Application of Machine Learning for Mineralogy Prediction from Well Logs in the Bakken Petroleum System

Abstract

Abstract One of the significant unconventional oil reserves in the USA is the Bakken Petroleum System located in the Williston Basin. It is known for its complex lithology, composed of three prominent members, Upper and Lower Bakken, with similar properties of organic-rich shale relatively uniform compared to the middle member with five distinct lithofacies, formed mainly from calcite, dolomite, or silica. The higher properties variability makes the reservoir characterization more challenging with low permeability and porosity. Understanding lithology by quantifying mineralogy is crucial for accurate geological modeling and reservoir simulation. Besides that, the reservoir's capacity and the oil production are affected by the type and the mineral volume fractions, which impact the reservoir properties. Conventionally, to identify the mineralogy of the reservoir, the laboratory analysis (X-Ray Diffraction, XRD) using core samples combined with the well logs interpretation is widely used. The unavailability of the core data due to the high cost, as well as the discontinuities of the core section of the reservoir due to the coring failures and the destructive operations, are one of the challenges for an accurate mineralogy quantification. The XRD cores analysis is usually used to calibrate the petrophysical evaluation using well logs data because they are economically efficient. To remedy to these limitations, artificial intelligence and data-driven based models have been widely deployed in the oil and gas industry, particularly for petrophysical evaluation. This study aims to develop machine learning models to identify mineralogy by applying six different machine learning methods and using real field data from the upper, middle, and lower members of the Bakken Formation. Efficient pre-processing tools are applied before training the models to eliminate the XRD data outliers due to the formation complexity. The algorithms are based on well logs as inputs such as Gamma Ray, bulk density, neutron porosity, resistivity, and photoelectric factor for seven (07) wells. XRD mineral components for 117 samples are considered outputs (Clays, Dolomite, Calcite, Quartz, and other minerals). The results' validation is based on comparing the XRD Data prediction from the developed models and the petrophysical interpretation. The applied approach and the developed models have proved their effectiveness in predicting the XRD from the Bakken Petroleum system. The Random Forest Regressor delivered the best performance with a correlation coefficient of 78 percent. The rest of the algorithms had R-scores between 36 and 72 percent, with the linear regression having the lowest coefficient. The reason is the non-linearity between the inputs and outputs.