Pith. sign in

REVIEW 3 major objections 5 minor 34 references

IteraSim RAG: A Multi-Stage Retrieval-Augmented Agentic Back-End for OpenFOAM-Based Computational Fluid Dynamics

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A multi-stage retrieval pipeline with query expansion, rank fusion, diversity re-ranking, and a canonical-knowledge layer achieves 77.9% mean tag coverage for OpenFOAM case generation, with parameter-modification tasks above 90%.

desk verdict Solid RAG-for-CFD system with a released benchmark, but the Table 3 ablation contradicts its own Eq. (3), so the main attribution claim is unsupported. read the letter →

arxiv 2607.20346 v1 pith:UMVNK2KG submitted 2026-07-22 cs.CE physics.flu-dyn

Pratyush Kumar This is my paper
classification cs.CEphysics.flu-dyn
keywords retrieval-augmentedgenerationOpenFOAMcomputationalfluiddynamicsmulti-agentsystemsqueryexpansionreciprocalrankfusionmaximalmarginalrelevancecanonicalknowledgelayer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that the quality of retrieval-augmented generation for OpenFOAM case configuration depends on retrieval architecture more than on model choice. A flat single-query retrieval loses about 50 points of tag coverage compared with a pipeline that expands the query into physics, solver-keyword, and troubleshooting variants, fuses the ranked lists, and re-ranks for diversity. A deterministic router that separates tool-specific workflow queries from corpus-wide physics queries, and a static canonical-knowledge layer injected into every agent prompt, are the other two load-bearing components; the canonical layer is worth a comparable 50 points. The evidence is a 28-case benchmark spanning zero-shot, few-shot, parameter-modification, and turbulence-swap setups, on which the full pipeline scores 77.9% mean tag coverage, with the parameter-modification category exceeding 90%. The paper reports only the retrieval tier as its principal metric; end-to-end executability of agent-generated cases is deferred to a companion note, though six reference configurations run to completion and two corrupted cases are repaired inside the Reviewer loop.

What carries the argument

The central object is a three-stage retrieval pipeline that transforms a single user query into a diverse, deduplicated context block: an LLM rewrites the query into physics, solver-keyword, and troubleshooting variants; Reciprocal Rank Fusion (a rank-aggregation rule that adds inverse ranks across lists) merges the four embedded queries' candidate lists; and Maximal Marginal Relevance (a greedy relevance–diversity trade-off) re-ranks the fused set and selects the final context. Around this core sit two further carry mechanisms: a keyword-driven dual-mode router that locks workflow queries onto a solver-specific sub-corpus while sending physics-level queries to the full corpus, and a static

What would settle it

Run the full Architect–InputWriter–Reviewer orchestrator on the same 28-case benchmark and compare end-to-end executability (rubric score 3 or 4) against the retrieval-tier scores. If a sizable fraction of cases with retrieval coverage above 70% fail to produce a running case (e.g., more than half), then Eq. (3) is not a valid proxy for practical usefulness. Alternatively, a surgical test of the canonical-layer attribution: on the 30-query ablation set, replace the canonical-knowledge layer with an equal-length block of random text from the same corpus while keeping retrieval unchanged; if cov

Watch

Extended reading notes

Core claim

The paper's central discovery is that three persistent failure modes of earlier LLM-for-CFD systems—flat single-query retrieval, undifferentiated routing of operationally distinct requests, and entangling the drafter with the reviewer while lacking an authoritative reference—can each be addressed by a distinct architectural component, and that these components can be credited with specific, large improvements in retrieval coverage. When any one of the two main components is removed, coverage collapses by roughly half: disabling query expansion drops the ablation score from 75.6% to 25.9%, and disabling the canonical-knowledge pinning drops it to 25.5%. The pipeline's mean tag coverage on the

Load-bearing premise

The load-bearing premise is that the paper's exact-token-match scoring rule (Eq. 3) on the 28-case benchmark measures something that predicts whether the system can actually generate executable, physically correct OpenFOAM cases—so that a 77.9% tag coverage in the retrieved context block translates into useful case generation, rather than merely retrieving text that contains the right keywords.

