Water Saturation Prediction Using Machine Learning and Deep Learning. Application to Three Forks Formation in Williston Basin, North Dakota, USA

This chapter presents a study on the construction of a reliable machine learning model for water saturation prediction in thin beds reservoir using conventional logs. The study proposed and built two supervised machine learning algorithms and one deep learning algorithm to predict consistent results. The dataset was pre-processed, and the importance of input variables to model construction was discussed. The results showed that the conventional logs GR, log Rt, NPHI, DT, and RHOB are important input variables to the learning process. More attributes were added to the learning process, such as output volumes from petrophysical analysis, formation members, and the hydrocarbon column. The results demonstrated the effectiveness of applying support vector regression (SVR) in thin beds analysis with a correlation factor of 0.78. The backpropagation neural network and random forest regression algorithms were also applied to the same dataset, with almost similar performance results. Although the program could not model perfectly the peaks due to the complexity of the Three Forks Formation, the study showed that the models are valuable methods for thin beds water saturation prediction using only conventional logs and increasing input variables could improve prediction accuracy.