acceptodds
Withdrawn from ICLR 2027

FlowGEM: a multi-scale neural surrogate model based on geometry and physics enrichment

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Neural networks learn the low-frequency components of a target signal before the high-frequency ones, a phenomenon known as spectral bias. This failure mode also affects neural surrogate models used to simulate systems governed by PDEs. The bias is particularly consequential in automotive and aerospace aerodynamics, where the flow is shaped both by millimeter-scale geometric details and by meter-scale structures such as the windshield angle. To address this challenge, CFD surrogates commonly employ hierarchical modelling, mixing reads that target local geometry features with reads that target large-scale information, and combining them into a learned latent representation with little control over how these scales interact. In this study, we present FlowGEM, a neural surrogate model trained on surface and volume point clouds. FlowGEM reads the geometry at multiple spatial scales in sequence, with shared weights across scales, recursively conditioning through a scale-dependent gate. In addition, it learns global conditioning through a physics enrichment module, injecting learned physics information into the model's latent geometry representation. With this setup, we compare FlowGEM against state-of-the-art surrogate models on realistic aerodynamics benchmarks (DrivAerML and NASA CRM), where it outperforms them on 4 of the 6 metrics tested. To understand the impact of spectral bias on the model, we analyze the prediction error in spectral bands and show how the multi-scale attention reduces frequency-dependent errors. Finally, because the output is predicted independently at each input point, FlowGEM's inference cost grows linearly with the number of points evaluated, and a full mesh can be decoded in chunks that fit the available memory.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, PDEs, CFDs, multi-scale attention, physics enrichment

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_jgjDOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes FlowGEM, a CFD neural surrogate with 2 core components: multi-scale attention (MSA) and physics enrichment with anchor decoder features. Evaluated on DrivAerML and NASA-CRM, FlowGEM achieves the best results on 4/6 reported metrics compared to recent baselines. Prediction error across graph-frequency bands are presented and analyzed to show which frequency bands FlowGEM can best resolve compared to a baseline.

Critical strengths

strengths:

  1. Strong empirical result on evaluated 2 datasets: overall winning on 4/6 reported metrics in DrivAerML and NASA-CRM, compared to state-of-the-art models

  2. Frequency-resolved error analysis provides a useful additional evaluation perspective. Even though the spectral analysis does not directly establish the classical training-dynamics notion of spectral bias that the paper claims to investigate, analyzing prediction error across graph-frequency bands is still valuable and relatively uncommon in related work.

Critical weaknesses

weaknesses:

  1. MSA novelty and design are weakly justified.

Proposed MSA is computed conceptually close to RNN, with the time dimension being replaced by scale, so it's not inherently novel architectural design. The paper does not include intuition or justifications for such design. Besides, scale-specific representation learning is delegated to FiLM.

A natural baseline to justify the choice is, how is this formulation compare to an alternative and intuitive approach where each scale uses stacked MSA layers? Plus in this case FiLM modulation is conceptually not needed because each layer already encodes scale information by learning transformation on its assigned scale. This should be included as a control for the recurrent MSA design.

  1. Impact of physics enrichment is not isolated. Since both anchor and queries are point samples, and both ultimately use decoder feature learned to predict CFD quantity, the query decoder already learns physics, and I find it a stretch to call the inclusion of decoder feature from the anchor part "physics-enriched". The only difference between anchor/query decoder feature is that the anchor feature is trained to be helpful to the Phys-Queries, but the paper does not ablate whether improvements come from auxiliary supervision, extra transformation depth to the anchor feature by the decoder, or the full proposed mechanism. A controlled physics-anchor ablation is therefore necessary.

  2. Missing hyperparameter sensitivity ablations: The paper does not study the impact of geometry-pool size and anchor count even though they are the hyperparam to the core technical proposals, and their choices directly affect fidelity, capacity, and cost.

  3. Missing more context on experiment background. For example what's the parameter size for all models, and their training/inference cost as in latency and FLOPs.

  4. Limited dataset coverage: as both AB-UPT and Transolver-3 includes AHMEDML in the evaluation, it would be highly justified for FlowGEM to include this dataset in the evaluation.

  5. Inconsistent result for claims of network failure in high-frequency learning: the paper claimed that high-frequency features are much more challenged to be learned compared to low-frequency, but Fig 3 shows that the largest benefit of FlowGEM over its (L=1) variant and ABUPT occurs in the mid-frequency range, with much smaller improvement at the highest frequencies.

Questions for authors

Please see each weakness for requests for additional evidence or writing clarification.

Reviewer_AkLaOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents FlowGEM, a neural surrogate for predicting aerodynamic fields on large, irregular geometries. Its multiscale attention module enriches anchors and queries through sequential neighborhood reads, with each scale's representation conditioning the next read. The main attention operator is shared across scales, while feature modulation incorporates neighborhood size. A physics-enrichment module uses auxiliary supervision for field predictions at anchor locations and feeds the prediction head's hidden features back into the anchors. Queries independently read shared anchor representations, allowing complete meshes to be decoded in chunks.

Experiments cover DrivAerML and NASA-CRM, with explicit operating parameters supplied for NASA-CRM. The paper reports improvements on both NASA-CRM targets and the DrivAerML pressure targets, and evaluates drag and lift coefficients. Graph-spectral analysis on five sampled test meshes shows that the four-scale variant improves selected surface-pressure bands relative to the single-scale variant, with little corresponding gain for the volume total-pressure coefficient. (Sections 3–5; Tables 2–3; Figures 2–3; Appendix A.2.)

