<p>Lithology identification using well-logging data plays a critical role in mineral resource exploration by offering essential insights into subsurface geological formations. As the demand for efficient and accurate exploration continues to rise, advanced identification methodologies have attracted increasing attention. Among these, deep learning-based approaches have emerged as particularly promising due to their ability to capture complex nonlinear patterns from training data, thereby enabling automated lithology identification and enhancing exploration efficiency. However, a persistent challenge lies in the limited discriminative power of global models when confronted with ambiguous lithological assemblages that exhibit highly similar logging responses. This often results in misclassifications and a subsequent decline in prediction accuracy. To address this challenge, we propose a novel lithology identification framework based on a local classifier with adaptive weights. Specifically, the method first utilizes a Gaussian Mixture Model (GMM) to detect ambiguous lithological assemblages that are prone to confusion. Then, dedicated local classifiers are constructed to improve discrimination within these assemblages by emphasizing subtle yet informative feature distinctions. Experimental results demonstrate that the proposed method significantly outperforms five benchmark machine learning models—Bagging, AdaBoost, Bi-GRU, Random Forest, and a baseline local classifier—yielding accuracy improvements of up to 15.0%, 6.2%, 8.6%, 11.6%, and 5.8%, respectively. These findings underscore the effectiveness of localized adaptive learning in enhancing lithology identification, particularly in geologically complex settings characterized by overlapping logging responses.</p>