arxiv: 2605.15028 · v1 · submitted 2026-05-14 · 💻 cs.MA

Recognition: no theorem link

Multi-Agentic Approach for History Matching of Oil Reservoirs

Linar Samigullin , Sergei Shumilin , Evgeny Burnaev

Authors on Pith no claims yet

Pith reviewed 2026-05-15 02:59 UTC · model grok-4.3

classification 💻 cs.MA

keywords history matchingoil reservoirsmulti-agent systemslarge language modelsreservoir simulationECLIPSE input decksOPM Flow

0 comments

The pith

PetroGraph deploys specialized LLM agents to automate history matching and reduce reservoir mismatch by 95 percent on SPE1, 69 percent on SPE9, and 13 percent on Norne.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces PetroGraph, a multi-agent framework that splits the history-matching workflow into distinct agents responsible for model review, experimental planning, parameterization, optimization, simulation, and summarization. These agents work with large language models, retrieval-augmented simulator documentation, input-deck validation, and an OPM Flow backend so that users can direct the process through natural language while retaining explicit control over parameters and settings. On three reservoir models of rising complexity the system achieves large reductions in weighted normalized root-mean-square error. The design aims to lower the expertise barrier for operating complex simulation workflows and to supply a reusable base for domain-aware model adaptation.

Core claim

PetroGraph decomposes history matching into a set of cooperating agents that handle model review, planning, parameterization, optimization, simulation execution, and result summarization; when evaluated on the SPE1, SPE9, and Norne models it reduces mismatch by 95 percent, 69 percent, and 13 percent respectively under a weighted normalized root-mean-square-error objective.

What carries the argument

The multi-agent orchestration layer that couples LLM agents to validation routines for modified ECLIPSE input decks and an OPM Flow simulation backend.

If this is right

Users can launch and steer history matching through natural language while keeping explicit oversight of chosen parameters and optimizer settings.
The same agent structure supplies a reusable foundation for extending reservoir-model adaptation to new field data or simulator types.
Automation of parameter selection and optimizer tuning reduces the manual configuration steps that currently limit practical deployment.
Human-in-the-loop checkpoints remain available at key stages to preserve engineering control.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The framework could be tested on additional real-field models to determine whether the performance drop seen on Norne is typical of increasing geological complexity.
Replacing the current OPM Flow backend with a higher-fidelity commercial simulator might improve the absolute mismatch reductions without changing the agent architecture.
The approach may generalize to other inverse calibration tasks that involve large input decks and expensive forward simulations, such as groundwater or geothermal modeling.

Load-bearing premise

Large-language-model agents can consistently produce valid ECLIPSE decks and select physically admissible parameters without introducing systematic errors that the reported mismatch metrics would miss.

What would settle it

Run the generated input decks on an independent simulator and check whether any deck produces non-physical parameter values or simulator crashes that the weighted normalized root-mean-square-error score does not flag.

Figures

Figures reproduced from arXiv: 2605.15028 by Evgeny Burnaev, Linar Samigullin, Sergei Shumilin.

**Figure 1.** Figure 1: Conceptual overview of PetroGraph. A user interacts with a reservoir model through natural language; PetroGraph translates the request into a coordinated multi-agent workflow that reviews input data, selects parameters and bounds, runs optimization and simulation, and returns validated model updates and diagnostic feedback. An important applied domain in which methodological heterogeneity poses a signific… view at source ↗

**Figure 2.** Figure 2: ). These agents are the Reviewer Agent, Planner Agent, Parameterizer Agent, Optimizer Agent, Simulator Agent, and Summarizer Agent. All agents share information through a common state. Detailed descriptions of each agent are provided in the corresponding subsection. The framework is based on the open-source LangGraph framework. A detailed graph architecture is provided in Appendix A [PITH_FULL_IMAGE:figur… view at source ↗

**Figure 3.** Figure 3: SPE1 simple synthetic reservoir model with two wells: one producer and one injector. The operational limits follow the definitions given by Odeh26 and the OPM data repository25. The gas injector is constrained by a maximum surface gas rate of 100 MMscf/day and a maximum bottom-hole pressure (BHP) of 9014 PSIA. The production well operates under a maximum oil rate of 20,000 STBO/day and must maintain a mini… view at source ↗

**Figure 4.** Figure 4: SPE9 complex synthetic reservoir model with 26 wells: 25 producers and one injector. Well controls are based on the specifications of Killough27 and the OPM data repository25. The water injector is limited to a maximum water injection rate of 5,000 STBW/day and a maximum BHP of 4,000 PSIA, referenced to a depth of 9,110 ft. Each producer is initially subject to a maximum oil rate of 1,500 STBO/day, with a … view at source ↗

**Figure 5.** Figure 5: Norne field reservoir model with 31 wells: 22 producers and 9 injectors. a known ground truth while preventing the language model from exploiting any memorized solutions that may exist in its training corpus. For the Norne field case, actual production and injection history is available, and the initial model is compared directly against these field measurements. For SPE1, a three-layer oil-water-gas syste… view at source ↗

**Figure 6.** Figure 6: SPE1 mismatch plots illustrating the deviation between the initial model predictions and the pseudo-historical data. For SPE9, a black-oil model with a more complex well configuration, a uniform multiplier of 1.5 was applied to the horizontal permeability, 0.5 to the vertical permeability, and 1.1 to the porosity field-wide. The initial wNRMSE across the observed responses: WBHP, WOPR, water production rat… view at source ↗

**Figure 7.** Figure 7: SPE9 mismatch plots showing the initial discrepancy relative to the pseudo-historical data. 7/24 [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

**Figure 8.** Figure 8: Norne field mismatch plots comparing the initial model with the actual historical data. All experiments were performed using the Qwen3.5-397B-A17B-FP8 large language model, a FP8-quantized checkpoint of the 397B-parameter model, served locally via vLLM with the temperature set to zero 29 . Results and discussion [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗

**Figure 9.** Figure 9: SPE1 mismatch plots after history matching with PetroGraph, demonstrating a significant improvement [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗

**Figure 10.** Figure 10: SPE9 mismatch plots after history matching with PetroGraph, demonstrating a significant improvement. The complete example of a history matching workflow using the Norne field model and with existing historical data is provided in Appendix C [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗

**Figure 11.** Figure 11: Norne mismatch plots after history matching with PetroGraph, showing a slight but noticeable improvement. expertise. By decomposing the process into specialised agents (review, planning, parameterization, optimization, simulation, summarization) and enriching them with retrieval-augmented generation, domain-specific tools, and human-in-the-loop checkpoints, the system achieves flexible reconfiguration, ad… view at source ↗

read the original abstract

History matching is a central inverse problem in reservoir engineering, where uncertain reservoir parameters must be calibrated against observations. Although automated history matching can reduce manual effort, practical deployment remains difficult because engineers must still configure heterogeneous workflows involving parameter selection, physically admissible bounds, optimizer choice, hyperparameter tuning, simulator execution, and diagnostic reporting. We propose PetroGraph, a multi-agent framework for intelligent reservoir history matching that decomposes this workflow into specialized agents for model review, experimental planning, parameterization, optimization, simulation, and summarization. The system combines large language model agents with domain-specific tools, retrieval-augmented access to simulator documentation, validation of modified ECLIPSE input decks, human-in-the-loop checkpoints, and an OPM Flow-based simulation backend. This design enables users to initiate and steer history matching through natural language while preserving explicit control over selected parameters and optimization settings. We evaluate PetroGraph on three reservoir models of increasing complexity: the synthetic SPE1 model, the faulted SPE9 benchmark, and the real-field Norne model. Using weighted normalized root mean square error as the objective, PetroGraph reduces the mismatch by 95% on SPE1, 69% on SPE9, and 13% on Norne. These results demonstrate that multi-agent orchestration can automate key decisions in history matching, lower the expertise barrier for operating complex simulation workflows, and provide a flexible foundation for extensible, domain-aware reservoir model adaptation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PetroGraph shows a workable multi-agent LLM workflow for reservoir history matching that cuts mismatch on standard benchmarks, though real-field gains stay modest and supporting details are thin.

read the letter

The core takeaway is that this paper builds a multi-agent system, PetroGraph, that splits history matching into agents handling model review, planning, parameterization, optimization, simulation, and summarization. It wires those agents to ECLIPSE-style deck validation, RAG over simulator docs, an OPM Flow backend, and human checkpoints so engineers can drive the process in natural language while keeping explicit control over parameters. On the three test cases the numbers are 95% mismatch drop on SPE1, 69% on SPE9, and 13% on the real Norne field using weighted NRMSE. That decomposition and the tool integration look like the actual new piece; prior multi-agent LLM work exists in other domains, but this specific orchestration for reservoir inverse problems is not referenced as existing art. The practical angle is handled decently: the design directly attacks the usual headaches of setting bounds, choosing optimizers, and producing valid input files. The human-in-the-loop checkpoints and deck validation are sensible safeguards. The results on the synthetic models are clear enough to show the workflow can automate decisions that normally require manual tuning. The soft spots are mostly around evidence strength. The abstract gives no error bars, no ablation runs, and no breakdown of how often agents produced invalid decks or unphysical parameter sets that had to be corrected. The steep fall-off from SPE9 to Norne is consistent with real-field complexity, but it also flags that robustness under more heterogeneous geology or noisy data is not yet demonstrated. Without those checks it is hard to judge whether the reported gains are stable or partly the result of favorable agent behavior on these particular models. This is aimed at reservoir engineers and simulation teams who already run history matching and want to test whether LLM orchestration can reduce setup time. It is not a theoretical advance in optimization or uncertainty quantification, but it is a concrete applied system on standard benchmarks. The work is coherent on its own terms and the central claim rests on measurable mismatch reductions rather than circular fitting, so it deserves a serious referee. I would send it to review with the expectation that the authors add ablations, failure statistics, and more detail on the Norne case.

Referee Report

2 major / 1 minor

Summary. The paper proposes PetroGraph, a multi-agent LLM-based framework that decomposes reservoir history matching into specialized agents (model review, experimental planning, parameterization, optimization, simulation, summarization) augmented with RAG on simulator documentation, ECLIPSE deck validation, human-in-the-loop checkpoints, and an OPM Flow backend. On three benchmarks of increasing complexity (SPE1, SPE9, Norne), the system reports mismatch reductions of 95%, 69%, and 13% respectively under weighted normalized root mean square error, with the claim that natural-language steering can automate key workflow decisions while preserving user control over parameters and settings.

Significance. If the empirical reductions hold under more rigorous validation, the work shows that multi-agent orchestration with domain tools can lower the expertise barrier for operating complex reservoir simulators and offers an extensible template for other engineering inverse problems. The explicit handling of deck validity and human checkpoints is a practical strength that directly targets deployment risks in LLM-driven simulation workflows.

major comments (2)

[Experimental evaluation] Experimental evaluation (abstract and results section): the reported mismatch reductions of 95% (SPE1), 69% (SPE9), and 13% (Norne) are given as single-point figures with no error bars, no repeated runs, and no ablation studies isolating the contribution of RAG, deck validation, or human checkpoints; this leaves the robustness of the central quantitative claims unquantified.
[Methodology] Methodology section: insufficient detail is provided on post-hoc handling of agent-generated invalid decks or unphysical parameter selections, despite these being identified as core risks; without explicit metrics on failure rates or correction success, it is unclear whether the measured improvements reflect reliable automation or selective reporting.

minor comments (1)

[Abstract] The abstract and results could more explicitly state the precise definition of the weighted NRMSE objective and the number of simulation runs underlying each percentage reduction.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback, which highlights important aspects of experimental robustness and methodological transparency. We address each major comment below and indicate the planned revisions.

read point-by-point responses

Referee: [Experimental evaluation] Experimental evaluation (abstract and results section): the reported mismatch reductions of 95% (SPE1), 69% (SPE9), and 13% (Norne) are given as single-point figures with no error bars, no repeated runs, and no ablation studies isolating the contribution of RAG, deck validation, or human checkpoints; this leaves the robustness of the central quantitative claims unquantified.

Authors: We agree that presenting single-point results without error bars, repeated runs, or ablations limits the ability to assess robustness. The reported figures reflect single representative executions chosen to illustrate feasibility across benchmarks of increasing complexity, given the substantial computational cost of full-field simulations such as Norne. In the revised manuscript we will add a dedicated limitations subsection that explicitly discusses LLM stochasticity and the single-run nature of the experiments. We will also include ablation results isolating the contribution of RAG (by comparing runs with and without retrieval) and a brief analysis of human-checkpoint intervention frequency for the smaller SPE1 and SPE9 cases, where additional runs are computationally feasible. revision: partial
Referee: [Methodology] Methodology section: insufficient detail is provided on post-hoc handling of agent-generated invalid decks or unphysical parameter selections, despite these being identified as core risks; without explicit metrics on failure rates or correction success, it is unclear whether the measured improvements reflect reliable automation or selective reporting.

Authors: We acknowledge that the current methodology description does not sufficiently detail the post-hoc correction mechanisms or provide quantitative failure-rate metrics. The system relies on automated ECLIPSE deck validation combined with human-in-the-loop checkpoints to detect and correct invalid decks or unphysical parameter selections. In the revised version we will expand the methodology section with a new subsection that (i) describes the exact correction workflow (re-prompting of the parameterization agent, fallback to default bounds, or human override), (ii) reports the observed rates of invalid decks and unphysical selections across the three benchmarks, and (iii) quantifies the fraction of cases resolved automatically versus those requiring human intervention. This will make clear that the reported mismatch reductions incorporate the full safeguard pipeline rather than selective reporting. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper is an empirical systems paper that proposes a multi-agent LLM framework (PetroGraph) for history matching and reports measured mismatch reductions on fixed benchmark models (SPE1, SPE9, Norne) using weighted NRMSE. No derivation chain, first-principles equations, or predictions are presented that reduce to inputs by construction. The workflow description (agent decomposition, deck validation, RAG, human checkpoints) and results are presented as experimental outcomes rather than self-referential fits or renamed known results. No self-citation load-bearing steps or ansatz smuggling appear in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

The framework rests on the assumption that current LLM agents plus retrieval tools can perform reliable domain-specific engineering tasks; no explicit free parameters are introduced, but the system itself is a new composite entity.

invented entities (1)

PetroGraph multi-agent framework no independent evidence
purpose: Orchestrate the full history-matching workflow via specialized LLM agents and tools
New system introduced by the authors; no independent evidence provided beyond the reported benchmark runs.

pith-pipeline@v0.9.0 · 5559 in / 1137 out tokens · 35717 ms · 2026-05-15T02:59:23.871874+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages · 1 internal anchor

[1]

Samoil, S., Fare, C., Jordan, K. E. & Chen, Z. History matching reservoir models with many objective bayesian optimization. Appl. AI Lett.5, e99, DOI: https://doi.org/10.1002/ail2.99 (2024). https://onlinelibrary.wiley.com/doi/pdf/10.1002/ail2.99

work page doi:10.1002/ail2.99 2024
[2]

Oliver, D. S. & Chen, Y . Recent progress on reservoir history matching: a review.Comput. Geosci.15, 185–221, DOI: 10.1007/s10596-010-9194-2 (2011)

work page doi:10.1007/s10596-010-9194-2 2011
[3]

& Madasu, S

Chai, Z., Nwachukwu, A., Zagayevskiy, Y ., Amini, S. & Madasu, S. An integrated closed-loop solution to assisted history matching and field optimization with machine learning techniques.J. Petroleum Sci. Eng.198, 108204, DOI: https://doi.org/10.1016/j.petrol.2020.108204 (2021). 4.Naji, S. Multi-agent systems: A review study.Ibn Al-Haitham J. for Pure Appl...

work page doi:10.1016/j.petrol.2020.108204 2020
[4]

& Blum, C

Jimenez-Romero, C., Yegenoglu, A. & Blum, C. Multi-agent systems powered by large language models: Applications in swarm intelligence.Front. Artif. Intell.8, DOI: 10.3389/frai.2025.1593017 (2025)

work page doi:10.3389/frai.2025.1593017 2025
[5]

& Liu, T

Chen, S., Liu, Y ., Han, W., Zhang, W. & Liu, T. A survey on llm-based multi-agent systems.arXiv preprint arXiv:2412.17481(2024)

work page arXiv 2024
[6]

arXiv preprint arXiv:2503.23315(2025)

Elrefaie, M.et al.Ai agents in engineering design: A multi-agent framework for aesthetic and aerodynamic car design. arXiv preprint arXiv:2503.23315(2025)

work page arXiv 2025
[7]

& Borrmann, A

Du, C., Esser, S., Nousias, S. & Borrmann, A. Text2bim: Generating building models using a large language model-based multi-agent framework.arXiv preprint arXiv:2403.05615(2024)

work page arXiv 2024
[8]

Shi, L., Qi, B., Luo, J.et al.Aegis: An advanced llm-based multi-agent for intelligent functional safety engineering.arXiv preprint arXiv:2406.18157(2024). 11/24

work page arXiv 2024
[9]

& Kovalenko, I

Lim, J., V ogel-Heuser, B. & Kovalenko, I. Large language model-enabled multi-agent manufacturing systems.arXiv preprint arXiv:2406.18939(2024)

work page arXiv 2024
[10]

Zhang, Y ., Wang, X., Li, Z.et al.An llm-powered multi-agent system for power grid control.arXiv preprint arXiv:2501.01541(2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[11]

& Makris, S

Gkournelos, C., Konstantinou, C. & Makris, S. An llm-based approach for enabling seamless human–robot collaboration in assembly.CIRP AnnalsDOI: 10.1016/j.cirp.2024.04.002 (2024)

work page doi:10.1016/j.cirp.2024.04.002 2024
[12]

A multi-agent debate workflow for construction projects: A cross-stage decision framework.J Build Des Environ 3(2025)

Hao, Y . A multi-agent debate workflow for construction projects: A cross-stage decision framework.J Build Des Environ 3(2025). 15.Alada ˘g, H. & I¸ sık, Z. Multi agent system (mas) use in construction management literature (2019)

work page 2025
[13]

Accessed: 2025-10-15, 2501.01205

Mushtaq, A.et al.Harnessing multi-agent llms for complex engineering problem-solving: A framework for senior design projects (2025). Accessed: 2025-10-15, 2501.01205

work page arXiv 2025
[14]

Accessed: 2025-10-15, 2511.15752

Yan, H.et al.Build ai assistants using large language models and agents to enhance the engineering education of biomechanics (2025). Accessed: 2025-10-15, 2511.15752

work page arXiv 2025
[15]

& Mouzakis, K

Du, H., Thudumu, S., Nguyen, H., Vasa, R. & Mouzakis, K. A comprehensive survey on context-aware multi-agent systems: Techniques, applications, challenges and future directions (2024). Accessed: 2025-10-15, 2402.01968

work page arXiv 2024
[16]

& Pan, S

Yue, L., Somasekharan, N., Zhang, T., Cao, Y . & Pan, S. Foam-agent 2.0: An end-to-end composable multi-agent framework for automating cfd simulation in openfoam (2025). 2509.18178. 20.Chen, Y ., Zhu, X., Zhou, H. & Ren, Z. Metaopenfoam: an llm-based multi-agent framework for cfd (2024). 2407.21320

work page arXiv 2025
[17]

Hou, S., Johnson, R., Makhija, R., Chen, L. & Ye, Y . Autofea: Enhancing ai copilot by integrating finite element analysis using large language models with graph neural networks.Proc. AAAI Conf. on Artif. Intell.39, 24078–24085, DOI: 10.1609/aaai.v39i22.34582 (2025). 22.Zhang, T., Liu, Z., Xin, Y . & Jiao, Y . Mooseagent: A llm based multi-agent framework...

work page doi:10.1609/aaai.v39i22.34582 2025
[18]

opm-data: Datasets for testing opm simulator

OPM (Open Porous Media) Initiative. opm-data: Datasets for testing opm simulator. GitHub repository (2023). Includes the Norne field model case

work page 2023
[19]

Odeh, A. S. Comparison of solutions to a three-dimensional black-oil reservoir simulation problem (includes associated paper 9741 ).J. Petroleum Technol.33, 13–25, DOI: 10.2118/9723-PA (1981). https://onepetro.org/JPT/article-pdf/33/01/ 13/2229415/spe-9723-pa.pdf

work page doi:10.2118/9723-pa 1981
[20]

SPE Reservoir Simulation Sym- posium ofSPE Reservoir Simulation Conference

Ninth SPE Comparative Solution Project: A Reexamination of Black-Oil Simulation, vol. SPE Reservoir Simulation Sym- posium ofSPE Reservoir Simulation Conference. DOI: 10.2118/29110-MS. https://onepetro.org/spersc/proceedings-pdf/ 95RSS/95RSS/SPE-29110-MS/3130251/spe-29110-ms.pdf

work page doi:10.2118/29110-ms
[21]

& Oliver, D

Chen, Y . & Oliver, D. S. History matching of the norne full-field model with an iterative ensemble smoother.SPE Reserv. Eval. & Eng.17, 244–256, DOI: 10.2118/164902-PA (2014). https://onepetro.org/REE/article-pdf/17/02/244/2127718/ spe-164902-pa.pdf. 29.Team, Q. Qwen3.5: Accelerating productivity with native multimodal agents (2026). 12/24 A Appendix: Gr...

work page doi:10.2118/164902-pa 2014
[22]

DATA file consists of sections, sections have keywords

work page
[23]

Use get description for sections and keywords to be more precise

work page
[24]

Planner Agent System Prompt System: You are a Petroleum Reservoir History Matching expert in PetroGraph, specializing in Design of Experiments (DoE)

In the end, when you are done, add reservoir description by tool ‘ add_reservoir_description‘. Planner Agent System Prompt System: You are a Petroleum Reservoir History Matching expert in PetroGraph, specializing in Design of Experiments (DoE). Your task is to design a DoE strategy for two agents: Parameterizer and Optimizer, based on the model descriptio...

work page
[25]

Base estimator (e.g., Gaussian Process)

work page
[26]

Number of initial points

work page
[27]

Initial point generator (e.g., Sobol, Latin Hypercube, Random)

work page
[28]

Acquisition function (e.g., EI, LCB, PI, gp_hedge)

work page
[29]

- Maximum task dimension **< 40 ** (current dimension: {task_dimension})

Number of iterations **Constraints** 15/24 - If using **Sobol** initial point generator: - Number of initial points **must be a power of 2 ** (4, 8, 16, 32, ...). - Maximum task dimension **< 40 ** (current dimension: {task_dimension}). - If task_dimension > 40, **do not choose Sobol **. **Input from Planner Agent ** DoE (Design of Experiments) plan: {opt...

work page
[30]

Call ‘set_bayesian_optimizer_config‘ with your chosen values

work page
[31]

RUNSPEC",

Briefly justify each choice (1-2 sentences per hyperparameter), referencing the DoE plan and constraints. Be concise but precise. Prioritize sampling efficiency and robustness for reservoir history matching. Summarizer Agent System Prompt Summarize this result of PetroGraph‘s History-Matching Agent into a short report. ## Casename: {case_name} ### Reservo...

work page
[32]

Regional pressure behavior across 5 EQUIL regions

work page
[33]

Gas injection efficiency and sweep

work page
[34]

Water breakthrough timing

work page
[35]

Fault transmissibility uncertainties 17/24

work page
[36]

GRID", "PROPS

Well control changes throughout production history The reservoir description has been saved and is ready for the history-matching specialist to begin work. The model contains rich production history data suitable for comprehensive history matching exercises. Planner Parameterization Strategy •Fault Transmissibility Modifiers:Target flow barriers between r...

work page
[37]

Fault transmissibility modifiers — DONE (58 parameters)

work page
[38]

PORV", "MULTPV

Permeability anisotropy ratios — DONE (21PERMZparameters) 3.Water-phase relative permeability endpoints— Need to add 4.Region-based pore volume multipliers— Need to add 19/24 Parameterizer Let me check if there’s a way to add region-based pore volume multipliers. I should look for MULTREGP or similar keywords, or check if we can useMULTIPLYin theEDITsecti...

work page
[39]

target flow barriers between regions to correct water breakthrough timing

Fault Transmissibility Modifiers (58 parameters) All fault transmissibility multipliers in the MULTFLT keyword have been parameterized to control flow barriers between regions. This addresses the Planner’s recommendation to “target flow barriers between regions to correct water breakthrough timing.” Parameters include: •Major faults:FAULT_DE0, FAULT_DE1, ...

work page
[40]

control vertical versus horizontal flow to manage water coning near producers

Permeability Anisotropy Ratios (21 parameters) Vertical permeability multipliers (PERMZ) for all 22 layers have been parameterized to control vertical vs horizontal flow. This addresses the Planner’s recommendation to “control vertical versus horizontal flow to manage water coning near producers.” Parameters include: •Garn formation:PERMZ_GARN1, PERMZ_GAR...

work page
[41]

Add water relative permeability endpoint parameters?

work page
[42]

Add region-based pore volume multipliers?

work page
[43]

base_estimator

Proceed with the current parameterization? 21/24 User Yeap, lets optimize it! Optimizer Optimizer Configuration Based on the parameter dimension and user request, I have selected the following Design of Experiments (DoE) and Bayesian optimization hyperparameters: { "base_estimator": "GP", "n_initial_points": 32, "initial_point_generator": "lhs", "acq_func...

work page 2013
[44]

Executive Summary The PetroGraph History-Matching Agent successfully executed an automated history matching workflow for the Norne Benchmark Reservoir Model. By focusing on flow pathway uncertainties, the agent reduced the global mismatch metric (weighted NRMSE) by approximately13%, improving from an initial value of2.312to a best value of2.013. The optim...

work page
[45]

•Grid:46×112×22 cells (113,344 active cells)

Model Overview •Reservoir:Full-field Norne Model (SPE ATW 2013). •Grid:46×112×22 cells (113,344 active cells). •Fluids:Black oil with compositional effects (DISGAS, V APOIL). •Geology:5 equilibration regions (Garn and Ile-Tilje formations) with complex faulting. •Wells:Multiple producers and gas injectors under historical controls (WCONHIST/WCONINJE)

work page 2013
[46]

Conse- quently, the strategy focused on dynamic flow mismatches rather than static volumetrics

History Matching Strategy Parameterization The Planner Agent identified that water production errors exceeded oil and gas errors by an order of magnitude. Conse- quently, the strategy focused on dynamic flow mismatches rather than static volumetrics. •Primary Parameters:Fault Transmissibility Modifiers (60+ parameters targeting flow barriers between regio...

work page
[47]

Metric Value Status Initial wNRMSE 2.3121 Baseline Best wNRMSE 2.0128Optimized Improvement∼12.98% Successful Reduction

Performance Metrics The objective function was the summed weighted Normalized Root Mean Square Error (wNRMSE) for well rates (WOPR, WWPR, WGPR, WWIR, WGIR). Metric Value Status Initial wNRMSE 2.3121 Baseline Best wNRMSE 2.0128Optimized Improvement∼12.98% Successful Reduction

work page
[48]

Key Parameter Adjustments The Parameterizer Agent selected optimal values within defined bounds for over 70 parameters. Notable adjustments include: • Fault Transmissibility:Significant variations applied to major faults (e.g., FAULT_DE0 optimized to 32.34, FAULT_MEASTto 1.80) to correct water breakthrough timing. • Vertical Permeability:Zone-specific adj...

work page
[49]

•Well Comparisons: –WBHP_compare.png(Bottom Hole Pressure)

Deliverables & Output All simulation results and diagnostic plots are stored in the following directory: /output/2026-04-07-14-11-13/ Generated Artifacts: •Best Case Folder:.../tasks/100 •Metric Evolution:metric_evolution.png(Convergence plot). •Well Comparisons: –WBHP_compare.png(Bottom Hole Pressure). –WOPR_compare.png(Oil Production Rate). –WGPR_compar...

work page 2026
[50]

Recommendations Based on the optimization results:

work page
[51]

Validate Water Breakthrough:Review the WWPR matches specifically, as this was the primary driver for parameter selection

work page
[52]

Fault Connectivity:The high sensitivity to fault multipliers suggests significant uncertainty in the static fault model; geological review of major fault seals is recommended

work page
[53]

Consider refining parameter bounds around the current best values for a second optimization pass

Further Iteration:While improved, a wNRMSE of 2.01 suggests residual mismatch. Consider refining parameter bounds around the current best values for a second optimization pass. 24/24

work page