acceptodds
Under review as a conference paper at ICLR 2027

FSAN: Flow State Attention Network for Aerodynamic Prediction

Abstract

Accurate aerodynamic prediction is critical for designing fuel-efficient and safe transportation systems such as aircraft and automobiles, yet traditional computational fluid dynamics (CFD) simulations remain computationally expensive and expertise-intensive, severely limiting their use in iterative design and real-time analysis. Existing deep learning surrogates suffer from two major limitations: (i) they are evaluated on datasets with narrow flow-condition ranges, leaving their performance under complex flow conditions undemonstrated; (ii) they treat global flow conditions as a single vector injected uniformly across all surface points, ignoring that different geometric regions experience distinct local flow phenomena, which degrades prediction accuracy under complex flow conditions. To address these limitations, we propose the Flow State Attention Network (FSAN). FSAN separately encodes point cloud and flow conditions, then partitions the geometry into multiple flow states via learnable soft assignments, and uses flow features to update these state representations, which in turn influence point cloud features through state changes. This enables fine-grained, state-specific interaction between geometry and flow information. Extensive experiments on two well-recognized aerodynamic benchmarks demonstrate that FSAN achieves the highest accuracy among the methods compared in this work at a higher computational cost. On Emmi-Wing, FSAN reduces the Relative L2 (REL-L2) error by over 20% compared to the strongest baseline (Transolver), and on DrivAerNet++, it achieves a 10% reduction compared to the strongest baseline (AdaField). These results establish FSAN as a promising neural surrogate on public benchmarks with diverse flow conditions and geometries. The code is available in the supplementary material.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
aerodynamic prediction, neural surrogate model, attention mechanism

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KFSEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a new aerodynamic prediction network, FSAN, which adaptively divides the 3D geometric surface into multiple potential-flow states, allowing different regions to exhibit different feature modulations based on global inflow conditions, thereby improving the accuracy of surface pressure predictions.

Critical strengths
  1. The paper notes that most current methods apply operating condition parameters as global conditions to all surface points in the same way. However, these global conditions can affect different geometric regions differently. As a result, different surface regions may need to handle operating condition information in their own ways. This insight offers a fresh perspective on how to combine geometry and operating conditions.
  2. The paper introduces an attention mechanism designed for combining geometry and condition information. It uses soft assignment and weighted aggregation of point features, along with cross attention. This approach allows different latent states to interact with condition information separately, enabling region-specific conditioning.
Critical weaknesses
  1. At present, the division of latent states and the conditioning of operating conditions are mainly learned through predictive tasks using data, without direct physical supervision. As a result, current experiments do not clearly show whether performance improvements come from the model learning effective adaptive division and conditioning, or if they are mainly due to the more complex attention structure. It is also possible that simply making the model structure more complex leads to better predictive performance.
  2. The main module described in the article is structurally very similar to existing methods. Both FSA and Transolver's physics attention use similar ideas, such as soft allocation, weighted aggregation, and feature backpropagation [1]. What sets FSA apart is its use of condition tokens and the way it conditions latent states through cross-attention. This means its main innovation is in the conditional mechanism within the latent state interaction framework, rather than in how regions are divided. Other related methods, like Graph Transformer [2], also use similar structures. The paper should explain more clearly what makes FSA a significant improvement over existing methods.

References:

[1] Wu, Haixu, et al. "Transolver: A fast transformer solver for pdes on general geometries." ICML 2024.

[2] Janny, Steeven, et al. "Eagle: Large-scale learning of turbulent fluid dynamics with mesh transformers." ICLR 2023.

Questions for authors
  1. The ablation experiments show that FSA performs better than the other fusion mechanisms tested. How can we tell if this improvement comes from adaptive state partitioning and working condition modulation, or if it is mainly due to the more complex attention structure?
  2. FSA and Transolver both use similar physics attention methods, such as soft allocation, weighted aggregation, and feature backpropagation. Could you explain more clearly what the main innovations of FSA are compared to Transolver? Also, what specific benefits do condition tokens and cross attention bring?
  3. The paper uses data-driven learning through prediction tasks to achieve latent state partitioning. Besides showing visual results and their link to flow characteristics, what other evidence can show that the model has truly learned adaptive state partitioning that changes with operating conditions, and that this is a key reason for better performance?
  4. The paper has not yet included specific geometric OOD or operating condition OOD tests. Without these results, how do you assess FSA’s ability to generalize to new geometries and operating conditions? Is it necessary to further verify its advantages over existing methods through targeted out-of-distribution experiments?
2 comments

Thank you for recognizing the motivation and design of our work, and for raising questions about the method, experimental evidence, and generalization. We respond to each point below.


