Digitization of Daily Drilling Reports Using LLMs

Abstract This research aims to revolutionize Daily Drilling Reports (DDR) analysis in the energy sector by leveraging Large Language Models (LLMs) and advanced natural language processing (NLP) techniques. The methodology transforms unstructured data with inconsistent formats and terminologies into actionable insights, minimizing errors, accelerating analysis, and enhancing decision-making. Utilizing OpenAI models and Few-Shot Learning, this approach achieves efficient event documentation and operational excellence, optimizing drilling operations while mitigating risks with unparalleled speed and precision. The methodological framework integrates NLP techniques with robust data extraction pipelines specifically tailored for DDRs. Fixed positioned text is extracted using Python tools, enabling accurate retrieval of parameters like Date, Measured Depth, and Casing Size. To address the unstructured content, marked by manually written operational summaries of varying lengths and formats, a structuring mechanism is developed to identify and classify events into predefined categories using in-context learning with GPT models. Few-shot learning techniques are employed to fine-tune the models with a custom DDR dataset, strategically divided into training, testing, and blind sets, ensuring validation across diverse examples and enhancing reliability. Descriptive prompt engineering and hyperparameter optimization yielded an accuracy exceeding 95% and an F1 score above 85% on a dataset of 15 wells comprising over 3,800 DDRs in PDF format. Fine-tuned models demonstrate high precision and adaptability for context-sensitive data points, such as stuck pipe incidents and fluid losses. By integrating coordinate-based data extraction, few-shot learning, and systematic model evaluation, the workflow significantly reduced DDR analysis time from months of manual effort to mere hours while ensuring accuracy. The solution effectively structures unstructured DDR data into Excel spreadsheets, with each row representing a day's operations, containing extracted details such as date, measured depth, mud weight, casing details, and categorized operational events from manually written operational summaries, including stuck pipe incidents, losses, tight holes, reaming, inflows/kicks, leak off tests, and formation integrity tests. This enables operators to access actionable insights like incident depths, facilitating optimized strategies for drilling operations and future well planning. The modular architecture ensures compatibility with various LLMs and NLP techniques, supporting customization for diverse datasets and parameter configurations. Continuous ingestion of new data enables the framework to evolve dynamically, improving accuracy and reliability while ensuring robust validation through testing and blind datasets. The solution's adaptability to diverse datasets and real-time learning capabilities delivers a transformative approach, enabling optimized drilling strategies and actionable, data-driven insights for enhanced field performance in the oil and gas sector.