Petrophysical log-driven kerogen typing: unveiling the potential of hybrid machine learning

Abstract

The importance of characterizing kerogen type in evaluating source rock and the nature of hydrocarbon yield is emphasized. However, traditional laboratory geochemical assessments can be time-intensive and costly. In this study, an innovative approach was taken to bridge this gap by utilizing machine learning techniques to ascertain key parameters—Organic Oxygen Index (OI), Hydrogen Index (HI), and kerogen type—from petrophysical logs of a well in the Perth Basin, Western Australia. This approach assembled geochemical data from 138 cutting samples of the Kockatea and Woodada formations and petrophysical log data. Subsequently, six machine learning algorithms were applied to predict the OI and HI parameters. The efficacy of these methods was assessed using statistical parameters, including Coefficient of Determination (R2), Average Percentage Relative Error, Average Absolute Percentage Relative Error, Root Mean Square Error, and Standard Deviation. The Support Vector Machines method emerged as the standout performer, with an R2 of 0.993 for the OI and 0.989 for the HI, establishing itself as an optimal tool for predicting these indices. Additionally, six classifiers were employed to determine kerogen types, with accuracy tested using precision, recall, F1-Score, and accuracy parameters.The study's findings highlight the superiority of the Gradient Boosting method in kerogen-type classification, achieving an impressive accuracy rate of 93.54%. It is concluded that when utilized with petrophysical logs, machine learning methodologies offer a powerful, efficient, and cost-effective alternative for determining OI, HI, and kerogen type. The novelty of this approach lies in its ability to accurately predict these crucial parameters using readily available well-log data, potentially revolutionizing traditional geochemical analysis practices.Graphical abstract