Feature Selection-Based Artificial Intelligence Techniques for Estimating Total Organic Carbon from Well Logs

For shale oil and gas exploration total organic carbon (TOC) is the significant factors where TOC estimation considered as a challenges for geological engineers because direct laboratory coring analysis is costly and time consuming. Passey method and Artificial Intelligence (AI) technique have used on well logs extensively to determine TOC content. But, the prediction of Passey method is low and AI technique such as ANN, Support Vector Machine (SVM) trapped in local optima, overfitting and heavy computation work or error if the technique isn't reasonable. In this paper, for the first time in TOC prediction we propose three feature selection-based algorithm which are Decision Tree (DT), Gradient Boosting Regressor (GBR) and Random Forest (RF) respectively. This feature selection-based algorithm select the best attributes among the input parameters for TOC content prediction. Then those best attributes works as an input for AI models for training and testing the AI models which illustrates that making a correlation between well logs and TOC content for the prediction. Specifically, 2069 core shale sample and well logging sample data of the Texas University Lands of Kansas Geologic Society were divided into 1448 training sample and 621 validating sample to evaluate the proposed AI models. This proposed AI model and feature selection-based algorithm jointly allows TOC content to be accurately and continuously predicted based on conventional well logs.