Tree-based, boosting, and stacked models for accurate prediction of total organic carbon from conventional well logs

Rock-Eval pyrolysis provides accurate total organic carbon (TOC) measurements but are expensive, time-intensive, and reliant on the availability and quality of rock samples. The Passey method, a well log-based approach, serves as an alternative; however, it often underestimates TOC compared to laboratory analyses. To address these limitations, this study introduces several innovative machine learning (ML) frameworks for TOC estimation, such as ExtraTrees (ET), Gradient Boosting (GB), and XGBoost (XGB), along with three stacked hybrid models (HM1, HM2, and HM3) that integrate ET, GB, and XGB in different combinations, resulting in a total of six distinct models. Except for XGBoost, these models have not been previously applied to TOC prediction, underscoring the novelty of this study. Among the models, ET achieves the maximum prediction accuracy, with a correlation coefficient (R²) of 0.9825 and root mean square error (RMSE) of 0.2483, followed closely by GB, which yields similar performance. A feature importance analysis, conducted using the best-performing ET model through sequential parameter exclusion, identifies gamma ray as the most influential predictor, while resistivity has the least impact on TOC estimation. The Passey method exhibits significantly lower predictive accuracy (R² = 0.621, RMSE = 1.018), further demonstrating the superiority of ML models. The evaluation of TOC is conducted here utilizing 125 core samples and well log data from the Shahejie Formation in the Dongying Depression, Bohai Bay, China. Additionally, the proposed methodology has been evaluated in an entirely new region located in Alberta, Canada, to improve the generalizability of this work. • Six machine learning models have been developed for TOC prediction. • ExtraTrees performed best among the proposed models. • Traditional Passey method exhibited significantly lower predictive accuracy. • Acoustic log is the most significant feature, while neutron is the least. • Validation of the proposed methodology in an entirely new field is performed.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.