Pith. sign in

REVIEW 2 major objections 6 minor 43 references

A learned policy calibrates a full-physics reservoir simulator by choosing where to probe, without gradients or surrogate models.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 17:10 UTC pith:J4JTV65K

load-bearing objection A clean, honest proof of concept for SAC as a simulator-in-the-loop proposal distribution on PUNQ-S3—but without an equal-budget random/DFO baseline, the 'learning' claim is not yet supported. the 2 major comments →

arxiv 2608.01986 v1 pith:J4JTV65K submitted 2026-08-03 cs.CE

Physics-Constrained Soft Actor-Critic for Simulator-in-the-Loop Petroleum Reservoir History Matching

Nam-Phong Huu Nguyen , Duy-Dong Nguyen , Tho Quan This is my paper
classification cs.CE
keywords history matchingreservoir calibrationsimulator-in-the-loopsoft actor-criticblack-box optimizationPUNQ-S3physics-constrained RLwell-level diagnostics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that expensive scientific simulators can be calibrated by reinforcement learning in a simulator-in-the-loop fashion, without gradients and without training a surrogate model to replace the simulator. On the PUNQ-S3 reservoir benchmark, the authors wrap the CMG IMEX full-physics simulator as an environment and use Soft Actor-Critic to learn a stochastic proposal distribution over five continuous parameters: a porosity multiplier, three directional permeability multipliers, and a well-skin factor. The reward combines range-normalized production mismatch across oil, water, gas, and bottom-hole pressure with penalties for physically invalid properties. Within a 200-call budget, the best valid candidate reaches a category-macro NMSE of 0.0285, R2 of 0.9324, and a bounded match score of 97.23%. The deeper point is that the policy learns where to evaluate rather than how to replace physics, and that well-level diagnostics expose localized water-rate failures that pooled metrics hide.

Core claim

On its own terms, the paper claims a new mode of calibration: an RL agent interacting with a trusted simulator as a black-box environment can discover good candidates for history matching in a compact continuous parameter space. The central discovery is that under a fixed 200-simulator-call budget, the best physically valid candidate found by Soft Actor-Critic matches PUNQ-S3 production history with category-macro NMSE 0.0285, R2 0.9324, and bounded match score 97.23%, while the same results show that aggregate metrics conceal strong local mismatches—PRO-1's water-rate match score drops to 13.31 and PRO-15's to 56.87. The design separates the learned proposal policy, the physics-validity fee

What carries the argument

The load-bearing mechanism is the simulator-in-the-loop reward. Each action is a normalized five-vector mapped linearly to bounds for porosity, I-, J-, K-permeability, and skin; CMG IMEX evaluates the generated case; simulated and observed oil, water, gas, and bottom-hole-pressure records are matched by well and nearest time within a 30-day tolerance; and the reward is the negative of a range-normalized weighted squared error minus a nonnegative physics penalty. Soft Actor-Critic operates as an adaptive stochastic proposal distribution: off-policy replay reuses every simulator call, twin critics approximate the penalized simulator-return landscape, and maximum-entropy exploration keeps the p

Load-bearing premise

The whole result depends on the assumption that five global multipliers—one porosity, three directional permeabilities, and a skin factor—are enough to reproduce PUNQ-S3's production history; if a reservoir requires spatially varying fields, this compact parameterization cannot represent the needed controls.

What would settle it

Run the same 200-call pipeline on a reservoir where the true permeability field is spatially heterogeneous (e.g., a channelized model), and check whether any retained valid candidate produces a well-level water-rate match score above 90 for every well. Failure would show the five-global-multiplier parameterization, not the policy, is the limiting factor.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Full-physics calibration can be done without gradients or surrogate training: every retained candidate is verified by the simulator, and each candidate costs exactly one simulator call.
  • A 200-call budget suffices on PUNQ-S3 for the five-parameter space: the best valid candidate gives category-macro NMSE 0.0285, R2 0.9324, and bounded match score 97.23%.
  • Well-level reporting is necessary: pooled water-rate scores hide failures at PRO-1 and PRO-15, where match scores drop to 13.31% and 56.87%.
  • The same interface—case generator, output parser, mismatch metric, and validity check wrapping any domain simulator—can be repurposed for other expensive-simulator calibration tasks.

Where Pith is reading between the lines

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

  • The bandit-like, one-evaluation-per-episode design suggests SAC is being used as a stochastic search distribution rather than a temporal-credit-assignment learner; a fair test would swap in equal-budget random search or differential evolution to see what the learned critic actually adds.
  • The five-global-multiplier parameterization is likely the ceiling for this benchmark; extending to spatial fields would require latent or structured actions, which the paper does not claim to support.
  • The proposed robust reward (category loss plus worst-well penalty) is plausible but untested; rerunning the same budget with that reward would show whether PRO-1 and PRO-15 improve without hurting aggregate scores.
  • Because every candidate is full-physics validated, wall-clock cost is dominated by the simulator, so the practical win depends on whether the learned proposal finds good candidates earlier than derivative-free search; the paper does not report best-so-far curves, which would be the decisive evidence.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. This paper proposes a Soft Actor-Critic (SAC) approach to petroleum reservoir history matching, treating the CMG IMEX full-physics simulator as a Gymnasium environment. The agent proposes five global reservoir parameters (porosity multiplier, I/J/K permeability multipliers, and skin factor), each proposal is evaluated by one IMEX run, and a reward combining a range-normalized multi-response mismatch with a physical-invalidity penalty is used to train the policy. On the PUNQ-S3 benchmark under a 200-call budget, the best valid candidate achieves category-macro NMSE 0.0285, R2 0.9324, and a bounded match score 97.23%. The paper also reports well-level diagnostics showing localized water-rate failures at PRO-1 and PRO-15, and it compares GOR trajectories qualitatively with prior EnKF/ES-MDA and GP-VARS results. The authors explicitly frame the work as a proof of concept, not a claim of general superiority.

Significance. If the result holds, the paper demonstrates a useful integration of a commercial full-physics simulator with off-policy RL, and it is honest about the limits of a single-benchmark demonstration. Strengths include the explicit separation between learned proposal generation and simulator-grounded final selection, the use of replay to reuse expensive simulator calls, and the well-level analysis that reveals failures hidden by pooled metrics. The main unresolved question is whether SAC contributes anything over inexpensive search in the 5D parameter space; the paper itself identifies equal-budget multi-seed comparisons as necessary.

major comments (2)
  1. [Experiment Setup; Discussion] This is the decisive missing control. Table 1 reports the headline NMSE/R2/Cmatch for a single SAC run with 200 full-physics calls (Experiment Setup). The only numerical context is Table 3, which the authors themselves label as 'contextual' because protocols differ. The Discussion states that 'Controlled multi-seed comparisons with equal-budget random search and derivative-free optimizers remain necessary.' In a 5-dimensional bounded box, 200 evaluations from space-filling or evolutionary search can often reach comparable NMSE, especially if the response surface is smooth; without such a baseline, the metrics cannot be attributed to the learned proposal distribution, and the RL machinery could be superfluous. A single seed also provides no variance estimate. Please add best-so-far curves over simulator calls for SAC versus at least one space-filling/random baseline and one derivative-fre
  2. [Methodology, Objective and Reward Design] The method is named 'physics-constrained,' but the physics penalty P_phys(m) is never specified. In the Methodology subsection 'Objective and Reward Design,' r(m) = -E(m) - P_phys(m), but the text only says that P_phys penalizes 'violations of the prescribed physical-property constraints' and that failed simulations receive 'a large penalty.' The reader is not told which properties are checked, how violations are quantified, or what numerical values are used. This matters because Algorithm 1 uses P_t both in the reward and in the valid-candidate gate (line 13). Please provide the exact definition of P_phys, the validity checks, and the penalty magnitudes, or point to released code.
minor comments (6)
  1. [Evaluation Protocol] Equation (3) includes an epsilon term in the NMSE denominator, but the value of epsilon is never specified for the evaluation metric. State its value and whether it is the same epsilon as in the reward loss.
  2. [Algorithm 1] The algorithm requires the base-model error E0 but does not define how it is computed. Specify the computation (e.g., the same l_q at a0=0) and report E0 so readers can gauge the difficulty of the 5D problem.
  3. [Table 1] Clarify how the 'Category macro' row is computed. The NMSE and R2 values match the arithmetic mean of the four primary responses, but Cmatch appears to be 100/(1+macro NMSE) rather than the mean of per-response Cmatch values; state this explicitly.
  4. [Equation (6)] The notation 'lambda max max_{w,c}' is awkward; it should be lambda times the maximum over w,c of L_{w,c} (or similar).
  5. [Table 3 vs Table 1] Table 3 reports mean BHP R2 = 0.8452 over four wells, whereas Table 1 reports pooled BHP R2 = 0.9378. This is not necessarily inconsistent, but the discrepancy should be explained (e.g., by noting that pooling across wells changes the variance normalization) to avoid confusion.
  6. [Reproducibility] The paper does not report wall-clock time per IMEX run or the total experiment time, nor the number of invalid/failed simulator runs. It also does not mention a code repository. For a methods paper, releasing the Gymnasium wrapper and reward computation would aid reproducibility.

Circularity Check

0 steps flagged

No circularity: the reported metrics are post-hoc simulator outputs, the reward objective is explicitly distinguished from the evaluation metric, and no load-bearing self-citation or imported uniqueness claim appears.

full rationale

The derivation chain is self-contained and empirical: the SAC policy proposes a 5D parameter vector, CMG IMEX executes the full-physics case, a reward combining a range-normalized mismatch and a physics penalty is returned, and the best valid candidate is selected by the production mismatch E(m). The paper explicitly separates the training objective from the reported metric: 'We denote this loss by ℓq to distinguish the reward objective from the evaluation NMSE defined later.' The final NMSE, R2, and Cmatch are computed from simulator outputs after training, not from the reward formula, and Cmatch is defined as a monotone transform of NMSE (Eq. 5), not as a new predictive claim. No step defines a prediction in terms of the quantity it purports to predict, no fitted parameter is renamed as a prediction, and the paper contains no self-citations that carry a load-bearing argument. The acknowledged absence of equal-budget random-search or derivative-free baselines is an evidence limitation—explicitly stated in the Discussion as 'Controlled multi-seed comparisons with equal-budget random search and derivative-free optimizers remain necessary'—but it is not a circularity. The central claim is that this particular RL-based search found a strong calibration under a 200-call budget; that claim is supported by direct simulator evaluation and does not reduce to its inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

The central claim depends on hand-chosen parameter ranges, equal reward weights, a 30-day matching tolerance, and standard SAC hyperparameters. The key modeling risk is the assumption that a five-parameter global scaling of the base model can represent the true reservoir behavior. No new physical entities are introduced.

free parameters (4)
  • Parameter ranges for the five global multipliers = porosity [0.80, 1.20], permeability [0.30, 3.00] each direction, skin [-5.00, 15.00]
    Chosen by hand and not derived from a prior distribution or calibration data. These bounds define the search space and directly constrain what solutions are reachable.
  • Response weights w_q in the reward = all equal (0.25 each)
    The total mismatch is a weighted average with equal weights for BHP, oil, water, and gas. This design choice shapes the objective and the final candidate selection.
  • Time matching tolerance = 30 days
    Simulated and observed records are matched by nearest time within a 30-day tolerance; this affects the number of matched points and the computed per-response loss.
  • SAC hyperparameters = learning rate 3e-4, replay capacity 10,000, batch size 64
    These are standard values used by Stable-Baselines3 defaults, but they are not tuned for this problem and may influence the outcome substantially.
axioms (3)
  • domain assumption CMG IMEX provides physically faithful simulations of PUNQ-S3.
    The entire method treats the simulator as the authoritative source of truth. This is standard in reservoir engineering but is a fidelity assumption that is not independently verified in the paper.
  • domain assumption PUNQ-S3 is a representative benchmark for reservoir history matching.
    The proof of concept is demonstrated on a single synthetic reservoir. Generalization to other reservoirs is assumed but not tested.
  • ad hoc to paper The 5-parameter global multiplier parameterization is sufficient to match the observed history.
    The method succeeds only if the true reservoir variability can be captured by these global multipliers. The paper does not validate this against a richer parameterization or show that the failure wells could be fixed with more parameters.

pith-pipeline@v1.3.0-daily-deepseek · 203 in / 8936 out tokens · 140225 ms · 2026-08-04T17:10:44.877515+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Physics-Constrained Soft Actor-Critic for Simulator-in-the-Loop Petroleum Reservoir History Matching." pith.science (2026). https://pith.science/paper/J4JTV65K

@misc{pith2026260801986,
  author       = {Pith},
  title        = {Pith review of: Physics-Constrained Soft Actor-Critic for Simulator-in-the-Loop Petroleum Reservoir History Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J4JTV65K}},
  note         = {Machine review of arXiv:2608.01986}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Many scientific calibration problems expose only an expensive executable simulator, making gradients unavailable and large-scale training-data generation impractical. We study petroleum reservoir history matching as an instance of this broader AI problem and formulate it as physics-constrained, simulator-in-the-loop policy search. Our method wraps the CMG IMEX full-physics simulator as a Gymnasium environment and uses Soft Actor-Critic (SAC) to learn a stochastic proposal distribution over continuous porosity, directional-permeability, and well-skin parameters. Each interaction generates and executes a reservoir case, aligns simulated and observed production responses, and returns a reward that combines multi-response mismatch with penalties for physically invalid properties. Off-policy replay reuses costly simulator feedback, while maximum-entropy learning preserves exploration. Unlike forward-surrogate approaches, the policy learns where to evaluate rather than learning to replace the simulator; every retained candidate is validated by IMEX. Under a 200-call budget on PUNQ-S3, the best valid candidate achieves category-macro NMSE $0.0285$, $R^2=0.9324$, and a bounded match score of $97.23\%$. Well-level analysis further exposes localized water-rate failures hidden by pooled metrics. These results establish a full-physics proof of concept for reinforcement-learning-based calibration of expensive scientific simulators.

Figures

Figures reproduced from arXiv: 2608.01986 by Duy-Dong Nguyen, Nam-Phong Huu Nguyen, Tho Quan.

Figure 1
Figure 1. Figure 1: Simulator-in-the-loop history-matching workflow. The SAC agent transforms the current state into continuous parameter [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Simulated and observed GOR and water-cut trajectories for PRO-1 and PRO-12. Black curves show the matched [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: GOR comparison for PRO-1 and PRO-15 on PUNQ-S3. Panels (a,d) show RL-SAC; panels (b,e) reproduce [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 1 canonical work pages

  1. [1]

    Intelligent Systems with Applications , volume =

    Alguliyev, Rasim and Aliguliyev, Ramiz and Imamverdiyev, Yadigar and Sukhostat, Lyudmila , title =. Intelligent Systems with Applications , volume =. 2022 , doi =

    2022

  2. [2]

    , title =

    Rana, Sachin and Ertekin, Turgay and King, Gregory R. , title =. Computers & Geosciences , volume =. 2018 , doi =

    2018

  3. [3]

    Computers & Geosciences , volume =

    Badawi, Daniel and Gildin, Eduardo , title =. Computers & Geosciences , volume =. 2025 , doi =

    2025

  4. [4]

    and Chen, Yan , title =

    Oliver, Dean S. and Chen, Yan , title =. Computational Geosciences , volume =. 2011 , doi =

    2011

  5. [5]

    Floris, F. J. T. and Bush, M. D. and Cuypers, Maarten and Roggero, F. and Syversveen, Anne Randi , title =. Petroleum Geoscience , volume =. 2001 , doi =

    2001

  6. [6]

    and Jurman, Giuseppe , title =

    Chicco, Davide and Warrens, Matthijs J. and Jurman, Giuseppe , title =. PeerJ Computer Science , volume =. 2021 , doi =

    2021

  7. [7]

    Kvålseth , title =

    Tarald O. Kvålseth , title =. The American Statistician , volume =. 1985 , doi =

    1985

  8. [8]

    and Reynolds, Albert C

    Emerick, Alexandre A. and Reynolds, Albert C. , title =. Computers & Geosciences , volume =. 2013 , doi =

    2013

  9. [9]

    and Stordal, Andreas S

    Evensen, Geir and Raanes, Patrick N. and Stordal, Andreas S. and Hove, Jeffrey , title =. Frontiers in Applied Mathematics and Statistics , volume =. 2019 , doi =

    2019

  10. [10]

    and Nœvdal, Geir and Oliver, Dean S

    Aanonsen, Sigurd I. and Nœvdal, Geir and Oliver, Dean S. and Reynolds, Albert C. and Vallès, Brice , title =. SPE Journal , volume =. 2009 , doi =

    2009

  11. [11]

    A review of automatic history matching , journal =

    Farah Nabeel Abdulrazzaq and Omar Falih Hasan , keywords =. A review of automatic history matching , journal =. 2023 , note =. doi:10.1016/j.matpr.2021.07.395 , url =

    work page doi:10.1016/j.matpr.2021.07.395 2023

  12. [12]

    National Science Review , volume =

    Tang, Ke and Yao, Xin , title =. National Science Review , volume =. 2024 , doi =

    2024

  13. [13]

    Protection and Control of Modern Power Systems , volume =

    Hu, Chunchao and Cai, Zexiang and Zhang, Yanxu and Yan, Rudai and Cai, Yu and Cen, Bowei , title =. Protection and Control of Modern Power Systems , volume =. 2022 , doi =

    2022

  14. [14]

    IEEE Transactions on Neural Networks and Learning Systems , volume =

    Wang, Xu and Wang, Sen and Liang, Xingxing and Zhao, Dawei and Huang, Jincai and Xu, Xin and Dai, Bin and Miao, Qiguang , title =. IEEE Transactions on Neural Networks and Learning Systems , volume =. 2024 , doi =

    2024

  15. [15]

    Proceedings of the 35th International Conference on Machine Learning , series =

    Haarnoja, Tuomas and Zhou, Aurick and Abbeel, Pieter and Levine, Sergey , title =. Proceedings of the 35th International Conference on Machine Learning , series =. 2018 , publisher =

    2018

  16. [16]

    Journal of Machine Learning Research , volume =

    Bergstra, James and Bengio, Yoshua , title =. Journal of Machine Learning Research , volume =. 2012 , doi =

    2012

  17. [17]

    Journal of Global Optimization , volume =

    Storn, Rainer and Price, Kenneth , title =. Journal of Global Optimization , volume =. 1997 , doi =

    1997

  18. [18]

    Heliyon , volume =

    Ma, Xianlin and Zhong, Rong and Zhan, Jie and Zhou, Desheng , title =. Heliyon , volume =. 2024 , doi =

    2024

  19. [19]

    and Jung, Hyungsik and Pyrcz, Michael J

    Jo, Honggeun and Pan, Wen and Santos, Javier E. and Jung, Hyungsik and Pyrcz, Michael J. , title =. Journal of Petroleum Science and Engineering , volume =. 2021 , doi =

    2021

  20. [20]

    Efficient deep-learning-based history matching for fluvial channel reservoirs , journal =

    Suryeom Jo and Hoonyoung Jeong and Baehyun Min and Changhyup Park and Yeungju Kim and Seoyoon Kwon and Alexander Sun , keywords =. Efficient deep-learning-based history matching for fluvial channel reservoirs , journal =. 2022 , issn =. doi:https://doi.org/10.1016/j.petrol.2021.109247 , url =

    arXiv 2022

  21. [21]

    SPE Journal , volume =

    Zhang, Zhen and He, Xupeng and AlSinan, Marwah and Kwak, Hyung and Hoteit, Hussein , title =. SPE Journal , volume =. 2023 , doi =

    2023

  22. [22]

    Applied Energy , volume =

    Li, Hao and Misra, Siddharth , title =. Applied Energy , volume =. 2021 , doi =

    2021

  23. [23]

    and Alomar, Abdullah O

    Alolayan, Omar S. and Alomar, Abdullah O. and Williams, John R. , title =. Energies , volume =. 2023 , doi =

    2023

  24. [24]

    Advances in Geo-Energy Research , volume =

    Liu, Piyang and Zhang, Kai and Yao, Jun , title =. Advances in Geo-Energy Research , volume =. 2023 , doi =

    2023

  25. [25]

    Processes , volume =

    Liu, Botao and Xu, Tengbo and Xu, Yunfeng and Zhao, Hui and Li, Bo , title =. Processes , volume =. 2025 , doi =

    2025

  26. [26]

    Processes , volume =

    Tong, Kaijun and Song, Wentong and Chen, Han and Guo, Sheng and Li, Xueyuan and Sun, Zhixue , title =. Processes , volume =. 2024 , doi =

    2024

  27. [27]

    SPE Journal , volume =

    Aslam, Billal and Zhang, Yanhui and Hoteit, Ibrahim and Yan, Bicheng , title =. SPE Journal , volume =. 2025 , doi =

    2025

  28. [28]

    SPE Journal , year =

    Abdulkareem, Usman and Nguyen, Quang Minh and Adeyemi, Ahmed and Onur, Mustafa , title =. SPE Journal , year =

  29. [29]

    and Gross, Stephen J

    Li, Boxiao and Bhark, Eric W. and Gross, Stephen J. and Billiter, Travis C. and Dehghani, Kaveh , title =. SPE Journal , volume =. 2019 , doi =

    2019

  30. [30]

    and Emerick, Alexandre A

    Lima, Mateus M. and Emerick, Alexandre A. and Ortiz, Carlos E. P. , title =. Computational Geosciences , volume =. 2020 , doi =

    2020

  31. [31]

    and Lorentzen, Rolf J

    Luo, Xiaodong and Stordal, Andreas S. and Lorentzen, Rolf J. and Nævdal, Geir , title =. SPE Journal , volume =. 2015 , doi =

    2015

  32. [32]

    and Oliver, Dean S

    Li, Ruijian and Reynolds, Albert C. and Oliver, Dean S. , title =. SPE Journal , volume =. 2003 , doi =

    2003

  33. [33]

    and Arts, R

    Peters, E. and Arts, R. J. and Brouwer, G. K. and Geel, C. R. , title =. SPE Reservoir Evaluation & Engineering , volume =. 2010 , doi =

    2010

  34. [34]

    2020 , issn =

    Development of ensemble smoother–neural network and its application to history matching of channelized reservoirs , journal =. 2020 , issn =. doi:https://doi.org/10.1016/j.petrol.2020.107159 , url =

    arXiv 2020

  35. [35]

    , title =

    Sun, Wenyue and Durlofsky, Louis J. , title =. Mathematical Geosciences , volume =. 2017 , doi =

    2017

  36. [36]

    , title =

    Chen, Yan and Oliver, Dean S. , title =. Computational Geosciences , volume =. 2013 , doi =

    2013

  37. [37]

    , title =

    Emerick, Alexandre A. , title =. Journal of Petroleum Science and Engineering , volume =. 2016 , doi =

    2016

  38. [38]

    Computational Geosciences , volume =

    Mohd Razak, Syamil and Jafarpour, Behnam , title =. Computational Geosciences , volume =. 2020 , doi =

    2020

  39. [39]

    Canchumuni, Smith W. A. and Emerick, Alexandre A. and Pacheco, Marco Aur. History Matching Geological Facies Models Based on Ensemble Smoother and Deep Generative Models , journal =. 2019 , doi =

    2019

  40. [40]

    , title =

    Tang, Meng and Liu, Yimin and Durlofsky, Louis J. , title =. Journal of Computational Physics , volume =. 2020 , doi =

    2020

  41. [41]

    Journal of Petroleum Geology , volume =

    Motaei, Eghbal and Ganat, Tarek and Tabatabai, Mehdi and Umer, Mian and Krishna, Shanker , title =. Journal of Petroleum Geology , volume =. 2026 , doi =

    2026

  42. [42]

    Journal of Petroleum Science and Engineering , volume =

    Ma, Xiaopeng and Zhang, Kai and Zhang, Jinding and Wang, Yanzhong and Zhang, Lei and Liu, Piyang and Yang, Yuyang and Wang, Jianchun , title =. Journal of Petroleum Science and Engineering , volume =. 2022 , doi =

    2022

  43. [43]

    Petroleum Science , volume =

    Cui, Jia-Wei and Sun, Wen-Yue and Jeong, Hoonyoung and Liu, Jun-Rong and Zhou, Wen-Xin , title =. Petroleum Science , volume =. 2025 , doi =

    2025