SCLRAD: Semi-supervised contrastive learning using random replacement of adjacent depths for lithology identification

• Published in: Journal of applied geophysics Vol. 241; p. 105795
• Main Authors: Zhao, Fengda, Zhai, Haobing, Zhou, Zihan, Zhang, Pengwei, Li, Xianshan
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.10.2025
• Subjects: Contrastive learning; Lithological identification; Sample pairs; Semi-supervised; Lithological identification; Sample pairs; Contrastive learning; Semi-supervised
• ISSN: 0926-9851
• DOI: 10.1016/j.jappgeo.2025.105795

Lithological identification is significant in logging interpretation work, as it is the foundation for evaluating reservoirs and describing reservoirs and basins. In lithological identification, due to the difficulty in obtaining lithological data and the high cost of labeling, there is often a situation of needing labels. How to fully utilize the lithological data with missing labels is a problem that needs to be solved. Therefore, this paper introduces a semi-supervised contrastive learning model, SCLRAD (Semi-supervised contrastive learning using random replacement of adjacent depths), aiming to fully utilize lithology data, extract better feature representations, and improve the accuracy of lithology recognition. Lithological data exhibits a discernible pattern of variation in depth. Sample pairs for contrastive learning are constructed by harnessing depth information to unearth deeper features within the data. The loss function weights for various proxy tasks are fine-tuned through a semi-supervised joint training framework. The combined effects of contrast loss and classification loss act on the encoder, enabling it to learn the intrinsic characteristics of the lithological data and capture its inherent differences and similarities, thus enhancing the classifier’s performance. Accordingly, the experimental approach presented in this study is grounded in blind wells testing. The lithology identification accuracy on the China Daqing Fields datasets and Hugoton and Panoma Fields datasets reached 82.02% and 68.55%, respectively. By manipulating the ratio of unlabeled data on the Hugoton and Panoma Fields dataset, even when the labeling ratio is below 60%, comparing favorably with the 1D-CNN, the accuracy improves by 2.77% to 4.22%. •A novel data augmentation approach has been developed.•A joint training architecture enhances the encoder’s feature extraction capability.•Experiments confirm SCLRAD’s effectiveness and competitiveness on lithological data.