1. How do we verify the role of adaptive state partitioning?

To further verify the role of adaptive state partitioning, we trained a fixed-partition control model whose soft assignments are computed solely from geometric features and therefore do not vary with operating conditions. The Emmi-Wing results for FSAN and the fixed-partition model are as follows:

MethodMSE (×10⁻³)MAE (×10⁻²)REL-L1 (%)REL-L2 (%)
FSAN0.2970.4561.402.59
Fixed-partition control0.3540.5361.653.06

The fixed-partition model performs worse on all four metrics, indicating that allowing the state partition to vary with operating conditions improves prediction performance.


2. What are FSAN’s innovations and advantages over Transolver?

The original Transolver does not have a dedicated flow-condition encoding and interaction branch. When adapting it to this task, one can broadcast the global flow conditions to every point and concatenate them with the point features. This places semantically different information—local geometric coordinates and global operating parameters—in the same point-feature representation. The model must then learn to distinguish these inputs and determine how to combine the global conditions with geometric features from different regions. The concatenated conditions also enter subsequent point-feature processing and soft partitioning, coupling geometric grouping with condition information from the outset and making it harder to learn effective partitions and region-specific responses to conditions.

FSAN encodes flow conditions separately as condition tokens and uses cross-attention to interact with the geometric latent states. Each latent state can select and integrate relevant information from the condition tokens based on its own geometric features, enabling region-specific condition modulation. The soft partition can also be formed from geometric features first, without simultaneously processing geometric coordinates and global operating parameters during partitioning.


3. What evidence shows that the partition varies with operating conditions and that this variation improves performance?

FSAN computes its soft assignments from features that have interacted with the operating conditions. Thus, the same geometry can produce different partitions under different conditions. The fixed-partition control lacks this condition dependence; as shown in the results under Question 1, it performs worse on all four prediction metrics. This supports our view that allowing state partitions to adapt to operating conditions is an important factor in FSAN’s performance gains.


4. How do we assess generalization to out-of-distribution geometries or operating conditions?

We agree that targeted OOD tests are important for assessing generalization to unseen geometries and operating conditions. However, publicly available aerodynamic datasets remain limited in geometric coverage, operating-condition ranges, and data quality, making it difficult to construct reliable and representative OOD evaluations. Therefore, our results support FSAN’s performance only on the distributions covered by the current datasets; they do not establish stronger generalization to unseen geometries or operating conditions. Developing high-quality OOD benchmarks and conducting systematic evaluations are important directions for future work.


We thank the reviewer again for the suggestions, which have helped us clarify the contributions of our method and the scope of the conclusions supported by our experiments.

Reviewer_KFSEOpenReview ↗

Official Comment by Reviewer KFSE

I thank the authors for the additional experiments and clarifications. The fixed-partition ablation provides useful evidence for the benefits of adaptive state partitioning. However, I remain unclear about the methodological novelty of FSAN beyond existing physics-attention mechanisms such as Transolver, particularly whether condition tokens and cross-attention constitute a sufficiently substantial advance. Additionally, the lack of targeted OOD evaluations on unseen geometries and operating conditions leaves its generalization advantages unverified. Therefore, I maintain my original rating.

Reviewer_h9ewOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FSAN predicts surface pressure coefficients from geometry and operating conditions. It combines local point attention with learned soft partitions, updates the pooled states through cross-attention to flow tokens, and projects them back to points. Experiments on Emmi-Wing and PIDA-augmented DrivAerNet++ report lower prediction errors than the evaluated baselines.

Critical strengths
  1. A reasonable architectural idea.
  2. Measurable empirical gains.
Critical weaknesses
  1. Insufficient evidence for flow-condition learning. DrivAerNet++ uses a fixed original velocity, while the additional conditions are introduced through PIDA. It remains unclear whether FSAN learns meaningful flow-dependent behavior or mainly benefits from augmentation. The augmentation protocol and automotive results need clearer validation, ideally with a constant-condition ablation.

  2. Limited methodological novelty and questionable baseline fairness. FSAN largely combines existing attention, latent partitioning, and condition-encoding mechanisms. The improvement over Physics-Attn is modest (2.68% to 2.59% REL-L2), leaving the benefit of the proposed mechanism insufficiently established. The unusually weak PointTransformerV3 results and lack of model-specific tuning also raise concerns about baseline fairness.

Questions for authors
  1. What exact PIDA transformation and sampling range are used, how are coordinates and CpC_p normalized, and is augmentation applied to every baseline and at test time? Please distinguish the original fixed-speed evaluation from augmented-condition evaluation.
  2. Does the automotive gain persist when flow-dependent tokens are replaced by learned constant tokens? For Emmi-Wing, can you report the seed variability of Table 3 and explain what the existing ablations establish about the flow encoder, rather than only the fusion mechanism?
  3. For PointTransformerV3, what are the actual feature and coordinate inputs on each dataset? Please clarify that serialization/voxelization uses geometry, and provide a convergence or limited tuning check supporting the reported baseline performance.
1 comment

Thank you for recognizing the architectural idea and empirical gains, and for raising questions about flow-condition evaluation, ablations, and baseline comparisons. We address them below.


1. PIDA augmentation and evaluation

The original DrivAerNet++ dataset uses a fixed velocity. Following AdaField’s PIDA setup, we introduce geometric-scale and velocity variations through G′=G/cG'=G/c and v′=cvv'=cv, where c∼U(0.2,3.2)c\sim\mathcal{U}(0.2,3.2). We sample the scale during training and fix it, along with point sampling, at test time. Coordinates are centered along the x-axis; CpC_p is clipped to [−1351,451][-1351,451] and divided by 450. All compared methods use the same augmentation.

This paired scaling preserves CpC_p. Thus, the PIDA-augmented DrivAerNet++ evaluation does not test predictions of CpC_p changes caused by flow-condition changes. Emmi-Wing evaluates CpC_p responses to varying flow conditions.


2. Constant condition tokens, Table 3 variability, and the flow-condition encoder

Replacing flow-dependent tokens with constant tokens removes information needed to distinguish operating conditions. For the same geometry under different conditions, this cannot provide meaningful condition-aware predictions. PIDA’s paired transformations preserve CpC_p, while Emmi-Wing directly tests pressure responses under varying conditions.

All Table 3 variants use the same flow-condition encoder design and differ only in the geometry–flow fusion mechanism. The ablation therefore compares fusion mechanisms under a shared condition representation; it does not isolate the encoder’s contribution. The encoder maps physical parameters to condition representations. Removing it would remove flow-condition information, so it is not treated as a separate ablation.

Updated Table 3 results:

VariantMSE (×10−3\times 10^{-3})MAE (×10−2\times 10^{-2})REL-L1 (%)REL-L2 (%)
FSAN (flow state attention)0.297 ± 0.0010.456 ± 0.0011.40 ± 0.002.59 ± 0.01
FSAN + FiLM0.334 ± 0.0050.501 ± 0.0021.65 ± 0.032.83 ± 0.04
FSAN + AdaLN-Zero0.325 ± 0.0060.499 ± 0.0051.61 ± 0.032.80 ± 0.05
FSAN + Cross-Attn0.374 ± 0.0010.602 ± 0.0001.87 ± 0.003.16 ± 0.01
FSAN + Physics-Attn0.298 ± 0.0020.484 ± 0.0001.49 ± 0.002.68 ± 0.02

3. PointTransformerV3 inputs and baseline comparisons

For both datasets, each PointTransformerV3 point uses normalized XYZ coordinates and normalized flow conditions as features. Emmi-Wing conditions are angle of attack, Mach number, and Reynolds number; DrivAerNet++ uses the corresponding PIDA flow conditions. The point cloud is normalized to a unit cube. XYZ coordinates determine geometric voxelization and spatial serialization; flow conditions do not determine voxel assignments or serialization order.

PointTransformerV3 is a general-purpose point-cloud architecture, not one designed specifically for surface-pressure prediction. We use its default configuration and train all methods with the same hyperparameters, without model-specific tuning. Its results therefore reflect the common protocol, not a task-specifically optimized performance ceiling. To broaden the baseline comparison, we add AB-UPT and Transolver++ to the updated Table 1:

DatasetMethodMSE (×10−3\times 10^{-3})MAE (×10−2\times 10^{-2})REL-L1 (%)REL-L2 (%)
Emmi-WingPointTransformerV32.722 ± 0.0081.637 ± 0.0055.17 ± 0.0211.65 ± 0.01
Emmi-WingTransolver0.375 ± 0.0030.517 ± 0.0011.58 ± 0.003.24 ± 0.03
Emmi-WingAdaField0.439 ± 0.0050.620 ± 0.0011.92 ± 0.003.67 ± 0.02
Emmi-WingAB-UPT0.441 ± 0.0050.671 ± 0.0042.07 ± 0.013.54 ± 0.02
Emmi-WingTransolver++0.432 ± 0.0020.651 ± 0.0032.02 ± 0.013.60 ± 0.01
Emmi-WingFSAN (Ours)0.297 ± 0.0010.456 ± 0.0011.40 ± 0.002.59 ± 0.01
DrivAerNet++PointTransformerV39.970 ± 0.0194.547 ± 0.00217.33 ± 0.0029.51 ± 0.01
DrivAerNet++Transolver4.638 ± 0.0093.617 ± 0.00413.78 ± 0.0220.06 ± 0.03
DrivAerNet++AdaField4.559 ± 0.0373.701 ± 0.00214.10 ± 0.0119.89 ± 0.04
DrivAerNet++AB-UPT5.968 ± 0.0333.984 ± 0.01315.17 ± 0.0522.74 ± 0.06
DrivAerNet++Transolver++4.160 ± 0.0493.388 ± 0.01812.92 ± 0.0719.02 ± 0.11
DrivAerNet++FSAN (Ours)3.675 ± 0.0213.273 ± 0.02412.54 ± 0.0217.91 ± 0.02

Thank you again for your constructive feedback. We hope these clarifications and additional results address your concerns.

Reviewer_CUwkOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FSAN predicts surface pressure from geometry and flow conditions. It combines local point attention with learned latent states that aggregate geometric features, interact with flow-condition tokens, and distribute the updated information back to individual points. Experiments on Emmi-Wing and DrivAerNet++ show improved accuracy over the evaluated baselines, using fewer parameters but more computation than Transolver.

Critical strengths

The accuracy improvements are consistent and meaningful. Across three runs, FSAN reduces relative L2 error from 3.24% to 2.59% on Emmi-Wing and from 19.89% to 17.91% on DrivAerNet++. Improvements also appear across the other reported metrics.

The paper includes useful checks beyond the headline results. I appreciate the comparisons between conditioning mechanisms on the same backbone, the alternative conditioning experiments for Transolver, and the full-surface evaluation. The authors also clearly report the computational tradeoff: FSAN uses approximately four times the inference time and 5.8 times the inference memory of Transolver at the standard point count.

Critical weaknesses

The distinction from existing conditioning methods is overstated IMHO. Section 2.2 argues that sharing flow tokens makes cross-attention spatially uniform. However, different geometric queries can attend differently to those same tokens, producing different responses across the surface. FSAN therefore introduces a particular way of grouping and conditioning features, rather than a previously unavailable ability to model regional responses. The controlled ablation is helpful, but the closest physics-attention variant already achieves 2.68% relative L2 error versus FSAN’s 2.59% (about a 3.4% improvement (Table 3)). The paper should distinguish this incremental gain from the 20% whole-model comparison and provide the implementation details and variability needed to assess it.

The flow-condition evaluation needs a more precise interpretation. Emmi-Wing provides a meaningful multiparameter test, although the evaluation uses similar condition distributions across training and testing. For DrivAerNet++, the cited PIDA augmentation jointly rescales geometry and velocity while preserving the pressure coefficient. This adds useful similarity-based examples, but does not independently introduce new dimensionless flow regimes. The automotive gains remain valuable; their relevance to the central flow-conditioning claim needs clearer explanation.

The authors have not compared their approach against other state-of-the-art models such as abupt, transolver++, geotransolver.

Questions for authors

How are flow conditions introduced into the Physics-Attn variant in Table 3? Please report its precise configuration.

What are the exact PIDA transformation, scaling range, coordinate normalization, and training/testing protocols? Which parts of this evaluation require predicting changes in pressure coefficients as flow conditions change?

1 comment

Thank you for recognizing the consistent accuracy improvements and the value of our controlled comparisons. We also appreciate your suggestions on clarifying the scope of our contribution, interpreting the PIDA-augmented evaluation, and adding comparisons with other methods. We first answer your two questions, then respond to the broader concerns.


