Comparative analysis of machine learning and empirical approaches for total organic carbon prediction in the J1d formation, Sichuan Basin, China

Traditional total organic carbon (TOC) prediction methods, such as multiple linear regression, ΔLogR and GR-based methods, often yield low accuracy due to inadequate account for subsurface heterogeneity in the J1d formation, which includes mineral composition, organic matter distribution, and depositional characteristics. Traditional approaches fail to explore the nonlinear relationships between TOC content and geological parameters. To address this issue, the current study employs machine learning (ML) methods such as extreme gradient boosting (XGBoost), random forest (RF), and support vector regression (SVR) to improve the accuracy of TOC estimates in the J1d shale in the Jiannan area, Sichuan Basin, China. By incorporating multiple well log data and optimizing each algorithm's hyperparameters through grid search cross-validation, we enhanced the models' ability to describe the complex, nonlinear relations between geological variables and TOC content. The results demonstrate that ML algorithms significantly outperform traditional methods. Among the models, XGBoost achieved the highest accuracy with R2 = 0.9704, RMSE = 0.0001, MAE = 0.060, and MSE = 0.006. Random forest followed with R2 = 0.9700, RMSE = 0.082, MAE = 0.069, and MSE = 0.007, while SVR showed lower accuracy (R2 = 0.8650, RMSE = 0.096, MAE = 0.078, MSE = 0.0095). In comparison, conventional methods exhibited poor correlation with actual TOC values. These findings highlight the potential of ML algorithms to improve TOC prediction in shale of J1d formation, Sichuan Basin, China.