Modular Framework Integrating Large Language Models with Drilling Hazard Detection Systems to Provide Operational Context-Informed Interpretations and Recommended Actions

Abstract Large Language Models (LLMs) have emerged as transformative Artificial Intelligence tools for advancing how computers process and generate text data, enabling improvements in applications ranging from natural language understanding and translation, to summarization, content creation and decision support. This work describes a modular framework integrating open-source LLMs with drilling hazard detection systems that generate early warnings, in order to provide interpretations and recommendations for mitigating risks, informed by context programmatically retrieved from external knowledge bases. The developed solution connects an existing stuck pipe risk detection system (OTC-32169-MS, SPE-217963-MS) to a module responsible for summarization of observed risks, generating interpretations of the likely sticking mechanism, and recommending mitigating actions according to best practices, in real-time. Interpretations and recommendations are generated by an LLM utilizing Retrieval Augmented Generation (RAG) to obtain relevant contextual information from a vectorized document collection forming an expert knowledge base. The collection contained a combination of general industry best practices, documentation for the linked risk detection system, and may utilize an operator's Standard Operating Procedures applicable to the field where drilling takes place. The stuck pipe detection system reliably generates early warnings, however there are many different stuck pipe risk indicators flagged. This can lead to some required effort by humans to determine the most likely stuck pipe mechanism and poses challenges in determining the right course of action in response to risk warnings. By automatically summarizing a recent history of risk indicators and past risks at nearby depths and generating interpretations and recommended actions using the LLM-based system, this time-consuming process can be completed in near-real time. In order to assess the prototype system's utility, the generated summaries and interpretations were cross-checked by drilling engineers against historical stuck pipe risk scenarios with known context. This supported verification of the interpretations’ accuracy, as well as the utility and appropriateness of the recommended actions. Interpretations of combinations of stuck pipe risk alerts raised during well construction were found by domain experts to be sensible, given their understanding of the operational context, and advised actions were considered appropriate for the scenario. With its ability to utilize relevant context from curated knowledge bases, generate human-like responses, and provide real-time assistance, the system enhances situational awareness, interpretations and information retrieval during drilling operations. This work describes a framework for integrating hazard detection systems with LLMs in order to provide contextually informed interpretations and appropriate recommended actions for risk mitigation. This includes a system architecture, a proof-of-concept implementation focused on stuck pipe, and an approach leveraging RAG to dynamically retrieve and integrate relevant information from multiple sources in real-time. The framework supports systems that respond effectively to drilling hazards, and provide specific, timely, and actionable insights to operations teams.