Exploratory Analysis of Machine Learning Methods for Total Organic Carbon Prediction Using Well-Log Data of Kolmani Field

Machine Learning methods have shown significance in automating the prediction of Total organic carbon (TOC) for determining source rock potential in oil and gas exploration. Higher TOC contents could indicate greater potential for oil and gas generation. Making accurate TOC predictions, therefore, is crucial in measuring hydrocarbon deposits for a prospective geological formation. Automating this procedure can save time and resources when compared to conventional geochemical methods. In this study, we explore machine learning methods on a Frontier-Basin well-log data for TOC Prediction. Firstly, we employ feature selection methods to ensure optimal feature usage in regression and classification tasks. Additionally, we compare several supervised Machine Learning methods including Logistic Regression, Decision Tree, K-Nearest Neighbor, Gradient Boosting, and Naive Bayes Methods to categorize the quality of TOC using its defined standard ranges on the well-log data of Kolmani River 2 and 3, respectively. Furthermore, Random Forest method was utilised for the regression task on both data. A 5 fold nested cross validation was utilised for both classification and regression task. We show an exploratory analysis and the prospect of using Machine Learning methods to effectively classify TOC distribution using continuous well log data. Results show that Machine Learning methods are efficient for TOC prediction for non-geochemical approaches. The regression analysis shows an acceptable R² value of 0.62 and 0.76 for the respective welllogs with an MSE score of 0.05 and 0.07, respectively. For the classification task, evaluation metrics including F1 score, Precision, Accuracy, and Recall were investigated for the different ML classification methods, and Naive Bayes’ was concluded to outperform others using standard metrics.