• Iorliam Aamo
  Department of Mathematics/Computer Science, Rev. Fr. Moses Orshio Adasu University, Makurdi (formerly, Benue State University, Makurdi), Nigeria
  




• Tyemimongo Iorlumun Cyprian
  Department of Mathematics/Computer Science, Rev. Fr. Moses Orshio Adasu University, Makurdi (formerly, Benue State University, Makurdi), Nigeria
  

Accurate estimation of Total Organic Carbon (TOC) is essential for evaluating sourcernrock quality, hydrocarbon generation potential, and reservoir characterization.rnHowever, conventional TOC determination methods are often costly,rntime-consuming, and limited by data availability. This paper proposes anrnexplainable hybrid machine learning and deep learning framework for reliablernTOC prediction using petrophysical, elastic, and geochemical attributes derivedrnfrom the Kolmani River-2 (KR-2) and Kolmani River-3 (KR-3) wells. Correlationrnanalysis revealed strong nonlinear relationships among reservoir variables,rnhighlighting the suitability of advanced data-driven modelling approaches. Fourrnpredictive models, namely Random Forest (RF), XGBoost, Decision Tree (DT), andrnBidirectional Long Short-Term Memory (BiLSTM), were developed and evaluated.rnThe results demonstrate that RF consistently outperformed the other models,rnachieving R² values of 0.9907 and 0.9977 for KR-2 and KR-3, respectively, withrnthe lowest MAE, MSE, RMSE, and MAPE values. Ensemble learning methods showedrnsuperior predictive capability compared with deep learning, particularly forrnstructured well-log datasets with limited sample sizes. To enhance modelrntransparency, SHAP-based explainability analysis was employed, revealingrnMeasured Depth (MD) as the most influential predictor, alongside ΔLogR, GammarnRay, Neutron Porosity, TOC (Rk-Eval), and P-sonic. The proposed frameworkrnsuccessfully combines predictive accuracy with interpretability, providing arnrobust and transparent solution.