Language Model Agents Retrieve Feasible Reservoir Development Plans From Knowledge Graphs

Engineering researchers have introduced an artificial intelligence framework named PHARAOH that deploys a large language model agent to retrieve multi-objective oilfield development plans from hierarchical knowledge graphs. Developing an oilfield presents a complex decision problem where candidate operational strategies must carefully balance overall hydrocarbon recovery against expected economic return while adhering strictly to operational constraints. Traditional retrieval techniques that rank evidence based solely on textual or structural similarity cannot represent such trade-offs or screen out infeasible proposals, whereas running full-physics reservoir simulations remains computationally prohibitive for interactive exploration.

To address these challenges, the PHARAOH framework couples an autonomous agent with a simulation-trained surrogate model capable of estimating critical performance metrics, including ultimate hydrocarbon recovery and net present value. Rather than treating plan generation as a continuous numerical optimization task, the system structures development alternatives across a multi-resolution knowledge graph and screens candidates within this pre-built decision space. The agent navigates this graph by pairing global anchoring with type-filtered hierarchical refinement, systematically narrowing down potential strategies to locate balanced interventions.

During its search process, the agent executes parallel trajectories across the decision network and subsequently aggregates the resulting exploration paths into a diverse set of candidate recommendations. To enable practical deployment without continuous access to costly computational infrastructure, the authors distilled the exploratory search policy of the primary teacher agent into a compact student model containing 7 billion parameters.

The researchers evaluated the system on PetroGraph, an experimental benchmark constructed from three public reservoir simulation models. Across these test environments, PHARAOH demonstrated the highest average retrieval performance among the evaluated baseline methods. Specifically, the framework improved Hit@1 by 8.8 and Recall@20 by 7.1 when compared against a surrogate-guided non-agentic search routine, demonstrating that agentic navigation across structured knowledge graphs yields more accurate plan identification than standard surrogate ranking.

To verify the practical viability of the selected strategies, the researchers conducted full-physics re-simulations on the highest-ranked candidates. These validation runs confirmed that 14 out of the 15 top-ranked development plans successfully satisfied the necessary operational constraints, although the authors noted that candidate plans situated near boundary thresholds remained sensitive to estimation errors from the surrogate model. Furthermore, tests on the distilled 7-billion-parameter student model revealed that it retained approximately 85 percent of the original teacher model’s Hit@1 performance while operating at a fraction of the baseline inference cost.

The authors emphasized that their empirical findings remain limited to evaluations conducted across three specific simulation benchmarks.

Subject of Research: Artificial intelligence framework for retrieving multi-objective oilfield development plans from knowledge graphs

Article Title: Large language model agents with simulation-trained surrogates retrieve multi-objective reservoir development plans from hierarchical knowledge graphs

Article References: Liu, Z., & Zhou, F. (2026). Large language model agents with simulation-trained surrogates retrieve multi-objective reservoir development plans from hierarchical knowledge graphs. Scientific Reports. https://doi.org/10.1038/s41598-026-78530-1

Image Credits: AI Generated

DOI: 10.1038/s41598-026-78530-1

Keywords: artificial intelligence, reservoir engineering, oilfield development, large language models, knowledge graphs, multi-objective optimization, surrogate modeling, PHARAOH, PetroGraph, decision support, model distillation, reservoir simulation, net present value, hydrocarbon recovery, computational retrieval, autonomous agents