An integrated machine learning framework with quantile-based uncertainty for shale volume estimation, lithofacies classification, and reservoir interval identification from well logs: Datta Formation, Bannu Basin, Kohat Plateau, Pakistan

Accurate estimation of shale volume, lithofacies distribution, and reservoir intervals is critical for reservoir characterisation in structurally complex fold-and-thrust belt settings. In this study, an integrated machine learning framework is developed to characterise the Jurassic Datta Formation in the Kohat Plateau, Pakistan, using conventional well logs. A dataset of 840 samples was used for this study. Multiple supervised regression models are applied to predict shale volume (Vsh), while lithofacies are identified through a combination of unsupervised K-means clustering and supervised classification. Quantile-based uncertainty estimation is incorporated using bootstrap resampling to generate depth-wise P10, P50, and P90 Vsh envelopes, enabling probabilistic assessment of prediction reliability. Model performance is evaluated using standard regression and classification metrics, residual analysis, and confusion matrices. Results demonstrate that tree-based models, particularly Random Forest and Decision Tree, provide R2 and RMSE Vsh predictions and stable lithofacies classification with reduced uncertainty. Lithofacies predictions are geologically consistent with the known interbedded sandstone – shale architecture of the Datta Formation. Integration of Vsh thresholds, facies classification, and uncertainty filtering enables robust delineation of gross and net reservoir intervals. The proposed framework provides an uncertainty-aware, data-driven approach for reservoir characterisation in tectonically complex clastic systems.