PetroGraph multi-agent system cuts oil reservoir mismatch by 95%

A team of researchers from Skoltech and other institutions has introduced PetroGraph, a multi-agent framework designed to automate the complex inverse problem of history matching in oil reservoir engineering. Traditionally, engineers must manually configure heterogeneous workflows—selecting parameters, setting physical bounds, choosing optimizers, tuning hyperparameters, running simulators, and generating reports. PetroGraph decomposes this process into six specialized agents: model review, experimental planning, parameterization, optimization, simulation, and summarization. These agents are powered by large language models and access domain-specific tools, including retrieval-augmented generation (RAG) for simulator documentation, validation of modified ECLIPSE input decks, and a backend built on OPM Flow.

In evaluations on three reservoir models of increasing complexity—the synthetic SPE1, the faulted SPE9 benchmark, and the real-field Norne model—PetroGraph achieved weighted normalized root mean square error reductions of 95%, 69%, and 13% respectively. Users can initiate and steer the entire matching process through natural language while retaining explicit control over selected parameters and optimization settings. The framework demonstrates that multi-agent orchestration can automate key decisions, significantly lower the expertise barrier for operating complex simulation workflows, and provide a flexible foundation for extensible, domain-aware reservoir model adaptation. This work, submitted to arXiv in May 2026, highlights a promising application of AI agents in geoscience and energy production.