Physics-Informed Lithology Identification From Well Logs via Enhanced Adaptive Synthetic Sampling Balancing and Sparrow Search Optimization

Summary

To enhance the accuracy and generalization of well-log-based lithology classification under water-based mud (WBM) conditions, an integrated framework is proposed by combining physics-informed composite feature construction, class balancing based on enhanced adaptive synthetic sampling (Enhanced-ADASYN), an improved variant of adaptive synthetic sampling (ADASYN), and metaheuristic hyperparameter optimization via the Sparrow Search Algorithm (SSA). The framework couples physical interpretation with adaptive data-driven learning. First, physics-informed composite features are constructed from the contrast between deep-investigation AT90 and shallow-investigation AT20 resistivity responses to characterize mud-invasion-related resistivity differences under WBM conditions. These features provide physically interpretable auxiliary information for lithology discrimination by incorporating the depth-of-investigation contrast between resistivity logs. Then, the proposed data-balancing strategy, an Enhanced-ADASYN algorithm integrating exponential ratio adjustment and feature-weighting mechanisms, adaptively balances minority lithology classes and improves the feature-direction consistency of synthetic sample generation. Additionally, SSA-based metaheuristic optimization is introduced to adaptively tune random forest (RF) hyperparameters. In the main test-set comparison, the proposed integrated framework achieved improved overall performance across the reported metrics compared with the baseline RF model. Specifically, the precision, recall, F1-score, and Kappa-score (K-score) improved from 0.821, 0.825, 0.817, and 0.803 to 0.8469, 0.8496, 0.8472, and 0.8316, respectively, corresponding to absolute improvements of 2.46–3.02 percentage points. Furthermore, validation on an independent test data set confirmed the generalization capability of the proposed framework, with absolute improvements of 0.77–1.74 percentage points across the reported metrics compared with the baseline RF model. The repeated ablation results further indicate that the selected BD composite-feature combination, consisting of the deep-shallow resistivity difference and the resistivity amplitude ratio, is the dominant contributor to the performance gain, while Enhanced-ADASYN and the SSA-based metaheuristic module provide complementary effects.