ONE CYLinder: A Benchmark for Graph-Based Surrogate Modeling of Unsteady Bluff-Body Flows
Théodore Michel, Antoine Campos, Alban Dujardin, Henry Areiza, Philippe Meliga, Elie Hachem
physics.flu-dyn · cs.LG
Quick verdict
Claims a 450-sim, 3-regime cylinder-flow benchmark with Graph Transformer baseline; solid and honest, but only self-compared and data release unverified.
Summary
The paper introduces ONECYL, a benchmark for graph-based surrogate modeling of two-dimensional unsteady flow past a circular cylinder. It comprises 450 Variational Multiscale finite-element simulations (270,000 snapshots) across three Reynolds-number regimes (roughly Re 60–200, 600–2,000, and 6,000–20,000), each with 150 randomized cylinder diameters (0.6–2.0) and positions on ~1,750-node body-fitted meshes, providing velocity, pressure, level-set, node-type, Reynolds, and drag/lift data. A unified evaluation protocol combines rollout RMSE at 1, 50, and 600 autoregressive steps, nine virtual probes, and integrated drag/lift forces. The authors also present a sparse-attention Graph Transformer baseline (~3.3M parameters) trained with noise injection and an initializer–iterator split so rollouts start from prescribed conditions alone. Ablations show level-set geometry inputs consistently help (e.g., Re2 600-step error 51.87 vs 73.40 ×10⁻³, about 29%), while a divergence loss mainly benefits higher regimes (Re3: 132.63 vs 153.61 ×10⁻³); combined-pipeline errors match ground-truth-initialized rollouts. Limitations include only 15 test cases per regime with large variance, no external baseline (e.g., MeshGraphNets) run on this data, two-dimensionality, coarse meshes at the highest Reynolds numbers, phase/amplitude errors, and no visible data or code availability statement.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- ONECYL provides 450 high-fidelity VMS finite-element simulations (270,000 snapshots) spanning laminar, transitional, and high-Reynolds regimes with randomized cylinder geometries and a unified multi-level evaluation protocol.
- Explicit level-set geometry encoding consistently improves long-horizon prediction accuracy and generalization to unseen geometries across all three Reynolds regimes.
- Divergence-based regularization becomes increasingly beneficial as flow complexity (Reynolds number) increases.
- The initializer–iterator strategy yields stable 600-step end-to-end rollouts from prescribed initial and boundary conditions alone, with long-horizon errors matching ground-truth-initialized rollouts.
- One-step errors are insufficient for evaluating autoregressive flow surrogates; full rollouts reveal large differences concealed at one step.
Strengths
Multi-level, physically grounded evaluation protocol
Sec. 5.1 defines rollout RMSE at 1/50/600 steps, nine geometry-relative virtual probes, and drag/lift integration (Appendix B). Table 7 shows one-step errors of 3.12–3.45×10⁻³ across all variants while 600-step errors span 51.87–187.58×10⁻³ — a concrete demonstration that one-step accuracy cannot rank autoregressive surrogates.
Unusually complete benchmark specification
Tables 1–3 fix mesh parameters (hmin=D/40, ~1750±170 nodes), simulation parameters (T=120s, Δt_simu=0.01s, 601 stored snapshots), and per-regime statistics; Appendices A.9–A.11 enumerate every case ID with D and (x0,y0); identical 120/15/15 splits are reused across all three Reynolds regimes for direct comparison.
Honest, specific limitations section
Actionable ablation findings
Initializer–iterator decoupling works
Weaknesses
No external baseline evaluated on ONECYL
Thin statistics: n=15 test cases, no significance testing
Reference-data fidelity at high Re is asserted, not demonstrated
Internal inconsistencies impede reimplementation
Two of three physics losses never evaluated in the visible results
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Preliminary Review — ONECYL (arXiv 2609.08947)
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the ONECYL benchmark?
ONECYL is a benchmark of 450 two-dimensional simulations of unsteady flow past a circular cylinder, generated with a variational-multiscale finite-element solver across three Reynolds-number ranges (about 60–200, 600–2,000, and 6,000–20,000). Cylinder diameters (0.6–2.0) and positions are randomized per case. It provides time-resolved velocity and pressure fields on unstructured meshes (~1,750 nodes), mesh connectivity, level-set geometry descriptors, and drag/lift time series, intended for training and evaluating graph-neural-network surrogate models.
How large is the ONECYL dataset?
450 simulations with 601 stored snapshots each, totaling 270,000 snapshots. Each simulation covers 120 seconds of physical time with a 0.2 s dataset sampling interval. The three Reynolds regimes share the same 150 geometries and the same train/validation/test splits (120/15/15 cases), enabling direct cross-regime comparison.
What are the main findings of the ONECYL paper?
Three findings stand out: (1) feeding a level-set geometry descriptor to the model consistently improves long-horizon accuracy in all three regimes (e.g., Re2 600-step error 51.87 vs 73.40 ×10⁻³, about 29% better); (2) a divergence-based physics loss mainly helps at higher Reynolds numbers (Re3: 132.63 vs 153.61 ×10⁻³) but not the laminar regime; (3) a two-model initializer–iterator pipeline predicts full 600-step trajectories from boundary conditions alone, with long-horizon errors essentially matching rollouts started from CFD-computed initial states.
Does ONECYL include 3D or fully turbulent flow?
No. The benchmark is strictly two-dimensional. The Re4 subset (Re up to 20,000) exhibits irregular wake dynamics treated through the VMS formulation acting as implicit LES, but three-dimensional wake instabilities are not represented. The authors list this among the limitations, noting that extensions to higher Reynolds numbers would require finer meshes and potentially 3D data.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The excerpt contains three '[...sampling gap...]' breaks; parts of §3.1.2 (case variability), §5.1 (metric definitions), the §5.4 force analysis, and mid-list references were not shown, so judgments on those specifics rest on surrounding text only.
- No data/code availability statement appears anywhere in the provided text; a link may exist in an unsampled section, so codeAvailable=false reflects the excerpt rather than a verified absence in the full paper.
- Figure contents (curves in Figs. 12–25, C.26–C.44) cannot be assessed from extracted text; judgments about visualizations rely on captions and body text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.