Logging Prediction Method of Organic Carbon in Mixed Deposits Based on Machine Learning

The quantitative evaluation and prediction of total organic carbon (TOC) content is an important part of source rock quality evaluation. However, the traditional TOC content prediction methods, such as geochemical analysis, nuclear magnetic resonance response and empirical formula, have some problems, such as high prediction cost and difficult to deal with complex strata and lithology overlap. The machine learning method has a significant advantage in solving the complex nonlinear relationship between data because of its powerful nonlinear mapping ability. In this paper, three machine learning methods, namely XGBoost, random forest and support vector regression (SVR), are used to predict TOC content in the study area by selecting the logging properties sensitive to TOC content, such as natural gamma ray, sonic time difference, neutron and compensation density. XGBoost, random forest and support vector regression were used to predict TOC content, with R2 of 0.77, 0.76 and 0.77, and MAE of 1.25, 1.25 and 1.21, respectively. The results show that the machine learning method can effectively evaluate the quality of source rocks and accurately identify the "sweet spot" horizon, which provides a reliable basis for reservoir evaluation and reservoir development.