Abstract

Understanding the geology of a reservoir requires accurate identification of its lithological facies. Lithological facies classification is the process of determining rock lithology by analysing indirect measurements such as well logs. Well logging is a standard tool for determining reservoir parameters. The interpretation of well logging data, on the other hand, may take some time but machine learning (ML) can automate the process in less time.

The objective of the study is to investigate the use of machine learning and well logs in classifying rock facies and to improve the accuracy of different machine learning models compared to previously published research on the same dataset. A total of 118 wells, with 12 known lithology from a field in the North Sea were available for analysis and classification, the dataset is divided into three for training, testing and validation purpose.

Nine machine learning models which include Random Forest (RF), Decision tree, Support vector machine (SVM), Gradient boosting (GB), Categorical boosting (CatBoost), Light gradient boosting (LGBM), Extreme gradient boosting (XGB), K-Nearest Neighbour (KNN), and logistic regression was designed and trained with a sample of the dataset. The random forest model outperforms the other methods in classifying the lithofacies in the dataset with 91% accuracy. For each model, various scenarios were tested including feature engineering, machine learning imputation, data augmentation and outlier removal to improve the performance of the models.

Random Forest, CatBoost, XGB and Gradient boosting performed best without any form of feature engineering, outlier filter and ML imputation. Logistic regression, decision tree and SVM performed best when the dataset has ML imputation (scenario 2), KNN and LGBM performed best with feature engineering. The top three models were then used to train the entire train set and evaluated with the open and hidden datasets. The RF base model performed best with an accuracy of 78% and 80%, on the open and hidden data respectively, the accuracy on the open data increased by 1% when the data is trained with hyperparameters and augmented data. The accuracy of the XGB model increased by 4% and 3% when the data was trained with additional features.

According to the findings, machine learning algorithms have a significant application in the automatic classification of lithology facies with high accuracy and efficiency. It has the potential to significantly improve the rock physical property estimation process while simultaneously reducing the amount of manual labour required.

Classification of lithology facies using machine learning model