Critical strengths
  • Good results on industrial aerodynamic benchmarks The paper evaluates FlowGEM on DrivAerML and NASA-CRM, including predictions over complete meshes. Table 3 reports approximately 25–28% lower errors on the two NASA-CRM targets than the strongest previously reported results listed in the table, as well as improvements on both DrivAerML pressure targets.

  • Evaluation beyond aggregate field errors The evaluation includes drag and lift coefficients, graph-spectral error analysis, and a local reproduction of AB-UPT on DrivAerML within the authors' training and evaluation pipeline. These complementary checks make the empirical results more informative.

Critical weaknesses
  • The benefit of recursive multiscale conditioning is not isolated. The four-scale versus single-scale comparison changes the number of reads, neighborhood coverage, and processing depth simultaneously. It therefore cannot show whether the improvement comes specifically from recursive conditioning, rather than from access to more geometric information or additional computation. Comparisons with parallel multiscale fusion and repeated single-scale reads, using comparable information and computational budgets, would test the main design claim more directly (Section 3.2; Figure 3).

  • Physics Enrichment lacks targeted validation, and its interpretation is too broad. PE combines auxiliary supervision, feedback of the prediction head's hidden features, and additional anchor processing. Without controls that separate these elements, the independent value of feature feedback remains unclear. Furthermore, Section 3.3 suggests that operating conditions need not be explicitly supplied, whereas the NASA-CRM experiments use six operating-condition inputs and DrivAerML has fixed conditions. The results therefore do not establish that PE can recover varying, unprovided operating conditions (Section 3.3; Appendix A.2 and B).

  • The spectral evidence supports a narrower conclusion than the motivation suggests. Figure 3 analyzes only five test geometries and two pressure-related quantities. Relative to the single-scale variant, the largest reported benefit is in intermediate surface-pressure bands, while the volume total-pressure results are nearly indistinguishable. This is useful evidence of a frequency-dependent error reduction in a specific setting, but it does not, by itself, establish an improvement in frequency-dependent learning dynamics or a general mitigation of high-frequency underfitting (Sections 1 and 5.2; Figure 3).

  • The presentation is fragmented and obscures the methodological argument. Section 3.1, titled “Problem Description,” focuses largely on scalability choices instead of clearly formalizing the prediction task. The overall architecture diagram and important training details appear only in Section 4.2 (“Architecture”), after the individual modules have already been described. Section 3.4 includes “Loss” in its title without clearly stating a unified objective for the main and auxiliary predictions. In addition, Section 3.2.1 discusses accuracy and runtime equivalence between two neighborhood-sampling strategies, although Table 1 reports only neighborhood distances. A clearer separation of problem formulation, method, training objective, and empirical evidence would substantially improve readability (Sections 3–4; Table 1; Figure 1).

Questions for authors

Please refer to weaknesses.

Reviewer_1ZpVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FlowGEM, a neural surrogate architecture based on multi-scale attention that reads a pool around anchor points for geometry conditioning and has an additional physics enrichment mechanism that is trained with an auxiliary decoder. FlowGEM is trained on two CFD datasets, DrivAerML and NASA CRM and evaluated against recent baseline methods. FlowGEM achieves competitive performance and outperforms baselines in 4 out of 6 cases.

Critical strengths
  • The proposed mechanisms for multi-scale attention and physics enrichment seem sound and it's plausible that they achieve real improvements over comparable architectures like AB-UPT or SMART.
  • The paper addresses a relevant problem to improve the accuracy of neural surrogates for general million-scale geometries.
  • Strong performance against recent baselines on relevant dataset DrivAerML
Critical weaknesses
  • The manuscript appears rushed, and the overall presentation can be improved substantially. The writing is refined and good in many places, and I quite enjoyed reading some parts like the introduction, but then there are many text passages that are hard to follow and can be supplemented with illustrations. I'm not sure if I would be able to reproduce the architecture based on the description in the manuscript.
  • The manuscript lacks in-depth ablations of the core contributions: the multi-scale attention with different scale levels and the physics enrichment with auxiliary decoder. How do training and inference costs compare to the closest relatives and Transolver-3? An interesting comparison would be GeoTransolver (https://arxiv.org/pdf/2512.20399) which also uses multi-scale attention.
  • Ideally include additional datasets and baselines for more comprehensive benchmarking results.

I think the core ideas presented in the paper could be very useful and I'm happy to increase my rating if the above points are addressed. In the current state, I think the manuscript is not ready for publication and needs more time.

Questions for authors

See questions raised in weaknesses. Additionally:

  • Can you give more details on the auxiliary decoder and loss? Have you just used the supervision loss, but on the anchor points here?
  • Why did you not show the reproduced numbers of Transolver-3 and AB-UPT if they were close to the ones reported in the paper?

Comments on the paper

1 comment

Dear reviewers and area chairs,

We thank you sincerely for your thorough reading of our work and your suggestions to improve it. Unfortunately, we will not be able to address all of the comments at this time, so we have decided to withdraw the submission. We have already started improving the work, including additional datasets, baselines and more thorough ablation studies. We will take all of your feedback into account in the next version of this work.

Best regards, The authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.