<p>Lithology identification plays an essential role in reservoir evaluation and characterization. Inferring lithology from well-logging curves has been widely used in subsurface characterization. However, well-logging data generally exhibit insufficient discrimination between geologically adjacent lithologies and notable distribution discrepancies across wells, which limit the classification accuracy and generalization of lithology identification models. In this study, we introduce DPA-MSCL, a novel contrastive representation learning method that enhances adaptability to unseen distributions via dynamic prototype alignment. Our method adopts supervised contrastive learning as the model framework and derives multi-scale representations from well-logging data to integrate geological information from different depth ranges. A balanced contrastive loss function is employed to reshape the representation distribution in the well-logging feature space while alleviating feature representation bias caused by class imbalance. During training, dynamically updated class prototypes are introduced to further guide the model to learn inherent lithological features. Moreover, we employ a two-stage learning paradigm that decouples representation learning from lithology classification to avoid semantic interference from the classifier. Extensive experiments on datasets from the Hugoton and Panoma fields and the China Daqing fields demonstrate that the proposed method achieves significant improvements over state-of-the-art approaches and validate its effectiveness for cross-well lithology identification.</p>