Large Language Model Empowered Automated Reservoir Agent: A Win-Win Strategy for Reservoir Intelligent Management and Transformation

Abstract

Abstract The upstream oil and gas industry is currently undergoing digital transformation, presenting numerous opportunities as well as significant challenges, particularly in the digitalization of reservoir management with Large Language Model (LLM). Meanwhile, reservoir management is a field rich in specialized knowledge and technical complexity. Most of the LLMs available todayare designed for general domains and lack a deep understanding of specialized knowledge, resulting in suboptimal performance in the realm of automated reservoir agents.To address this issue, we introduce the Knowledge-Retrieval–DeepSeek System (KRDS), an LLM-powered automated reservoir agent that integrates a two-stage knowledge retriever with a retrieval-augmented conversational DeepSeek LLM deployed on-premises to ensure data sovereignty. The KRDS automated reservoir agent consists of two components: the two-stage knowledge retrieval and recall model and the retrieval-augmented conversational DeepSeek LLM framework. By injecting the top-ranked passages returned by the retriever, the LLM acquires reservoir-engineering context that is otherwise absent from general-purpose models.Then the retrieval-augmented conversational DeepSeek LLM frameworkinvokes tool functions to retrieve reservoir properties, production history, future conditions, and more, in order to provide feedback on the query of reservoir engineer. In this way, the automated reservoir agent calls the relevant tool functions in the system based on the query, completes reservoir management tasks, and presents the results to the reservoir engineer in a question-and-answer format. To verify the effectiveness and reliability of the KRDS automated reservoir agents, the intelligent management of a typical carbonate reservoir in the Persian Gulf Basin of the Middle East is used as an example. The establishment and application of KRDS automated reservoir agentsconsists of four sections: dataset, evaluation index, experimental facility, and main procedures. KRDS is furnished with eighteen specialized reservoirmanagement tool functions including ‘get_permeability,' ‘get_porosity,' ‘get_production_rate,' ‘well trajectory,' and others, tasked to answer forty-five scenario-based questions reflecting daily operational practice. For the query ‘Check the production history of A1 well,' KRDS automatically renders oil-, water- and injection-rate curves, achieving the goal of utilizing the LLM to empower the automated reservoir agent for reservoir intelligent management and transformation.Quantitatively, the success rate and correct path per hundred experiments are 83% and 87%, respectively. To the best of our knowledge, KRDS represents the first domain-tailored LLM agent that autonomously executes reservoir-management tasks while maintaining enterprise-level data security. The results demonstrate that coupling targeted retrieval with an open-access LLM offers a feasible pathway toward intelligent reservoir management and digital transformation in the upstream sector.

Ähnliche Arbeiten