Advancing shale geochemistry: Predicting major oxides and trace elements using machine learning in well-log analysis of the Horn River Group shales

This study evaluates machine learning algorithms for predicting geochemical compositions in the Middle to Upper Devonian Horn River Group shales. The minor textural variations within shale successions necessitate a detailed understanding of their geochemical composition for accurate interpretation of depositional environments and stratigraphic relationships. Geochemical analysis is essential for unconventional reservoir shales but is traditionally labor-intensive. Machine learning offers a cost-effective alternative, streamlining geochemical interpretation and stratigraphic correlation. Five models, Random Forest Regressor, Gradient Boosting Regressor, XGBoost, Support Vector Regressor, and Artificial Neural Networks (ANN), were assessed using well-log data to predict major oxides and trace elements. Training and validation used data from two wells, with model performance tested on an unseen well to evaluate generalizability. Tree-based models, particularly Random Forest Regressor, demonstrated high accuracy for major oxides such as K₂O and CaO, while Gradient Boosting Regressor excelled for Al₂O₃ and TiO₂. However, SiO₂ and Na₂O were less predictable due to their complex origins and low concentrations. For trace elements, Random Forest Regressor effectively predicted Th, Zr, Co, and total rare earth elements (∑REE). Redox-sensitive elements such as Mo, Cu, U, and Ni had lower accuracy due to their weaker correlation with well-log data; however, Random Forest Regressor still achieved the best performance among the models for these elements. Blind tests confirmed the generalizability of the models, with tree-based models maintaining strong predictive performance for several major oxides, trace elements, and REEs, while ANN and Support Vector Regressor exhibited robustness in Al₂O₃, K₂O, and TiO₂ predictions. This study highlights tree-based models as reliable tools for predicting geochemical compositions, supporting chemostratigraphy and reservoir characterization. Integrating machine learning with well-log data offers a promising solution for efficient geochemical analysis and subsurface characterization in Devonian shale reservoirs.