Mechanism-Integrated Predictive Analytics for Stuck Pipe Prevention: A Physics-Guided Machine Learning Framework with Operational Decision Support

Stuck pipe is responsible for 10–30% of total non-productive time in drilling operations worldwide and arises from distinct physical mechanisms — differential sticking, pack-off, mechanical obstruction, and hydrostatic imbalance — that interact in ways conventional monitoring cannot detect in advance. Existing machine learning approaches improve on threshold-based alarms but remain disconnected from the engineering reasoning required to act on their outputs. This study develops a stuck pipe prediction framework that integrates physics-guided feature engineering, time-series behavior capture, and an operational decision-support layer. Engineered precursor features — including a torque fluctuation index, hole-cleaning efficiency proxy, differential pressure indicator, and hook-load deviation — encode the dynamic signature of stuck pipe onset rather than instantaneous parameter snapshots. Applied to a dataset of 14,232 depth-indexed records across 24 wells (class imbalance ratio 27:1), the XGBoost classifier achieves an AUC of 0.844 under time-based validation and 0.767 under well-based validation, with mean AUC of 0.793 ± 0.081 across five independent well-split configurations. Compared to the logistic regression baseline (AUC 0.743, F1 0.137), the physics-guided model represents a 13.6% AUC improvement and a 114% F1 improvement on time-based evaluation. The framework's primary operational contribution is a mechanism-attribution decision layer that routes risk outputs to specific parameter interventions — WOB reduction for differential sticking, flow rate correction for pack-off, reaming protocols for mechanical sticking — enabling field deployment that closes the loop between prediction and response. An illustrative rig scenario demonstrates 40+ minutes of actionable lead time ahead of conditions that would otherwise result in stuck pipe. The framework is designed for integration with standard WITSML rig data infrastructure and provides a reproducible, interpretable basis for real-time drilling risk management.