposed method achieved 100% accuracy in intra-well testing, 89% accuracy in inter-well testing, and a weighted F1 score of 0.88, which demonstrated significant improvements over existing igneous rock lithology identification methods. This study shows that the deep integration of geological prior knowledge with deep learning effectively enhances the engineering applicability and interpretive reliability of lithology identification in igneous rock, and providing a novel technical solution for refined evaluation of complex reservoirs.

Keywords

• log interpretation
• lithology identification
• deep forest algorithm
• logging data fusion
• igneous
• class imbalance handling