Predictive Modelling and Application of Total Organic Carbon Content in Geophysical Logging of Deep and Thin Clastic Formations in the Southern Ordos Basin

In order to analyze the specific significance of stratigraphic total organic carbon (TOC) distribution for gas field development, as well as to optimize the existing TOC prediction methods, carbonaceous mudstones and coal seams from the southern margin of the Ordos Basin was selected as the experimental object, and TOC testing experiments were carried out with coaly stratigraphic core. The core TOC experiment and its corresponding logging curve data were applied. Utilizing logging curves as the input features, experimental TOC values as the target label, and assessing the model using Mean Absolute Error (MAE) and Coefficient of Determination (R2), a refined TOC logging prediction model was developed through the application of Homogeneous Clustering (K-Means) and Genetic Algorithms (GA) in combination with GA-GBDT. The findings from the experiment indicate that incorporating K-Means can enhance the precision of the machine learning model for predicting TOC. The model that incorporates K-Means grouping for TOC content prediction, known as GA-GBDT model, demonstrates superior performance, especially in single-well applications. Therefore, the grouped GBDT_0/GBDT_1 model with K-Means was selected as the TOC content prediction model. The peak of TOC content in the study area has a gradually decreasing trend from southwest to northeast, which provides a reference basis for efficient and relatively accurate assessment of hydrocarbon potential and favorable zone preference in the study area.