Reservoir Lithology Identification Based on Improved Adversarial Learning
Lei Song, Xingyao Yin, Linjie Yin
Abstract
Reservoir lithology identification is critical to reservoir characterization, reserves calculation, and geological modeling. The deep learning lithology identification method is a data-driven algorithm for establishing the relationship between lithology-sensitive properties and litho-types from a large amount of observed data. The lithology label is inadequate for high drilling and core recovery costs. Consequently, we propose a reservoir lithology identification method based on improved adversarial learning to relieve the overfitting problem and the multi-solution problem caused by inadequate labeled data and massive learnable parameters in training. Firstly, a probabilistic lithology classification neural network (PLCNN) is constructed to predict lithology from density, P-velocity, and S-velocity. In addition, we design an improved adversarial learning (IAL) lithology identification workflow to train the PLCNN with limited labeled data and large-scale unlabeled data. In the workflow, a lightweight discrimination network is established to ensure that the prediction result of the PLCNN is consistent with the data distribution characteristics of real underground lithology. Finally, the proposed method is successfully applied to the Book cliffs model. Compared with the conventional supervised learning workflow, the misclassification of sand and sandy shale can be relieved efficiently with the IAL workflow, and the classification accuracy can be improved to 92.71%.