Editorial extensions

If this is right

  • A flat single-query RAG baseline is at a roughly 50-point coverage disadvantage for OpenFOAM case configuration, because disabling query expansion (which reduces the pipeline to flat retrieval) drops the ablation score from 75.6% to 25.9%.
  • The canonical-knowledge layer is equally load-bearing: removing it costs 50.1 points, so retrieval of documentation alone is not enough to stabilise generation of solver dictionaries and boundary conditions.
  • Tool-conditioned routing is what drives the 90.7% mean and 100% median on parameter-modification tasks: when a query names its active solver, the router locks retrieval to that solver's sub-corpus and recovers the relevant dictionary tags.
  • The median retrieval latency of 2.09 s shows that four-way query expansion does not push the system out of the interactive regime, so multi-query retrieval is practical in an agentic loop.
  • The low variance (σ=8.8%) and 60% floor across turbulence-model swaps indicate that rank fusion plus diversity re-ranking keeps the wall-function and field-equation tags present even when their similarity profiles differ.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: the two ~50-point ablation drops imply that query rewriting and authoritative-reference injection are nearly interchangeable in their contribution, which suggests a transferable design rule for other scientific software domains: pair query expansion with a small canonical reference layer rather than relying on corpus size alone.
  • Editorial extension: the paper's own synonym-aware re-scoring raises one low outlier from 33% to ~67%, which hints that the true operational coverage of the pipeline is higher than the reported 77.9% exact-match score; a fuzzy-matching evaluation would likely raise the headline figure.
  • Editorial extension: since the Reviewer repaired the two probe failures using only the solver log and the canonical layer, without consulting retrieval, a plausible testable hypothesis is that the canonical layer alone may be sufficient for a bounded class of syntax-repair tasks; if confirmed, the retrieval tier could be bypassed for such corrections, saving latency.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper presents IteraSim RAG, a retrieval-augmented agentic back-end for OpenFOAM case generation. The system combines an LLM-based three-stage retrieval pipeline (query expansion, reciprocal rank fusion, MMR re-ranking), a deterministic dual-mode router, and an Architect–InputWriter–Reviewer agent triad supplemented by a static canonical-knowledge layer. The main measured result is a mean retrieval coverage of 77.9% (median 79.1%) on a released 28-case benchmark, with a leave-one-out ablation attributing the largest contributions to multi-query expansion (−49.7 points) and canonical-layer pinning (−50.1 points). End-to-end executability of the agent-generated cases is deferred to a companion note; the paper reports that all six underlying reference configurations run to completion on OpenFOAM v2506 and that two injected FATAL errors are repaired by the Reviewer loop.

Significance. If the claims were fully supported, this would be a useful contribution to the LLM-for-CFD literature: it explicitly targets three identified limitations of flat RAG systems, and it ships a reproducible 28-case benchmark, scoring rubric, ablation records, corrupted-case probes, and figure scripts. The authors are also transparent about the commercial competing interest and about the fact that the benchmark is small and tag-based. The main reservation is that the central component-attribution claim is internally inconsistent as written, and the headline retrieval metric has not yet been tied to the downstream executability that motivates the system. These issues are fixable but require substantive revision.

