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

Optimizing oilfield development is a high-dimensional, multi-objective decision problem: candidate plans must balance hydrocarbon recovery against economic return while satisfying operational constraints. Retrieval methods that rank evidence by textual or structural similarity cannot express such trade-offs or reject infeasible plans, whereas full-physics simulation is too costly to run inside an interactive search. We present PHARAOH, a framework in which a large language model agent retrieves development plans over a multi-resolution knowledge graph, guided by a simulation-trained surrogate model that estimates objectives such as ultimate recovery and net present value. The agent combines global anchoring with type-filtered hierarchical refinement and aggregates parallel search trajectories into a set of diverse recommendations; its policy is then distilled into a compact student model. PHARAOH screens and ranks candidates within a pre-built decision space rather than solving a continuous optimization problem. On PetroGraph, a benchmark built from three public reservoir simulation models, PHARAOH attains the best average performance among the compared baselines, improving Hit@1 by 8.8 and Recall@20 by 7.1 over a surrogate-guided non-agentic search. Full-physics re-simulation shows that 14 of 15 top-ranked plans satisfy the operational constraints, although near-boundary plans remain sensitive to surrogate error. The distilled 7B student retains about 85% of the teacher’s Hit@1 at a fraction of the inference cost. These findings are limited to three simulation benchmarks; field-scale validation remains future work.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-12
DOI
https://doi.org/10.1038/s41598-026-78530-1
Primary Topic
Reservoir Engineering and Simulation Methods
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

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

刘志坚, Fei Zhou
Scientific Reports
Reservoir Engineering and Simulation Methods
article

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

刘志坚, Fei Zhou
article en

Abstract

Optimizing oilfield development is a high-dimensional, multi-objective decision problem: candidate plans must balance hydrocarbon recovery against economic return while satisfying operational constraints. Retrieval methods that rank evidence by textual or structural similarity cannot express such trade-offs or reject infeasible plans, whereas full-physics simulation is too costly to run inside an interactive search. We present PHARAOH, a framework in which a large language model agent retrieves development plans over a multi-resolution knowledge graph, guided by a simulation-trained surrogate model that estimates objectives such as ultimate recovery and net present value. The agent combines global anchoring with type-filtered hierarchical refinement and aggregates parallel search trajectories into a set of diverse recommendations; its policy is then distilled into a compact student model. PHARAOH screens and ranks candidates within a pre-built decision space rather than solving a continuous optimization problem. On PetroGraph, a benchmark built from three public reservoir simulation models, PHARAOH attains the best average performance among the compared baselines, improving Hit@1 by 8.8 and Recall@20 by 7.1 over a surrogate-guided non-agentic search. Full-physics re-simulation shows that 14 of 15 top-ranked plans satisfy the operational constraints, although near-boundary plans remain sensitive to surrogate error. The distilled 7B student retains about 85% of the teacher’s Hit@1 at a fraction of the inference cost. These findings are limited to three simulation benchmarks; field-scale validation remains future work.

Scientific Reports
Southwest Petroleum University (CN)
Reservoir Engineering and Simulation Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon