Neural Field Ensembles for Aerodynamic Surface Prediction: Winning Solution to the ONERA CRM Wall Distribution 2025 Challenge
Lionel Salesses, Caroline Sainvitu, Tariq Benamara
cs.LG · physics.flu-dyn
Quick verdict
Winning ONERA CRM 2025 entry: a conditional neural-field ensemble (8.81 vs 8.64 baseline); solid stepwise ablation, but selection was leaderboard-driven.
Summary
This paper documents the first-place solution to the ONERA CRM Wall Distribution Regression Challenge: predicting pressure and skin-friction coefficients (Cp, Cfx, Cfy, Cfz) over the NASA CRM wing-body-pylon-nacelle geometry from coordinates, surface normals, and operating conditions (Mach, angle of attack, stagnation pressure), trained on 312 RANS simulations of 260,774 surface points each, without mesh connectivity. The method is a conditional neural field: an MLP with Fourier feature encoding (Ne=128), five hidden layers of 700 neurons (~2.15M parameters), a Relative Squared Error loss aligned with the challenge's worst-case relative MAE metric, the SOAP optimizer, and a five-fold cross-validation ensemble of 20 models (~7.3×10⁷ parameters). On the hidden 146-simulation test set, the score rises stepwise from 7.80 (prototype) to 8.81, beating the organizer's Global MLP baseline (8.64, ~1.7×10¹⁰ parameters); final R² is 0.973 vs 0.956 and wrMAE 0.210 vs 0.228, though wrMAE for Cfx and Cfy is not better than the baseline. Limitations: the stepwise ablation is reported on leaderboard (hidden-test) scores with inconsistent validation/hidden-test labeling; no seed variance is reported; no external baselines beyond the organizers' models and no code release; and ensemble parameter counts are internally inconsistent (20 × 2.15M ≠ 7.3×10⁷).
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 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
- A conditional neural field ensemble (FFE-MLP with RSE loss, 5-fold CV, 20 models, ~7.3×10⁷ parameters) scored 8.8141 on the hidden test set, winning the challenge ahead of the Global MLP baseline at 8.6401.
- Fourier feature encoding was the single largest improvement (+0.54 score, 7.80→8.34), attributed to mitigating spectral bias for shocks and localized flow structures.
- Replacing MSE with a Relative Squared Error loss aligned to the worst-case metric cut wrMAE from 0.276 to 0.251 at essentially unchanged R² (Section 6.4).
- The winning solution is orders of magnitude more parameter-efficient than the baseline while also improving global accuracy (R² 0.973 vs 0.956).
- Mesh-free coordinate-based neural fields are a strong default when connectivity is unavailable and training simulations are limited (312 configurations).
Strengths
Metric-aligned loss is a genuine, well-isolated finding
Section 6.4 shows MSE→RSE loss leaves R² essentially unchanged (0.965→0.967) but cuts wrMAE from 0.276 to 0.251 (+0.13 score), directly demonstrating that matching the training objective to the worst-case relative error metric matters more than capacity at that stage.
Honest documentation of discarded approaches
Section 6.8 reports three failed directions (GNNs on reconstructed nearest-neighbor graphs, AoA-based zonal decomposition with threshold rationale from drag-coefficient convergence, learnable frequency encodings) with concrete reasons, which is rare in competition papers and useful for practitioners.
Quantified stepwise progression
Strong parameter efficiency with transparent comparison
Weaknesses
Ablation substrate is the hidden test set, not internal validation
Overclaim contradicted by own Table 3
Internal arithmetic inconsistency in ensemble size
Parameter-efficiency claim ambiguous in the abstract
No uncertainty quantification and no code
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a competition-solution paper: the authors (Cenaero) document the method ranked first (8.8141, Table 1) among 64 participants and 75 submissions in the ONERA CRM Wall Distribution Regression Challenge. The task is fixed-geometry surrogate modeling: predict Cp, Cfx, Cfy, Cfz at 260,774 surface points of the NASA CRM-WBPN configuration from coordinates, surface normals and three operating parameters (M∞, AoA, pi), trained on 312 RANS simulations with 146 held out as a hidden test set. Central claims: (i) a conditional neural field ensemble beats the organizer's Global MLP baseline (8.81 vs 8.64, Table 3) at far fewer parameters; (ii) Fourier feature encoding was the largest single gain (7.80→8.34, Table 2); (iii) aligning the training loss with the worst-case metric (RSE, Section 5.4) is decisive for robustness. Claims are appropriately scoped to this benchmark; multi-geometry generalization is explicitly deferred to future work (Section 8).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 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 score did the winning solution achieve in the ONERA CRM wall distribution challenge?
The Cenaero team's neural-field ensemble scored 8.8141 on the hidden test set, ranking first out of 64 participants and 75 submissions. The best organizer-provided baseline (a Global MLP) scored 8.6401; the maximum possible score is 10.
What is the CRM-WBPN dataset used in this challenge?
It is a database of 468 RANS simulations of the NASA Common Research Model wing-body-pylon-nacelle configuration, each providing pressure and skin-friction coefficients at 260,774 surface points across a range of Mach numbers (0.3–0.96), angles of attack (−15° to 15°), and stagnation pressures. 312 simulations were released for training and 146 kept as a hidden test set. After the challenge it was publicly released on recherche.data.gouv.fr (doi:10.57745/K1UK6Z).
How large is the winning model compared to the baseline?
Per the paper, the winning ensemble comprises 20 models of ~2.15 million parameters each (stated total ~7.3×10⁷, though 20 × 2.15M arithmetically gives 4.3×10⁷ — an inconsistency this review flags). The baseline Global MLP uses four networks of ~4×10⁹ parameters each, ~1.7×10¹⁰ in total. So the winner is roughly 233× smaller ensemble-vs-ensemble, or ~1860× smaller per network.
Is code available for this winning solution?
No. The paper states the CRM-WBPN dataset was publicly released after the challenge, but no code repository or implementation release is mentioned. The architecture (FFE-MLP, 5×700 layers) and training recipe (RSE loss, SOAP, 5-fold CV ensemble) are described in enough detail for reimplementation, though some details (epochs, ε in the loss) are missing.
Why did the authors use neural fields instead of GNNs or neural operators?
The challenge does not provide the CFD mesh, only point coordinates and surface normals, so mesh-connectivity-based methods (GNNs like MeshGraphNets) cannot be applied directly. The authors tried reconstructing approximate graphs from the point cloud (Section 6.8) but found the results too rough and the effort too large. Neural operators were cited as related work but not empirically compared, which is a limitation since the dataset is now public.
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.
- Figures 1–5 (geometry, train/test split maps, qualitative Cp/friction field comparisons) cannot be assessed from the extracted text; claims about improved shock reproduction rest on imagery not verifiable here.
- Table values were parsed from text extraction; small digit-level OCR errors cannot be fully excluded (e.g., 8.8141 in Table 1).
- Leaderboard scores (Table 1) are as reported on Codabench and were not independently verified.
- The exact composition of the final ensemble is ambiguous in the source text itself (parameter counts do not reconcile), which limits precision on that point rather than the extraction.
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.