A Hybrid Physics-Informed Machine Learning and Causal-Analysis Framework for Real-Time Prediction and Prescriptive Minimization of Non-Productive Time Using Reference Wells

Abstract

Non-productive time (NPT) continues to represent one of the largest controllable costs in drilling operations, frequently consuming 20–30 % of total rig time and generating losses that can easily reach several hundred thousand dollars per well. The work described here introduces a hybrid analytical framework that combines physics-informed machine learning, causal root-cause attribution, and a lightweight digital twin to deliver real-time NPT prediction together with ranked prescriptive actions. Multi-well reference datasets are used to construct transferable, uncertainty-quantified performance standards that guide field decisions and reduce variability between shifts. Telemetry and event logs are continuously ingested and pre-processed to produce labeled activity timelines. Sequence labeling models working together with operation-specific rules identify individual drilling tasks and compute relevant KPIs. Transfer learning from a broad set of historical wells creates tailored baseline expectations for the active well, while a physics-informed regularization term enforces consistency with established torque-and-drag relationships and hydraulic models (ECD, swab/surge effects). Causal inference, implemented through intervention-based counterfactual reasoning, pinpoints the controllable factors responsible for observed deviations. A near-real-time digital twin then runs Monte Carlo simulations incorporating parameter uncertainty to evaluate corrective options and present the most promising recommendations on field dashboards. The system was applied to a large historical multi-well dataset and subsequently deployed in a twelve-month live pilot on operating rigs. Compared with conventional statistical alerting thresholds, the predictive component achieved roughly 35 % higher true-positive detection of impending NPT events and reduced false alarms by 30–50 %, allowing interventions at an earlier stage. The prescriptive recommendations, ranked by simulated NPT impact, delivered an average well-by-well NPT reduction of 12–18 %, with the most pronounced improvements occurring during tripping and connection sequences. At typical rig day rates of 120 000–200 000 USD, these time savings translated into direct cost reductions estimated between 100 000 and 440 000 USD per well, depending on total measured depth and well design. Shift-level visibility and causal feedback loops decreased inter-shift performance dispersion by approximately 22 % and improved compliance with well-specific operational targets. Overall, the integrated approach converts reliable forecasting into field-verified, prioritized corrective steps that meaningfully lower NPT and well construction expenditure while remaining scalable across rig fleets. The principal contributions are: Creation of transferable, physics-constrained, uncertainty-aware operational benchmarks derived from reference wells via transfer learning.Conversion of predictive NPT signals into causally grounded prescriptive actions validated by rapid digital-twin scenario runs.Documented field results showing sustained reductions in NPT and shift-to-shift variability. Machine learning models and physics-informed neural networks were used for real-time operation detection and NPT prediction to reduce ILT(invisible lost time) with around 100 well data started and revised through time.