major comments (3)
  1. [Table 3; Secs. 2.5, 3.2, 3.5; Eq. (3)] The −50.1 point ablation for disabling the canonical-knowledge layer is internally inconsistent with the paper's own definition of the retrieval metric. Eq. (3) scores tokens(X_i), where X_i is 'the context block returned by the three-stage retrieval pipeline, before any generation occurs.' Sec. 2.5 states that the canonical layer K is 'appended to the system prompt of every agent in addition to the retrieved context,' and Sec. 3.5 states K is 'by design invisible at the retrieval tier, since it is injected at the orchestration stage rather than at retrieval.' If K is not part of X_i, removing it cannot affect s_i, so Table 3's −50.1 row cannot measure any retrieval-tier quantity. If K is part of X_i, then the 77.9% headline silently includes a non-retrieved, proprietary component and is no longer a retrieval-coverage number. Either way, the claim in Sec. 3.2 that 'canonical pinning' is
  2. [Sec. 3.1; Eq. (3); Secs. 3.3 and 3.8] The headline 77.9% is an exact-token-match score against author-defined tags on an author-built benchmark, with no evidence that tag coverage predicts executable or physically correct OpenFOAM cases. Sec. 3.3 explicitly defers agent-driven executability to a companion technical note, and Table 5 compares self-selected benchmarks with non-comparable metrics (retrieval vs. end-to-end success). As a result, the paper's stated practical motivation—lowering the CFD setup barrier—rests on an unvalidated proxy. A concrete remedy would be to report agent-driven executability on the same 28 cases, or to show a correlation between s_i and downstream success on a held-out set. In addition, the manual synonym-aware re-scoring of A3 after seeing the results (raising 33.3% to ≈66.7%) should be pre-registered or applied uniformly to all cases; otherwise it is a post-hoc scoring change.
  3. [Table 3; Sec. 2.3; Sec. 2.4] The leave-one-out ablation in Table 3 is scored on a separate 30-query set, and the text says the full-pipeline score 'lies within one standard error' of the 28-case mean but no standard error is reported. More importantly, the retrieval pipeline contains several tunable hyperparameters whose sensitivity is not assessed: the MMR lambda (0.6/0.5), the RRF smoothing constant k=60, the retrieval depth K/K_fetch, the chunk size/overlap, and the query-expansion temperature and token budget. Because the headline metric depends on these choices, the paper should either report a small sensitivity sweep or state explicitly which values were used and why. This is less severe than the canonical-layer issue but still affects the robustness of the central quantitative claim.
minor comments (5)
  1. [Table B.1] The column headings appear garbled: 'Solvers(%) τ(%)' is not defined and is likely a formatting artifact; 'A vg.' should be 'Avg.' The table also mixes solver names in different cases (e.g., 'icofoam' vs. 'simpleFoam'), which should be normalized.
  2. [Sec. 3.3] The text says 'all 28 base configurations meshed cleanly' immediately after stating that the 28 cases reduce to six distinct base runs. Please clarify whether the executability check was run on six distinct configurations inherited by the rest, or on all 28 individually.
  3. [Sec. 3.5] This section twice states the canonical layer is invisible at the retrieval tier but then attributes Category D's tight standard deviation to the wall-function/y+ table in K. Given the major comment above, the attribution language should be revised to avoid implying a retrieval-tier effect.
  4. [Sec. 3.1 / Sec. 3.8] The paper lists 'three residual caveats' after earlier stating 'two limitations affect the numbers above' (Sec. 3.1). The numbering should be made consistent and all caveats collected in one place.
  5. [Sec. 2.2 / Ref. [29]] The expert markdown notes are proprietary and only their section-level structure is released. This limits reproducibility of the retrieval corpus; please state explicitly which parts of the corpus are public and which are not, and whether the benchmark can be run without access to the proprietary notes.

Circularity Check

1 steps flagged · score 6.0 of 10

Table 3's −50.1 canonical-layer ablation is unmeasurable under Eq. (3): K is defined outside X_i, so the central attribution is by construction.

  1. self definitional [Table 3 in Sec. 3.2, contrasted with Eq. (3) in Sec. 3.1, Sec. 2.5, and Sec. 3.5]
    "Sec. 3.1: 'The retrieval score for case i is s_i = 0.80·|T_i ∩ tokens(X_i)|/|T_i| + 0.20·⊮[solver_i ∈ tokens(X_i)] … where X_i is the context block returned by the three-stage pipeline.' Sec. 2.5: 'a static canonical-knowledge layer that is appended to the system prompt of every agent in addition to the retrieved context.' Sec. 3.2: 'Disabling the canonical-knowledge pinning is even more damaging (−50.1 points), which quantifies the contribution of the static knowledge layer of Sec. 2.5.' Sec. 3.5: 'The canonical-knowledge layer (contribution L3) is by design invisible at the retrieval tier.'"

