Prediction of total organic carbon at Rumaila oil field, Southern Iraq using conventional well logs and machine learning algorithms

Total organic carbon (TOC) is an important parameter for assessing the hydrocarbon potential of source rocks. The standard method for analysis of TOC is the Rock-Eval pyrolysis on cutting and core samples. The coring process is always expensive and time consuming. Therefore, researchers around the world focused on developing techniques to estimate TOC and other organic parameters from readily available well logs data that are almost available in all wells. In this study, we evaluated the use of three machine learning models namely, random forest (RF), rotation forest (rF), k nearest neighbors (KNN) to estimate TOC based on conventional well logs data. The well logs involved gamma ray, acoustic, density, neutron, and deep resistivity. The efficacy of the models was tested against the most widely used backpropagation artificial neutral network (BPANN) and support vector regression (SVR) models. North Rumaila oilfield in southern Iraq was taken as a case study. The models were trained and tested using data from two wells in the field, namely R-167 and R-172. The number of TOC measurements used for training and testing were 40 (R-167) and 18 (R-172), respectively. The efficacy of the used algorithms was evaluated using mean absolute error (MAE), root mean squared error (RMSE), and correlation of determination (R-2). The models are also visually compared using Taylor diagram and violin plot to distinguish the best performance model. Results indicated the KNN was the best followed by RF and then rF. The worst performance models were BPANN and SVR models. This study confirmed the ability of machine learning models for building efficient model for estimating TOC from readily available borehole logs data without the need for very expensive coring process.