Learning Across Wells: A Robust Framework for Lithology Classification and Prediction

Abstract Automated lithology classification from wireline logs plays a crucial role in enhancing subsurface characterization, reservoir delineation, and completion design efficiency. This paper develops two widely used supervised machine-learning classifiers, Random Forest (RF) and K-Nearest Neighbors (KNN), for predicting multi-well lithology. The evaluation uses four standard wireline logs: gamma ray (GR), bulk density (RHOB), corrected sonic transit time (DT CORR ), and effective porosity (PHIE), collected from four North Sea wells (F02-01, F03-02, F03-04, and F06-01). Four lithology classes Shale, Sandstone, Tight Sand, and Carbonate were used to conduct the evaluation of the models, which were conducted without core-calibrated ground truth or fixed petrophysical cut-offs. The assessment involved blind-well testing and Leave-One-Well-Out (LOWO) cross-validation. Hyperparameters were improved through the application of stratified five-fold cross-validation. In the blind F06-01 well, Random Forest (RF) demonstrated an accuracy of 100.0% (weighted F1 = 1.000), in contrast to KNN's accuracy of 98.45%. Additionally, LOWO validation produced mean accuracies of 99.58% for RF and 97.10% for KNN. GR, RHOB, and PHIE were the primary features of significance, while DT CORR had a negligible contribution. The models reached a consensus on 98.4% of blind-well samples, with RF accurately addressing all discrepancies. Even without feature scaling, Random Forest's full, feature-randomized ensemble structure surpasses the distance-based KNN classifier in inter-well log variability robustness. We address the limitations of rule-based labeling without core data, review machine learning research for lithology and facies classification, and propose resistivity logs, ensemble boosting, and core-calibrated labels for future research.