    Eq. (3) scores only tokens(X_i), the context block returned by the three-stage retrieval pipeline. Sec. 2.5 places K outside X_i ('appended to the system prompt of every agent in addition to the retrieved context'), and Sec. 3.5 says K is 'by design invisible at the retrieval tier.' Hence bypassing K in the –canonical row cannot change s_i; the −50.1 loss is unobservable in the stated metric. If the ablation instead removes K from the scored context, then the 'retrieval-tier' metric silently includes a non-retrieved, author-written reference, so K's contribution is baked into the definition of X_i. Either way, the claim that canonical pinning is one of 'the two load-bearing components' is forced by the paper's own definitions rather than measured.

full rationale

The headline 77.9% retrieval coverage is itself an honest, self-contained measurement against a released rubric, and the multi-query/MMR/RRF pipeline effects are plausible retrieval-side quantities. The circular step is the component-attribution claim for the canonical-knowledge layer. Eq. (3) defines the retrieval score on tokens(X_i), the context block returned by the three-stage pipeline. Section 2.5 defines K as injected at the orchestration stage, in addition to the retrieved context, and Sec. 3.5 explicitly concedes K is 'by design invisible at the retrieval tier.' Therefore Table 3's −50.1 point loss from disabling canonical pinning cannot be produced by the stated metric—the intervention is outside the measured object. If the ablation instead removed K from the scored context, the metric would no longer be a pure retrieval-tier measure and K's contribution would be defined into the score by construction. Because this row is the paper's central quantitative basis for calling canonical pinning one of the two load-bearing components, that specific claim reduces to a definitional artifact. No self-citation chain is load-bearing here; the problem is an internal definitional contradiction in the ablation logic. The score reflects partial circularity: the primary retrieval-coverage number stands, but a central component-level attribution is forced by definitions rather than evidence.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new physical, mathematical, or ontological entities are introduced. The canonical-knowledge layer and tool-dependency graph are software components, not postulated entities. The main load-bearing inputs are the author-defined benchmark tags, the proprietary canonical layer, and the chosen hyperparameters.

free parameters (5)
  • MMR relevance-diversity tradeoff lambda = 0.6 (workflow), 0.5 (expert)
    Hand-picked in Sec. 2.3.3; no sensitivity analysis; directly controls the diversity of the final context block and appears in the ablation as a −10.6 point loss when MMR is removed.
  • RRF smoothing constant k = 60
    Taken from Cormack et al. (Sec. 2.3.2); affects fusion weights and candidate ordering; no robustness test reported.
  • Retrieval depth K and K_fetch = K=5 workflow, K=8 expert; K_fetch=max(2K, K+4)
    Chosen per mode in Sec. 2.3.2; determines context-block size and candidate pool, directly affecting measured tag coverage.
  • Chunk size and overlap = 400 words (docs), 800 words (long notes), 100-word overlap
    Ingestion choices in Sec. 2.2; affect whether expected tags fall inside a retrieved chunk.
  • Query-expansion LLM temperature and token budget = 0.4, 120 tokens
    Stage-1 settings in Sec. 2.3.1; query expansion is the largest single contributor (−49.7 points when removed), yet these settings are not ablated.
assumptions (5)
  • standard math RRF and MMR formulas (Eqs. 1–2) operate as described in the cited information-retrieval literature
    The paper relies on these formulas without independent derivation in Sec. 2.3.
  • domain assumption The 28-case benchmark and its expected physics/solver/keyword tags T_i are a valid ground truth for OpenFOAM case-generation quality
    Tags and categories are author-defined (Sec. 3.1, Appendix B); no external validator or inter-rater check is used, and manual synonym-aware re-scoring was needed for two outliers.
  • domain assumption The proprietary canonical-knowledge layer is authoritative and correct
    The layer is injected into every agent prompt (Sec. 2.5) and is credited with the largest ablation loss (−50.1); its full text is not released, so its content cannot be audited.
  • domain assumption Tag coverage in the retrieved context predicts successful OpenFOAM case generation
    The paper measures only the retrieval tier (Sec. 3.1) and explicitly defers agent-driven executability to a companion note (Sec. 3.3); the link between retrieval coverage and runnable cases is asserted, not measured.
  • domain assumption The corpus (tutorials, expert notes, FAQ pairs, FoamGPT pairs, WEC corpus) is representative of the LLM-for-CFD request space
    Only five sources are ingested (Sec. 2.2); two of the 28 cases fail from corpus gaps, which the paper attributes to coverage rather than algorithm.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IteraSim RAG: A Multi-Stage Retrieval-Augmented Agentic Back-End for OpenFOAM-Based Computational Fluid Dynamics." pith.science (2026). https://pith.science/paper/UMVNK2KG

