Semi-Supervised Neural Network for Complex Lithofacies Identification Using Well Logs

Lithofacies identification is crucial for characterizing subsurface reservoirs in the development and exploration of oil and gas resources. A common approach is to develop a lithofacies prediction model using well log data annotated with core observations. However, limited availability of labeled data often leads to reduced accuracy in prediction. There is a significant amount of unlabeled logging data without corresponding rock cores, and leveraging unlabeled data has shown promise in enhancing prediction models in other domains such as computer visualization. In this work, a semi-supervised ladder network (SLN) is introduced to improve lithofacies identification, particularly in cases with a small number of labeled data. The SLN combines supervised learning (using labeled data) and unsupervised learning (using unlabeled data) through lateral connections during training. Furthermore, an improved version, improved SLN (iSLN), is proposed, which enhances feature extraction by incorporating oblique connections. To evaluate the effectiveness of the proposed method in complex lithofacies identification, comparison experiments using datasets from continental and marine reservoirs are conducted. The experimental results demonstrate that the iSLN model achieves approximately 5% higher accuracy than the SLN model. Moreover, when only a few labeled samples are available, the iSLN outperforms the supervised model by over 6%. In addition, the optimized determination of key parameters in the iSLN model is discussed. These findings highlight the potential of our approach for improving lithofacies identification when labeled data are limited.