Data-driven total organic carbon prediction using feature selection methods incorporated in an automated machine learning framework

An accurate assessment of shale gas resources is highly important for the sustainable development of these energy resources. Total organic carbon (TOC) analysis thus becomes fundamental for understanding the distribution and quality of hydrocarbon source rocks within a shale gas reservoir. The eleva...

Full description

Bibliographic Details
Main Authors: MacĂȘdo, Bruno da Silva, Wayo, Dennis Delali Kwesi, Campos, Deivid, De Santis, Rodrigo Barbosa, Martinho, Alfeu Dias, Yaseen, Zaher Mundher, Saporetti, Camila M., Goliatt, Leonardo
Format: Article
Language:English
Published: Nature Publishing Group 2025
Subjects:
Online Access:http://umpir.ump.edu.my/id/eprint/45116/
http://umpir.ump.edu.my/id/eprint/45116/1/Data-driven%20total%20organic%20carbon%20prediction%20using%20feature%20selection.pdf
Description
Summary:An accurate assessment of shale gas resources is highly important for the sustainable development of these energy resources. Total organic carbon (TOC) analysis thus becomes fundamental for understanding the distribution and quality of hydrocarbon source rocks within a shale gas reservoir. The elevation of the TOC is often associated with the presence of source rocks, indicating the potential for oil and gas production. TOC assessment is performed using laboratory methods, which can be time-consuming and costly. Data-driven models have been successfully applied to model the relationship between TOC and other constituents and to predict the TOC content. However, these methods depend on extensive parameter adjustments that must be carefully conducted in different sedimentary environments. In this context, Automated Machine Learning (AutoML) is an alternative for accurately predicting TOCs, saving time-consuming fine-tuning steps in model development. This study aims to develop an AutoML strategy for estimating TOC using well log data. This procedure automatically preprocesses the search for the best method parameters, reducing the execution time. Among the methods evaluated, Extremely Randomized Trees (XT) performed best (R = 0.8632, MSE = 0.1806) in the test set. The proposed strategy provides a powerful data-driven method, which allows real-world use of the well to assist in data analysis and subsequent decision-making.