@misc{pith2026260720346,
  author       = {Pith},
  title        = {Pith review of: IteraSim RAG: A Multi-Stage Retrieval-Augmented Agentic Back-End for OpenFOAM-Based Computational Fluid Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UMVNK2KG}},
  note         = {Machine review of arXiv:2607.20346}
}
read the original abstract

Configuring a computational fluid dynamics (CFD) case in OpenFOAM requires assembling a multi-directory input deck of mutually consistent solver, discretisation and boundary-condition dictionaries -- a task that remains a substantial barrier to non-specialist use of open-source CFD software. Large language models (LLMs) coupled with retrieval-augmented generation (RAG) can lower this barrier, but existing systems retrieve with a single flat query, apply one retrieval strategy to operationally distinct requests, and let a single agent both draft and review its own output. We present IteraSim RAG, a retrieval-augmented software back-end for automated OpenFOAM case generation built around these three limitations. An LLM first expands the query into physics, solver-keyword and troubleshooting variants, Reciprocal Rank Fusion then merges the resulting ranked lists, and Maximal Marginal Relevance re-ranks the fused candidates against an HNSW-indexed dense vector store. A deterministic keyword router dispatches tool-conditioned workflow queries and corpus-wide physics queries down separate retrieval paths, and generation is split across an Architect, an InputWriter and a Reviewer agent, backed by a static canonical-knowledge layer covering solver selection, turbulence closures, boundary conditions and finite-volume defaults. On an openly released 28-case benchmark spanning zero-shot setup, few-shot generalisation, single-parameter modifications and turbulence-model swaps, the pipeline attains a mean retrieval coverage of 77.9% (median 79.1%), with the parameter-modification category exceeding 90%. All six reference configurations run to completion on OpenFOAM v2506, and two synthetically corrupted cases are diagnosed and repaired within the bounded Reviewer loop using only the solver log and the canonical layer. The benchmark, scoring rubric and figure scripts are released for reproducibility.

Figures

Figures reproduced from arXiv: 2607.20346 by the authors.

Figure 1
Figure 1. The figure can be read top-to-bottom as a single pass: a user query [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 1
Figure 1. Architecture of IteraSim RAG (each block is walked through, top to bottom, in [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Schematic representation of the three-stage retrieval pipeline of IteraSim RAG. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figures from the paper (5 more)
Figure 3
Figure 3. Figure 3: Retrieval-tier performance on the 28-case IteraSim RAG benchmark. [PITH_FULL_IMAGE:figures/full_fig_p016_3.png]
Figure 4
Figure 4. Figure 4: Retrieval coverage per case within each category. [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: End-to-end IteraSim RAG workflow: prompt [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Final-time velocity magnitude at t = 0.5 s for the lid-driven cavity: unmodified OpenFOAM tutorial reference case (left) and the case generated by IteraSim RAG from the natural-language prompt (right). FATAL; identify and correct the fault such that the case runs to en…
Figure 7
Figure 7. Figure 7: Converged simulation fields for the six zero-shot Category A reference cases. [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 5 canonical work pages

