Application of Self-Supervised Autonomous Agent Framework for Digital Transformation of Elder Well Potentials Discovery

Abstract Objectives/Scope The manuscript presents an innovative agent-based framework aimed at enhancing the reassessment of elder oil and gas wells. By integrating diverse generative artificial intelligence (AIGC) models, this framework harnesses self-supervised feedback mechanisms to continually enhance the reasoning prowess of large-scale models, thereby facilitating curve reconstruction, horizon forecasting, and potential layer interpretation for existing wells. Notably, this approach demonstrates the feasibility of agent-driven AI-native applications in bridging various oil and gas industry contexts without necessitating extensive fine-tuning. Methods, Procedures, Process The proposed methodology introduces a novel Explainable Self-Supervised Agent Group Framework (ESSA), which constructs an agent cluster to facilitate seamless interaction among business systems, AIGC models, and scenario-specific models. The ESSA's three-tiered agent system operates within a closed-loop configuration, leveraging continuous feedback and self-supervision agent frame as its core mechanism for aiding reinforced learning to deal with the complex task. This architecture encourages iterative refinement of AIGC models, mitigating issues such as inaccuracies, fabrications, and misleading information, while facilitating precise model selection, content generation, and refined evaluations of aged wells. Originality The manuscript underscores the novelty of its platform architecture through two primary contributions: (1) The synergy between the agent ensemble and AIGC models yields an AI-native solution capable of tackling multi-faceted, scenario-spanning challenges in the oil and gas sector. (2) The multi-agent framework fosters a closed-loop ensemble that promotes collective intelligence through self-learning and optimisation, selecting optimal models to continually refine well reassessment processes, significantly curbing operational costs in practical deployments. Reuslts To validate the framework's efficacy, we analysed a dataset comprising 50+25 elderly wells, all exceeding 30 years in age, located across two distinct geological blocks. Our evaluation encompassed model selection accuracy, logging curve reconstruction fidelity, horizon classification (encompassing oil, water, gas, sandstone, mudstone, and oil-water coexistence zones), and the capacity to recommend potential horizons. The outcomes underscore the framework's robust agent ensemble organisation, enabling AIGC models to unlock their full potential through self-supervised prompting engineered within the proposed agent framework. With human-in-the-loop interactions, the framework achieved horizon classification and potential horizon prediction accuracies ranging from 50% to 80% for the mature wells under assessment.