Physics-Informed Agentic AI: An Intelligent Assistant for Production Management and Optimization

Abstract This paper presents the development and deployment of the Digital Oilfield Agentic Production Engineering Assistant (DOAPE Assistant), an intelligent engineering co-pilot designed to enhance production management and optimization in unconventional oil and gas fields. By integrating agentic AI with physics-based modeling, the assistant supports engineers and operators in identifying and resolving suboptimal well and field performance. Built on a multi-agent architecture powered by Large Language Models (LLMs), the system orchestrates specialized agents and delivers context-specific insights tailored to individual wells and fields. The assistant underwent a rigorous year-long training process using high-quality production data and reinforcement learning techniques. Its architecture comprises multiple collaborative agents, each with distinct roles, coordinated by an Orchestration agent that dynamically selects the appropriate downstream agents and physics-based models. This enables the system to generate responses that are both data-informed and grounded in first-principles physics. Notably, the integration of gas lift (GL) and electrical submersible pump (ESP) nodal analysis capabilities allows the assistant to provide physically consistent engineering insights, enhancing its reliability and relevance in field operations. Unlike some conventional AI tools that rely solely on using ML techniques on historical data, the DOAPE Assistant leverages simulations and analyses that are physically and contextually accurate. Distinct from GPT-style chatbots built on closed foundation models, the DOAPE Assistant employs a suite of agents fine-tuned by subject matter experts through Reinforcement Learning from Human Feedback (RLHF). It further incorporates advanced techniques such as agentic retrieval-augmented generation (agentic RAG) and cache-augmented generation (CAG), and in-context learning (ICL), while seamlessly integrating bespoke tools that improve efficiency and precision. For the DOAPE Assistant to be effective, the team envisioned an AI system capable of reasoning through complex steps. The goal was to mimic some of the duties of a junior production engineer supporting a senior engineer in managing a large inventory of wells. In practice, the junior engineer may be asked to run simulations to evaluate different well or artificial lift design scenarios, or to mine data across multiple systems of record to better understand a field's behavior or review past performance to determine operational steps to bring wells back online. Piloted across two fields in Chevron's Permian operations and embedded within a broader Digital Oilfield (DOF) solution, the DOAPE Assistant has demonstrated impressive performance. It responds to complex production engineering queries in under 30 seconds with an accuracy rate exceeding 90%. During the pilot, this AI assistant demonstrated its ability to respond accurately to queries requiring a combination of data mining, simulation analysis and insight generation that can support engineers in their daily surveillance analysis and optimization tasks.