Machine learning approaches for formation matrix volume prediction from well logs: Insights and lessons learned

Abstract

Determining the formation rock type and their petrophysical properties using well-log data is necessary for resource assessment. Formation porosity, shale content, and saturations must be estimated to quantify hydrocarbon reserves. Interpreting well logs could be challenging, especially when the formation is highly heterogeneous. Conventional physics-based interpretation can be tedious and time-consuming. Machine learning (ML) models can automate the well-log interpretation process, saving time and reducing the need for the expertise of trained engineers. In this research, we utilized five supervised ML algorithms incorporating multiple well-log parameters to identify patterns and behaviours in the well-log dataset. Our objective is to provide accurate, real-time predictions of sandstone, carbonate, and shale mineral volumes in the formation. We used the data from open-hole well logs of four wells in the Volve field and trained ML models using Random Forest, Gradient Boosting, Support Vector Regression, K-Nearest Neighbour, and Artificial Neural Networks. We used a robust framework that included random and grid search and cross-validation techniques to find the best hyperparameters and minimize bias and variance in the ML models. ML models were built by training and testing data from two wells and then validated on unseen data from the next two wells.All the ML models were able to predict mineral volume both on testing data as well as on the validation set. K-Nearest Neighbour, Random Forest, and Gradient Boosting outperform the other ML models. The overall prediction accuracy is similar for these three models, and the Random Forest model can be taken as the best model based on Taylor's plot. Random Forest can reduce the variance of the model to minimize the overfitting of decision trees. Gradient boosting builds trees sequentially and focuses on weak predictors to minimize a loss function and achieve better accuracy. We observed that all predictive ML models could easily quantify shale content but have difficulties differentiating sandstone and carbonate. The KNN model results are better than the other ML models for predicting carbonate volumes. One challenge when using ML techniques in the current state is that the accuracy of the model is limited to wells with similar drilling conditions as the training data, without significant changes in geological conditions. This means that a domain expert is needed to assist with predictive modelling using machine learning algorithms. Despite this challenge, our results show that using a machine-learning framework for reservoir characterization can be effective. However, ML models should not be viewed as a replacement for traditional interpretation methods but rather as an additional tool that can be used by engineers involved in petrophysical evaluations.