1. Physics-Attn configuration

The Physics-Attn variant in Table 3 retains FSAN’s local point attention and flow-condition encoder. It encodes the three-dimensional flow condition into three 256-dimensional tokens, which are concatenated with the point features and fed into Physics Attention. The model uses 12 blocks, 4 attention heads, 32 slices, and 16 local neighbors.


2. PIDA augmentation and evaluation

The original DrivAerNet++ dataset uses a fixed velocity. We adopt the PIDA setup proposed by AdaField to extend it with geometric-scale and velocity variations. This augmentation is independent of FSAN’s model design and is applied consistently to the compared methods. Specifically, we use G′=G/cG'=G/c and v′=cvv'=cv, where c∼U(0.2,3.2)c\sim\mathcal{U}(0.2,3.2). Coordinates are centered along the x-axis, and pressure is clipped to [−1351,451][-1351,451] and then divided by 450. The paired scaling preserves CpC_p.

Thus, the PIDA-augmented DrivAerNet++ evaluation does not require predicting changes in CpC_p caused by changing flow conditions. Emmi-Wing provides the evaluation of CpC_p responses to varying flow conditions.


3. Response to the main concerns

Thank you for pointing out that we overstated the distinction from existing methods. We will narrow our claims, acknowledge that geometric queries can also produce spatially varying responses through cross-attention, and describe FSAN’s contribution as a design that groups geometric latent states for condition interaction.

We also provide variability results for Table 3 and add experimental results for AB-UPT and Transolver++ to the main comparison in Table 1.

Updated Table 1: Quantitative comparison on Emmi-Wing and DrivAerNet++

DatasetMethodMSE (×10−3\times 10^{-3})MAE (×10−2\times 10^{-2})REL-L1 (%)REL-L2 (%)
Emmi-WingPointTransformerV32.722 ± 0.0081.637 ± 0.0055.17 ± 0.0211.65 ± 0.01
Emmi-WingTransolver0.375 ± 0.0030.517 ± 0.0011.58 ± 0.003.24 ± 0.03
Emmi-WingAdaField0.439 ± 0.0050.620 ± 0.0011.92 ± 0.003.67 ± 0.02
Emmi-WingAB-UPT0.441 ± 0.0050.671 ± 0.0042.07 ± 0.013.54 ± 0.02
Emmi-WingTransolver++0.432 ± 0.0020.651 ± 0.0032.02 ± 0.013.60 ± 0.01
Emmi-WingFSAN (Ours)0.297 ± 0.0010.456 ± 0.0011.40 ± 0.002.59 ± 0.01
DrivAerNet++PointTransformerV39.970 ± 0.0194.547 ± 0.00217.33 ± 0.0029.51 ± 0.01
DrivAerNet++Transolver4.638 ± 0.0093.617 ± 0.00413.78 ± 0.0220.06 ± 0.03
DrivAerNet++AdaField4.559 ± 0.0373.701 ± 0.00214.10 ± 0.0119.89 ± 0.04
DrivAerNet++AB-UPT5.968 ± 0.0333.984 ± 0.01315.17 ± 0.0522.74 ± 0.06
DrivAerNet++Transolver++4.160 ± 0.0493.388 ± 0.01812.92 ± 0.0719.02 ± 0.11
DrivAerNet++FSAN (Ours)3.675 ± 0.0213.273 ± 0.02412.54 ± 0.0217.91 ± 0.02

Updated Table 3: Mechanism-isolation ablation on Emmi-Wing

VariantMSE (×10−3\times 10^{-3})MAE (×10−2\times 10^{-2})REL-L1 (%)REL-L2 (%)
FSAN (flow state attention)0.297 ± 0.0010.456 ± 0.0011.40 ± 0.002.59 ± 0.01
FSAN + FiLM0.334 ± 0.0050.501 ± 0.0021.65 ± 0.032.83 ± 0.04
FSAN + AdaLN-Zero0.325 ± 0.0060.499 ± 0.0051.61 ± 0.032.80 ± 0.05
FSAN + Cross-Attn0.374 ± 0.0010.602 ± 0.0001.87 ± 0.003.16 ± 0.01
FSAN + Physics-Attn0.298 ± 0.0020.484 ± 0.0001.49 ± 0.002.68 ± 0.02

Thank you again for your constructive feedback. We hope that these clarifications and additional results address your concerns, and we welcome any follow-up questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.