  1. [1]

    O. S. Ismail, G. T. Adewoye, Analyses and modeling of laminar flow in pipes using numerical approach, Journal of Software Engineering and Applications 5 (9) (2012) 653–658.doi:10.4236/jsea.2012.59076

    arXiv 2012

  2. [2]

    Alemi, S

    H. Alemi, S. A. Nourbakhsh, M. Raisee, A. F. Najafi, Effects of volute curvature on performance of a low specific-speed centrifugal pump at design and off-design conditions, Journal of Turbomachinery 137 (4) (2015) 041009.doi:10.1115/1.4028766

    work page doi:10.1115/1.4028766 2015

  3. [3]

    Safaei, R

    S. Safaei, R. Zeynali, B. Safaei, Modeling study of propane/propylene membrane separation using CFD method, Macromolecular Theory and Simulations 30 (3) (2021) 2000092.doi:10.1002/mats.202000092. URLhttps://doi.org/10.1002/mats.202000092

    work page doi:10.1002/mats.202000092 2021

  4. [4]

    Sochi, Fluid flow at branching junctions, International Journal of Fluid Mechanics Research 42 (1) (2015) 59–81.doi:10.1615/ InterJFluidMechRes.v42.i1.50

    T. Sochi, Fluid flow at branching junctions, International Journal of Fluid Mechanics Research 42 (1) (2015) 59–81.doi:10.1615/ InterJFluidMechRes.v42.i1.50

    2015

  5. [5]

    Kumar, S

    P. Kumar, S. K. Saha, A. Sharma, Experimental and CFD-DEM study on local packing distribution and thermofluidic analysis of binary packed bed, Chemical Engineering Science 282 (2023) 119372.doi:10.1016/ j.ces.2023.119372. URLhttps://doi.org/10.1016/j.ces.2023.119372

    arXiv 2023

  6. [6]

    Kumar, S

    P. Kumar, S. K. Saha, A. Sharma, D. Sharma, P. Chaudhuri, CFD investigation of helium gas flow in sphere packed (Pebble bed) in a rectangular canister using OpenFOAM, Fusion Engineering and Design 172 (2021) 112858.doi:10.1016/j.fusengdes.2021.112858. URLhttps://doi.org/10.1016/j.fusengdes.2021.112858

    arXiv 2021

  7. [7]

    Jasak, A

    H. Jasak, A. Jemcov, Z. Tukovic, OpenFOAM: A C++ library for com- plex physics simulations, in: International Workshop on Coupled Meth- ods in Numerical Dynamics, 2007, pp. 1–20

    2007

  8. [8]

    H. G. Weller, G. Tabor, H. Jasak, C. Fureby, A tensorial approach to computational continuum mechanics using object-oriented techniques, Computers in Physics 12 (6) (1998) 620–631.doi:10.1063/1.168744. URLhttps://doi.org/10.1063/1.168744 33

    doi:10.1063/1.168744 1998

Show all 34 references
  1. [9]

    S. K. Samal, P. Kumar, S. K. Saha, Comparative study of surface rough- ness models in the hydro-thermal characterization of flow over rough walls, Physics of Fluids 35 (9) (2023) 095106.doi:10.1063/5.0165497. URLhttps://doi.org/10.1063/5.0165497

    2023 doi

  2. [10]

    Kumar, S

    P. Kumar, S. Sarkar, S. K. Saha, Numerical study of bouncing Lei- denfrost viscoplastic drops, Physics of Fluids 38 (3) (2026) 032007. doi:10.1063/5.0316065

    2026 doi

  3. [11]

    Pandey, R

    S. Pandey, R. Xu, W. Wang, X. Chu, OpenFOAMGPT: A retrieval- augmented large language model (LLM) agent for OpenFOAM-based computational fluid dynamics, Physics of Fluids 37 (3) (2025) 035120. doi:10.1063/5.0257555. URLhttps://doi.org/10.1063/5.0257555

