Quantifying the drivers of well drilling efficiency using a hierarchical explainable machine-learning framework

Abstract

Optimizing drilling efficiency requires understanding not only which parameters influence the rate of penetration (ROP), but also the relative contribution of different drilling measurement domains to predictive capability. Most machine-learning studies emphasize prediction accuracy while providing limited insight into the engineering value of operational, hydraulic, formation, and drill string dynamic measurements. This study presents a hierarchical explainable machine-learning framework to quantify the incremental contribution of these measurement domains using real-time drilling data from the 15/9-F-5 well in the Norwegian Continental Shelf. Following quality control, a dataset comprising 606 complete observations and 24 predictor variables was analyzed. Random Forest (RF), extreme gradient boosting (XGBoost), and CatBoost were evaluated using a depth-blocked GroupKFold cross-validation strategy with hyperparameter optimization, after which the best-performing model was employed for hierarchical feature-domain analysis. Within the baseline hierarchy, the addition of hydraulic measurements produced a 29.87% increase in validation R2. Alternative formation-first and hydraulic-first hierarchies, together with pairwise feature-domain analyses, confirmed that incremental contributions depended on the sequence of feature integration. The robustness of these findings was independently verified using RF feature importance, permutation importance, SHapley Additive exPlanations (SHAP) analysis, and aggregated domain-level SHAP attribution. Aggregated SHAP analysis revealed that formation variables accounted for 48.1% of the aggregated feature attribution, whereas controllable operational and hydraulic parameters collectively accounted for 45.0%. This demonstrates that although formation-related variables exhibited the strongest predictive association, nearly half of the predictive information originates from drilling parameters that may support real-time operational decision-making. Overall, this work advances explainable machine learning in drilling engineering by providing a transparent framework for quantifying both the incremental predictive contribution and the overall feature attribution of drilling measurement domains.