Log facies identification in carbonate reservoirs using multiclass semi-supervised learning strategy

Abstract

Facies identification using conventional log data is one of the most important contents for reservoir fine characterization, where some machine learning methods have been applied due to the advantages of automatic interpretation. The application effect of commonly used supervised learning methods depends on whether there are large amounts of precisely labeled samples, but it is time-consuming and laborious to manually label enough log facies samples. Based on the idea of semi-supervised learning, we have proposed the multiclass positive and unlabeled machine learning (PU-learning) which uses labeled data and unlabeled data simultaneously to identify 5 types of carbonate log facies including carbonate bedrock, collapsed cave, cave-filling of sand and mud, dissolution fracture and unfilled cave. This method can further use the feature information of a large number of unlabeled samples to alleviate the overfitting caused by insufficient labeled samples without adding any additional labeling costs for model construction. An actual application for two carbonate well blocks in the Tahe Oilfield shows that when using only limited labeled log facies data, PU-learning outperforms supervised support vector machine (Supervised-SVM) and supervised artificial neural network (Supervised-ANN) in general. Especially for minority classes like cave-filling of sand and mud in the imbalanced dataset, the evaluation indexes Recall and F1 significantly improved by PU-learning show that the semi-supervised learning strategy is of great meaning to the actual automatic log interpretation work.