    2025 doi

  4. [12]

    L. Yue, N. Somasekharan, T. Zhang, Y. Cao, Z. Chen, S. Di, S. Pan, Foam-Agent: Towards automated intelligent CFD workflows, arXiv:2505.04997 (2025). URLhttps://arxiv.org/abs/2505.04997

    2025

  5. [13]

    Duraisamy, G

    K. Duraisamy, G. Iaccarino, H. Xiao, Turbulence modeling in the age of data, Annual Review of Fluid Mechanics 51 (2019) 357–377.doi: 10.1146/annurev-fluid-010518-040547. URLhttps://doi.org/10.1146/annurev-fluid-010518-040547

    2019 doi

  6. [14]

    A. Beck, D. Flad, C.-D. Munz, Deep neural networks for data-driven LES closure models, Journal of Computational Physics 398 (2019) 108910.doi:10.1016/j.jcp.2019.108910. URLhttps://doi.org/10.1016/j.jcp.2019.108910

    2019

  7. [15]

    S. L. Brunton, B. R. Noack, P. Koumoutsakos, Machine learning for fluid mechanics, Annual Review of Fluid Mechanics 52 (2020) 477–508. doi:10.1146/annurev-fluid-010719-060214. URLhttps://doi.org/10.1146/annurev-fluid-010719-060214

    2020 doi

  8. [16]

    D. A. Boiko, R. MacKnight, B. Kline, G. Gomes, Autonomous chemical research with large language models, Nature 624 (2023) 570–578.doi: 10.1038/s41586-023-06792-0. URLhttps://doi.org/10.1038/s41586-023-06792-0 34

    2023 doi

  9. [17]

    L. Wang, C. Ma, X. Feng, Z. Zhang, H. Yang, J. Zhang, Z. Chen, J. Tang, X. Chen, Y. Lin, et al., A survey on large language model based autonomous agents, Frontiers of Computer Science 18 (6) (2024) 186345.doi:10.1007/s11704-024-40231-1. URLhttps://doi.org/10.1007/s11704-024-40231-1

    2024 doi

  10. [18]

    Y. Chen, X. Zhu, H. Zhou, Z. Ren, MetaOpenFOAM: an LLM-based multi-agent framework for CFD, arXiv:2407.21320 (2024). URLhttps://arxiv.org/abs/2407.21320

    2024 arXiv

  11. [19]

    L. Yue, N. Somasekharan, T. Zhang, Y. Cao, S. Pan, Foam-Agent 2.0: An end-to-end composable multi-agent framework for automating CFD simulation in OpenFOAM, arXiv:2509.18178 (2025). URLhttps://arxiv.org/abs/2509.18178

    2025

  12. [20]

    L. Yue, P. Xu, T. Zhang, N. Somasekharan, S. Pan, FoamGPT: Fine-tuning large language model for agentic automation of CFD simulations with OpenFOAM, in: Machine Learning and the Physical Sciences Workshop, NeurIPS, 2025. URLhttps://ml4physicalsciences.github.io/2025/files/ Neu...

    2025

  13. [21]

    E. Fan, K. Hu, Z. Wu, J. Ge, J. Miao, Y. Zhang, H. Sun, W. Wang, T. Zhang, ChatCFD: A large language model-driven agent for end-to- end computational fluid dynamics automation with structured knowl- edge and reasoning, Advanced Intelligent Discovery 2 (3) (2026) e202500174.doi...

