acceptodds
Under review as a conference paper at ICLR 2027

AMG: Adaptive Multi-point Guidance for Aerodynamic Design via Diffusion Models

Abstract

Aerodynamic inverse design aims to find a wing geometry that achieves specified lift and drag coefficients across multiple flight conditions. Diffusion models have recently emerged as a promising tool for this task, but existing approaches coordinate multiple operating conditions through manually tuned or heuristic weighting schemes, which cannot adaptively resolve conflicts among targets. We propose Adaptive Multi-point Guidance (AMG), an inference-time guidance framework that steers a pretrained diffusion model by solving a linear system in the latent space. At each denoising step, AMG evaluates the denoised prediction with a differentiable surrogate model, constructs the physical Jacobian of the aerodynamic coefficients with respect to the latent variable, and solves for a latent displacement whose pseudo-inverse solution automatically determines the contribution of each target. This formulation replaces the noise-space scalar weighting of classifier guidance with a latent-space vector displacement, and enables adaptive coordination of competing aerodynamic objectives without manual weight tuning. On a set of multi-point design tasks constructed from the SuperWing dataset, AMG reduces the average MAPE by 69% relative to the best-performing baseline among those considered in this work, where the MAPE is evaluated by the same surrogate model used for guidance and reported at the best intermediate step of sampling. These results suggest that AMG has the potential to be applied in practical aerodynamic design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
aerodynamic inverse design, diffusion models, classifier guidance, adaptive weighting, multi-point design

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_W1B4OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces diffusion guidance mechanism (AMG) for the Aerodynamic Design via Diffusion Models. In introduces adaptive guidance weights based on the pseudo-inverse of the Jacobian (of the target properties with respect to the noisy input). Experiments on the SuperWing dataset report improvements in surrogate-evaluated prediction error over several generative baselines.

Critical strengths
  1. The paper addresses a practically relevant problem involving multiple aerodynamic operating conditions.

  2. The method is straightforward to implement and does not require retraining the diffusion backbone for different target combinations.

Critical weaknesses
  1. Novelty: The main contribution of this paper (Eqs. 15-17), is essentially a classical pseudoinverse correction applied during diffusion sampling. Closely related approaches already exist, including Pseudoinverse-Guided Diffusion Models [1] Physics-Constrained Flow Matching [2], and Geometry-Correct Diffusion Posterior Sampling [3]. The manuscript does not sufficiently distinguish its mathematical contribution from these approaches. Moreover, the formulation performs local least-squares correction without establishing a principled resolution of conflicting objectives.

  2. Insufficient independent physical validation: another critical weakness is that the entire experiments rely on the setup where the same learned surrogate model is used for guidance and evaluation. Consequently, the reported improvements may partially reflect optimization of surrogate-specific errors rather than actual aerodynamic performance. Independent CFD evaluations of generated designs would substantially strengthen the experimental conclusions.

[1] Pseudoinverse-Guided Diffusion Models for Inverse Problems (Song et al., 2023 ICLR).

[2] Physics-Constrained Flow Matching: Sampling Generative Models with Hard Constraints (Utkarsh et al., NeurIPS 2025).

[3] Geometry-Correct Diffusion Posterior Sampling with Denoiser-Pullback Curvature Guidance and Manifold-Aligned Damping, (Shin et al., ICML 2026).

Questions for authors
  1. Can authors elaborate the methodological novelty compared to the above related works [1]-[3]? To my understanding, this is training-free guidance [4]-[5] where instead of direct gradient used in the classical DPS [5], Pseudoinverse is utilized and this was investigated before in the literature.

  2. Satisfying multiple desired properties is an important problem, but that should keep the original diffusion sampler's basic properties (e.g., TACS [6]). For instance, training-free control of the target properties could break the basic properties of the generated samples (which indirectly proved as measuring ratios of Invalid ones but this heuristic base counting is not enough). Can authors further investigate whether the method does not break the basic ability of the trained diffusion models?

  3. Can author investigate another surrogate models (trained separately with another data) to check whether the current evaluation setup is reliable?

I am willing to raise my score if above concerns are resolved.

[4] TFG: Unified Training-Free Guidance for Diffusion Models (Ye et al., 2024 NeurIPS).

