| Abstract |
Estimating Shale Volume using conventional methods in the Bakken formation, a complex heterogeneous reservoir, is challenging due to the presence of other radioactive minerals. This study aims to evaluate the accuracy of several machine-learning algorithms in predicting shale volume variation using conventional well logs. For a detailed understanding of the shale volume variation in different fields in the Bakken formation, quad-combo logs with X-ray diffraction data were collected from six wells. The workflow consists of measuring shale volume using the gamma-ray method and calibrating it with values obtained from X-ray diffraction. The obtained dataset is then split into training and testing sets. In total, around 5225 data points were used for the purpose of this study. Several ML algorithms, namely support-vector regression, decision trees regression, random forest regression, gradient boosting, and deep neural networks, were tested and compared with the results of conventional methods without being calibrated with the shale volume derived from XRD. Machine learning models were able to predict shale volume with an accuracy higher than traditional methods that can be greatly influenced with radioactive minerals in formations such as the Bakken and produce erroneous results. Shale volume predicted using machine learning algorithms such as Gradient Boosting had a correlation coefficient of 0.98 compared to traditional methods which had correlation coefficients ranging from 0.27 to 0.52 for the six wells. Therefore, machine learning models are a more accurate predictor of shale volume in the Bakken formation, while relying only on the quad-combo logs data. |
, A. Dehdouh 
, Ilyas Mellal 
