Predicting total organic carbon in Wufeng–Longmaxi shales by coupling machine learning with multiple intelligent optimization algorithms: A case study from structurally complex Northern Guizhou

Total organic carbon (TOC) plays a crucial role in unconventional oil and gas exploration. The traditional ΔlogR method faces limitations under complex geological conditions, while conventional machine-learning models is sensitive to hyperparameter settings, validation design, and model interpretability. This study integrates core and well-log data from six wells in the structurally complex Wufeng–Longmaxi shale succession of northern Guizhou to develop a unified TOC prediction framework. Four metaheuristic optimization algorithms are combined with five machine-learning models. Recursive feature elimination with cross-validation is used for variable selection, and model performance and interpretability are assessed using stratigraphically withheld interval validation and SHapley Additive exPlanations (SHAP). The findings indicate: (1) eight logging parameters are retained as model inputs, and the GWO-CatBoost model demonstrates the best predictive performance, achieving a test-set R 2 of 0.930 and an RMSE of 0.373 wt%, outperforming the traditional ΔlogR method; (2) stratigraphically withheld interval validation in well A1 results in an R 2 of 0.82, indicating stable predictive performance across various intervals of the Wufeng–Longmaxi succession; (3) the model applied to uncored intervals produces continuous TOC curves and effectively identifies organic-rich shale intervals; and (4) SHAP analysis identifies uranium content and compensated neutron log as the most influential variables governing TOC predictions. The reproducible and interpretable TOC prediction framework developed in this study offers reliable technical support for continuous quantitative TOC evaluation and identification of organic-rich intervals in the structurally complex Wufeng–Longmaxi shale succession of northern Guizhou.