[5] Diffusion Posterior Sampling for General Noisy Inverse Problems (Chung et al., 2023 ICLR).

[6] Conditional Synthesis of 3D Molecules with Time Correction Sampler (Jung et al., 2024 NeurIPS).

2 comments

Thank you for recognizing the practical relevance and ease of implementation, and for your suggestions on related work, generation quality, and independent evaluation. We acknowledge overlooking relevant image-domain research. We added ΠGDM, PCFM, and CLAMP (the method proposed in Geometry-Correct Diffusion Posterior Sampling), and included MMPS at the request of other reviewers.


1. Differences from related methods and the adaptive mechanism

These methods are closely related to AMG; pseudoinverse correction itself is not unique to our work.

AMG aims to determine multi-target guidance adaptively. It first specifies the desired aerodynamic change for the current step based on the target residuals and remaining sampling steps. It then uses the full local sensitivities to jointly determine each target's contribution and the overall latent correction magnitude, without searching for target weights or an additional global guidance strength for different target combinations.

ΠGDM constructs guidance through conditional-score approximation, PCFM applies constraint corrections during flow-based generation, and CLAMP constructs updates using curvature and damping. Their correction magnitudes are determined by their respective sampling updates, step sizes, or damping configurations. AMG instead explicitly incorporates how much the aerodynamic targets should change in the current step into the displacement solve, giving the correction magnitude a direct target-based interpretation under the local linear model.

The distinction is jointly determining multi-target coordination and correction magnitude through the full aerodynamic mapping, rather than introducing a pseudoinverse. We compare these methods on the same design tasks.


2. Generation quality and additional experiments

We agree that visual inspection and invalid-geometry counts alone cannot fully establish whether guidance preserves the original model's generation quality. Although some methods produce visibly distorted shapes, visual inspection did not reveal obvious self-intersections, negative thickness, or similar geometric validity issues in the generated geometries.

To further quantify whether generated shapes depart from the dataset distribution, we use a vertex-distance-based shape similarity measure: the RMS distance between corresponding vertices of a generated geometry and its nearest training geometry, normalized by the reference chord length. This supplements geometric validity checks with a quantitative assessment of overall shape similarity to the training data.

We also performed independent CFD evaluations using ADflow, with the main physical settings based on SuperWing and computational meshes reconstructed from the generated geometries. The results are summarized below.

Method groupMethodSurrogate MAPE (%)CFD MAPE (%)Shape distance (%)
Previously includedStandard CG14.0325.791.80
NRCG22.7626.041.75
CcDPM27.6137.271.74
Airfoil-DDPM26.8142.311.91
Dflow-SUR18.6353.603.28
Newly addedΠGDM2.7539.922.97
MMPS8.5760.682.89
PCFM5.9478.485.41
CLAMP10.4350.413.29
OursAMG4.3234.271.92

The shape-distance reference for the held-out test geometries is 2.14%, compared with 1.92% for AMG, indicating that AMG-generated geometries remain close to the training data under this overall distance measure. Although ΠGDM achieves the lowest surrogate MAPE, visual inspection reveals severe geometric distortions, showing that minimizing surrogate target error alone does not ensure plausible designs.

This distance provides a coarse supplementary assessment of shape distribution shift. It is not a sufficient criterion for geometric validity and cannot by itself establish that the original generation distribution and diversity are fully preserved.

AMG's surrogate advantage is not retained in the current CFD evaluation. Some surrogate-guided methods exhibit lower surrogate target error but higher CFD target error, suggesting that optimization may exploit surrogate prediction errors. The underlying mechanism warrants further investigation.

Improving surrogate reliability is outside the main scope of this paper. This issue is not specific to AMG but is a shared risk for methods relying on surrogate models. In practical design, the surrogate can be replaced by a differentiable CFD solver to avoid this source of error from the learned predictor; we have not experimentally evaluated this replacement in the present work.


3. Independent evaluation

We use the ADflow evaluations above as the independent validation reference, rather than adding a second surrogate evaluation.


Thank you again. We welcome follow-up questions and will do our best to clarify any remaining concerns.

