_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 224618, “Application of Predictive Machine-Learning Optimization Enables Successful Delivery of Highly Challenging Wells, ” by Inaki Guenaga, SPE, Rohit Singhal, SPE, and Omar Al-Farisi, SPE, Dragon Oil, et al. The paper has not been peer-reviewed. _ This paper describes a case application of a software solution enabling prescriptive optimization of well delivery, using an industry-proven physics-informed machine-learning (ML) approach that requires no local training, for predictive identification and characterization of pending well-construction risks in a highly challenging, high-consequence operating environment. The solution is agnostic to geology, rig, bit, bottomhole assembly (BHA), or fluid, and requires only surface drilling parameters and trajectory data to provide advisory service for stuck-pipe avoidance, rate-of-penetration (ROP) optimization, and vibration monitoring. Methodology and Processes To leverage the advantages that an ML approach can bring to a complex system as a whole, it is advisable to strictly isolate specific concurrent operations and related challenges within that system and apply ML modeling to each individually. The output from each model, agent, or subsystem may be used as input for population of another alongside raw measured data, calculated engineering data, and traditional physics-based modeling. This deconstructed approach is considerably more sophisticated than replicating the process of an exhaustive, pure physics-based model in ML form, and also brings practical benefits. The number of determinant parameters per agent is reduced substantially, meaning that each agent becomes agnostic to the idiosyncrasies of local external causal factors. Thus, the agents can be pretrained robustly on a well-curated comprehensive data set representative of a variety of well-construction environments, and can be implemented in any operating environment without need for local training. The solution deployed for the operator is built on this approach, in which the underlying “engine” comprises multiple and interoperative ML agents. Each uses surface data exclusively to model one discrete facet of the well-construction process. Using information captured within their in-built training data, these agents monitor for risk symptoms and leading indicators present in these various modeled characteristics synchronously and collectively, and forecast the near-term evolution of a subset of these characteristics to predict, as early as possible, the occurrence of various hazards that may lead to nonproductive time (NPT). In practice, the ML solution is applied to address and ameliorate the two following key real-world challenges to well delivery, both forensically using historical data for future well planning, and in real time for forward-looking risk avoidance and immediate incremental optimization: - Predictive anticipation, identification, and characterization of pending stuck-pipe risk, categorized by static-, dynamic-, and fluid-friction-based dysfunction, informing fit-for-purpose prescriptive action to be taken in good time to avoid NPT - Proactive advisory for pragmatic ROP optimization, with operationally sensitive recommendations for weight on bit (WOB), rev/min, or flow provided to ensure minimal invisible lost time (ILT) through optimal ROP, albeit within operational, geological, or equipment constraints and including safeguards against induced vibration or excess solids loading.
Chris Carpenter (Fri,) studied this question.