A formation-response classifier for borehole deviation using conventional well log data

Accurate and timely identification of borehole deviation is essential in drilling, as unexpected trajectory shifts serve as key indicators of localised geomechanical failure, formation anisotropy, and wellbore integrity degradation. This study presents a systematic machine learning (ML) framework for classifying borehole deviation as Normal or deviated, utilising seven petrophysical and geomechanical well-log features from a 301-sample single-well dataset. Five supervised classifiers, K-Nearest Neighbours (KNN), Logistic Regression (LR), Gaussian Naive Bayes (GNB), Multilayer Perceptron (MLP), and Histogram Gradient Boosting (HGB) were evaluated. Preprocessing incorporated standard scaling, median imputation, and class imbalance mitigation using balanced class weights and the Synthetic Minority Oversampling Technique (SMOTE). To eliminate spatial data leakage, models were evaluated within a 10-fold depth-blocked cross-validation framework across five random seeds, yielding 95% confidence intervals for all performance metrics. The HGB model achieved the highest macro-averaged F1-score of 97.88% (±2.05% CI) and an out-of-fold cross-validation accuracy of 98.00%, outperforming KNN (94.39%), MLP (86.04%), GNB (81.04%), and LR (77.03%). Permutation-based feature importance highlighted Delta T, Gamma-ray, and Resistivity as the primary indicators of deviation risk, aligning with geomechanical principles. These findings demonstrate that Histogram Gradient Boosting with leakage-safe preprocessing provides a robust offline proof-of-concept for assessing wellbore trajectory quality before field validation.