Prediction of total organic carbon at the oil field using conventional well logs and machine learning algorithms

Quantifying total organic carbon (TOC) is critical for indirectly identifying organic-rich source rock intervals and serves as a vital component of petroleum system modeling. While Rock-Eval pyrolysis is the standard analytical method, coring processes are costly and time-consuming, making TOC estimation from readily available well-log data a practical alternative.Existing approaches face three key limitations: 1) Well-specific empirical models trained case-by-case require complex parameter tuning and lack generalization capability; 2) Performance heavily depends on training data quality and quantity; 3) Single models exhibit instability and overfitting risks. To address these challenges, we propose VOT, a multi-model ensemble learning framework for TOC prediction, comprising two innovations: 1) A data augmentation pipeline simulating real-world variability to expand dataset diversity and size; 2) Using the augmented training dataset, four machine learning methods and the VOT method were employed to predict TOC.Validated through 10 wells in southern Sichuan Basin (748 core samples for training, 191 for testing). The experimental results indicate that the best-performing single model is CatBoost, followed by LightGBM and Random Forest, with XGBoost being the least effective. However, XGBoost outperforms the other methods in the RE metric, suggesting potential complementarity in the predictive capabilities of the four methods. Indeed, the proposed VOT, which integrates the four models, outperforms any single model overall, demonstrating that the proposed method effectively addresses the aforementioned issues.