Abstract:  Total organic carbon (TOC) content is an important parameter for evaluating the geochemical characteristics of shale reservoirs, serving as a key indicator of organic matter content and hydrocarbon generation potential. Logging data, with its high vertical resolution, can compensate for the limitations of insufficient sampling and high testing costs. As a result, using logging data to predict TOC content has become a crucial technology in shale oil exploration.This paper proposes a combined model integrating the Gaussian mixture model (GMM) and the light gradient boosting machine (LightGBM) to achieve well-logging-based TOC content  prediction. Taking the first section of  Qingshankou Formation in  Changling depression, southern Songliao basin, as the study area, we collected acoustic time difference, neutron porosity, natural gamma, resistivity, and density logging data. Outlier detection and processing were conducted using the boxplot method. Cross-validation and grid search methods were employed to optimize the parameters and establish the TOC content prediction model. The proposed method was compared with five models: The improved Δlg R, K-nearest neighbors, decision tree, extreme gradient boosting, and LightGBM. The results indicate that the GMM-LightGBM model achieved the best prediction performance, with evaluation metrics of root mean square error (RMSE), mean absolute error (MAE), and the coefficient of determination (R2) being 0.254, 0.211, and 0.766, respectively. Furthermore, the model was applied to another core well, demonstrating satisfactory performance with RMSE, MAE, and R2 values of 0.547, 0.462, and 0.647, respectively.

Key words: machine learning, GMM-LightGBM model, TOC, shale, reservoirs, Songliao basin