AI-Driven Environmental Data Extraction for Energy Sector Assessment

Abstract The oil and gas industry continually seeks to enhance environmental stewardship and reduce its operation carbon footprint, with carbon intensity reduction being a key objective. However, the lack of comprehensive and accessible global real operation data hinders accurate environmental assessment, climate modeling, and decision making. This challenge is exacerbated by the dispersion of crucial information across various sources and the high costs and legal aspects associated with accessing proprietary databases. This paper introduces an innovative framework leveraging Large Language Models (LLMs) - specifically applying GPT-4 and GPT-4o - to extract critical oil and gas asset information from diverse literature sources. Our approach aims to bridge the gap between proprietary and open-source information, creating a more accessible database for different sectors such as industry stakeholders, policy makers, and investors. Our evaluation process rigorously assessed the LLM's performance, categorizing extractions into true positives, false positives, and false negatives. We employed key metrics including true positive rate, precision, and F1 score to balance the framework's ability to correctly identify relevant information with its capability to find all pertinent data. An iterative fine-tuning process allowed us to address specific challenges in oil and gas data extraction, such as distinguishing between active and production wells. We developed a novel LLM framework that conducts iterative comparisons between GPT-4's output and a dataset of 108 ground truth documents which were meticulously labeled by domain experts. Through 11 training and testing iterations, we fine-tuned prompts to optimize information extraction. The framework achieved impressive results, with a true positive rate of 81.44% and an F1 score of 76.78% on the testing dataset. The framework demonstrated significant efficiency gains, processing 32 documents in 61.41 minutes with GPT-4o, averaging 7.09 seconds per extraction. This represents a substantial improvement over traditional manual method. Moreover, the cost-effectiveness of the approach, with GPT-4o reducing extraction costs by a factor of 10 compared to GPT-4, makes it accessible to a wider range of researchers and organizations. This research has far-reaching implications for the oil and gas industry. By creating an organized, transparent, and accessible database, we aim to democratize access to critical information, potentially revolutionizing how data is extracted and utilized in this field and beyond. This framework could support more accurate climate modeling efforts, enhance decision-making processes for operations and investments, and contribute significantly to the sector's ability to meet environmental commitments, particularly in areas such as emissions reduction and energy transition strategies.