Enhancing Automated Sliding Performance Through the Integration of Physics Models and Machine Learning Algorithms

Abstract Automated sliding solutions have existed in the industry for years. In early deployments, simply meeting steering requirements with near 100% automation even at the cost of drilling performance was considered a technical success. Now, algorithms are expected to meet or exceed performance standards set by the best drillers. A multi-faceted approach leveraging precise rig control, physics models, and machine learning techniques aims to deliver consistent, best-in-class performance in a scalable manner. The dynamics of sliding was modeled using high fidelity time-domain finite element model across a broad range of operating conditions. A reduced order model, based on a set of key parameters, was fit to the resulting data set to efficiently estimate system dynamics in real-time. The quantity of parameters was restricted to enable reliable real-time calibration based on limited sensor feedback. Historical sliding data from different drilling programs and basins was then used to generate baseline estimates for these parameters. Grounding parameters in the fundamentals of torque and drag (T&D) modeling and bit-rock mechanics was key to producing a simple and robust model. When sliding begins, the parameters are continuously tuned using real-time T&D models and a suite of machine learning techniques to characterize system dynamics and bit mechanics. The reduced model, coupled with a newly designed controller, executes steering based on predicted dynamics and sensor feedback. Continuous estimation of the relationship between Top Drive quill orientation (Θquill) and tool face orientation (TFO) serves as a reference frame on which all steering actions build upon. The combined model and control system were refined by testing on a series of progressively challenging real world sliding scenarios. Core assumptions of physical behavior mostly held true. Other aspects of the system, such as interaction between nested control loops and steering strategies, benefited from substantial changes. A key to continuous improvement was a mature rig automation platform that facilitated a rapid deployment and testing cycle. With over 200,00 feet drilled as part field testing across multiple basins, this system has been shown to deliver steering performance and penetration rates comparable to high performing manually executed slides. In numerous hole sections, penetration rate was increased and steering efficiency improved, yielding measurable improvement in effective rate of penetration. While the manually executed slides during A/B testing required numerous human steering actions per foot to attain peak performance, this system requires an order of magnitude fewer interventions. In difficult sliding environments, high reliance on traditional sensor measurements presents a challenge for automated sliding algorithms. A new method relating quill position to TFO over the course of a BHA run coupled with automated torque release through synchronized drawworks reciprocation and top drive oscillation minimizes off-bottom time without compromising on steering. A model-based solution facilitates stable steering control and consistent performance but may be limited by incorrect assumptions. The fusion of physics models with machine learning presents a major step forward in automated directional drilling.