Machine Learning Approaches for Predicting Total Organic Carbon From Petrophysical Well Logs in the Wufeng–Longmaxi Shale, Southeast Sichuan Basin, China

ABSTRACT Total Organic Carbon (TOC) is a fundamental parameter for evaluating shale oil and gas potential, as it influences hydrocarbon generation and a wide range of petrophysical properties, including porosity, density, brittleness, and reservoir quality. Traditional TOC determination using laboratory geochemical analysis, though accurate, is slow, expensive, and unable to provide continuous depth profiles. Empirical log‐based estimation methods such as the Schmoker density model, gamma‐ray correlations, multiple linear regression, and the ∆logR technique offer faster alternatives but suffer from limited generalizability and poor performance across formations with variable mineralogy and depositional settings. This study develops and compares three machine learning (ML) models, Random Forest (RF), Multiple Linear Regression (MLR) and XGBoost, for predicting TOC from four common well logs (GR, NPHI, RHOB, DT). Data from four shale‐rich wells are preprocessed, scaled, analysed, and used to train, test, and validate each model. The models are further benchmarked against traditional log‐based empirical methods. Across all wells, the RF model consistently outperforms other approaches, achieving the lowest RMSE values and highest R 2 scores and demonstrating strong generalization capabilities. XGBoost shows moderate performance, while MLR performs the weakest due to its inability to model nonlinear relationships. The results confirm that ML‐based workflows, especially ensemble algorithms, substantially improve TOC prediction accuracy and reliability across diverse geological settings.