The total organic carbon content is one of the important indicators for evaluating shale gas, and it is crucial to the accurate prediction of the geological "sweet spot" of shale gas reservoirs and the selection of perforated and fractured wells in the later production process. In view of the shortcomings of the existing total organic carbon content prediction methods for shale gas reservoirs, method △logR has many artificial parameters, strong subjectivity, and multiple regression analysis method has the disadvantage of weak generalization ability. This paper took the marine shale gas in southern Sichuan as the research object, selected different parameters as the eigenvectors of model training by sorting out the logging data and laboratory core analysis results in the research block, established the BP neural network and support vector machine prediction models of the total organic carbon content, analyzed the differences between different models and the influencing factors such as the combination of model features. Finally, we used the selected model to predict the total organic carbon content of the wells that did not participate in the training, compared with the measured values. It is concluded that the BP neural network model without energy spectrum logging data can truly reflect the nonlinear relationship between the logging data and the reservoir, and the error between the predicted results and the actual results is small.