    2026 doi

  14. [22]

    Z. Xu, L. Wang, C. Wang, Y. Chen, Q. Luo, H.-D. Yao, S. Wang, G. He, CFDAgent: A language-guided, zero-shot multi-agent system for complex flow simulation, Physics of Fluids 37 (11) (2025) 117124. doi:10.1063/5.0294696. URLhttps://doi.org/10.1063/5.0294696

    2025 doi

  15. [23]

    C. Yang, Y. Wang, J. Tang, H. Qu, Z. Zou, Y. Liu, C. Deng, Z. Qiu, M. Ding, SwarmFoam: An OpenFOAM multi-agent system based on multiple types of large language models, arXiv preprint arXiv:2601.07252 (2026). URLhttps://arxiv.org/abs/2601.07252 35

    2026

  16. [24]

    G. W. Furnas, T. K. Landauer, L. M. Gomez, S. T. Dumais, The vo- cabulary problem in human-system communication, Communications of the ACM 30 (11) (1987) 964–971.doi:10.1145/32206.32212. URLhttps://doi.org/10.1145/32206.32212

    1987

  17. [25]

    Madaan, N

    A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, et al., Self- refine: Iterative refinement with self-feedback, in: Advances in Neural Information Processing Systems, Vol. 36, 2023, pp. 46534–46594

    2023

  18. [26]

    G. V. Cormack, C. L. A. Clarke, S. Buettcher, Reciprocal rank fusion outperforms Condorcet and individual rank learning methods, in: Proc. 32nd ACM SIGIR, 2009, pp. 758–759.doi:10.1145/1571941.1572114. URLhttps://doi.org/10.1145/1571941.1572114

    2009

  19. [27]

    Carbonell, J

    J. Carbonell, J. Goldstein, The use of MMR, diversity-based reranking for reordering documents and producing summaries, in: Proc. 21st ACM SIGIR, 1998, pp. 335–336.doi:10.1145/290941.291025. URLhttps://doi.org/10.1145/290941.291025

    1998

  20. [28]

    Y. A. Malkov, D. A. Yashunin, Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs, IEEE Transactions on Pattern Analysis and Machine Intelligence 42 (4) (2020) 824–836.doi:10.1109/TPAMI.2018.2889473. URLhttps://doi.org/10.11...

    2020

  21. [29]

    I. Team, IteraSim RAG benchmark and evaluation kit: canonical- knowledge-layer structure,https://github.com/iterasim/ iterasim-rag-public, section-level structure of the curated ex- pert notes on solver selection, boundary-condition idioms, mesh- quality criteria, numerical-st...

    2026

  22. [30]

    cfd-online.com/Forums/openfoam/, practitioner threads used as source material for the troubleshooting question–answer set

    CFD Online, OpenFOAM technical discussion forum,https://www. cfd-online.com/Forums/openfoam/, practitioner threads used as source material for the troubleshooting question–answer set. Accessed 13 July 2026

    2026

  23. [31]

    OpenCFD Ltd., OpenFOAM v2506 tutorial and reference case suite, https://www.openfoam.com/documentation/tutorial-guide, refer- 36 ence dictionaries and canonical case distribution bundled with the ESI- OpenCFD OpenFOAM release (2025)

    2025

  24. [32]

    SandiaNationalLaboratories, WEC-OpenFOAMopendatasets,https: //github.com/sandialabs/WEC-OpenFOAM-Open-Datasets, wave en- ergy converter templates, meshes, and Reference Model 6 (RM6) cou- pled multiphase / rigid-body examples (2024)

    2024

  25. [33]

    J. J. Pan, J. Wang, G. Li, Survey of vector database management sys- tems, in: The VLDB Journal, Vol. 33, 2024, pp. 1591–1615.doi: 10.1007/s00778-024-00864-x. URLhttps://doi.org/10.1007/s00778-024-00864-x

    2024 doi

  26. [34]

    Kumar, IteraSim RAG: benchmark and evaluation kit for retrieval- augmented agentic OpenFOAM case generation, Zenodo, version v1.0- submission (2026).doi:10.5281/zenodo.21346132

    P. Kumar, IteraSim RAG: benchmark and evaluation kit for retrieval- augmented agentic OpenFOAM case generation, Zenodo, version v1.0- submission (2026).doi:10.5281/zenodo.21346132. Appendix A. Summary of the RAG Model For convenience we summarise the end-to-end behaviour of It...

    2026 doi

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.