Thanks for the response. While the authors provide the relevant answer to my questions, it still doesn't resolve my original concerns of the novelty and in fact, the new experiments indicate the proposed result might be highly overfit to the surrogate reward models (very low Surrogate MAPE but not the lowest other metrics: CFD MAP, Shape distance). Therefore, i will keep my original score toward weak rejection.

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

This paper studies multi-point aerodynamic inverse design using diffusion models, focusing on satisfying multiple design targets while maintaining geometric validity. It proposes an inference-time guidance update that uses a Jacobian matrix and a pseudoinverse to jointly correct discrepancies across aerodynamic targets and adaptively determine their contributions to the update.

Critical strengths

The size of the linear system suits the current task. The shape has 594 degrees of freedom but only four targets, reducing the solve to a 4-by-4 matrix. The matrix solve is lightweight; the main computational cost lies in network evaluation and differentiation.

I particularly appreciate the authors’ disclosure of surrogate-based evaluation, best-intermediate-step and final-step results, and failed geometries. The scaling ablation illustrates the trade-off between target error and geometric validity.

Critical weaknesses

The same surrogate model provides the optimization gradients and evaluates the resulting designs. The method may improve aerodynamic performance, but it may also partly exploit surrogate errors on generated geometries. The current experiments cannot distinguish these possibilities.

Each task includes only two operating conditions, which limits the evaluation. It is unclear how representative these pairs of conditions are.

Questions for authors
  • Why is the multi-point evaluation limited to two operating conditions per task?

  • Given the importance of surrogate training, could the surrogate model suffer from collapse, or could the optimization exploit its prediction errors through reward hacking?

  • Given the differentiable surrogate and shape parameterization, why not directly apply damped least-squares optimization? What additional value does the diffusion prior provide under the same computational budget? Could you compare against direct optimization using the same FFD representation, surrogate model, and budget, initialized from dataset wings or diffusion-generated wings?

Please also address the concerns raised under Weaknesses.

2 comments

Thank you for recognizing the computational structure, experimental transparency, and scaling ablation, and for your specific suggestions regarding the number of operating conditions, surrogate reliability, and the value of the diffusion prior. We address these points below and report additional experiments.


1. Why two operating conditions per task?

The main reason is computational cost. Each additional operating condition requires further aerodynamic surrogate evaluations and differentiation. We therefore used two conditions to evaluate multi-point design within the available computational budget. Each condition includes CL and CD, giving four scalar targets that must be satisfied simultaneously.


2. Surrogate error exploitation and independent CFD validation

We acknowledge that optimization can exploit surrogate prediction errors through reward hacking. Our additional experiments reveal behavior consistent with this risk: lower surrogate target error does not necessarily translate into lower CFD target error.

We performed independent evaluations using ADflow, with the main physical settings based on SuperWing and computational meshes reconstructed from the generated geometries. To address requests raised in the other reviews, we also added ΠGDM, MMPS, PCFM, and CLAMP and measured shape distance. The results are summarized below.

Method groupMethodSurrogate MAPE (%)CFD MAPE (%)Shape distance (%)
Previously includedStandard CG14.0325.791.80
NRCG22.7626.041.75
CcDPM27.6137.271.74
Airfoil-DDPM26.8142.311.91
Dflow-SUR18.6353.603.28
Newly addedΠGDM2.7539.922.97
MMPS8.5760.682.89
PCFM5.9478.485.41
CLAMP10.4350.413.29
OursAMG4.3234.271.92

Shape distance is the RMS distance between corresponding vertices of a generated geometry and its nearest training geometry, normalized by the reference chord length; the corresponding value for the held-out test geometries is 2.14%. It measures overall shape similarity rather than providing a sufficient criterion for geometric validity. Although ΠGDM achieves the lowest surrogate MAPE, visual inspection reveals severe geometric distortions, showing that minimizing surrogate target error alone does not ensure plausible designs.

AMG's surrogate advantage is not retained in the current CFD evaluation. Among the methods using surrogate-based guidance, some results exhibit lower surrogate target error but higher CFD target error, suggesting that optimization may exploit surrogate prediction errors. The underlying mechanism warrants further investigation.

Improving surrogate reliability is outside the main scope of this paper. This issue is not specific to AMG but is a shared risk for methods relying on surrogate models. In practical design, the surrogate can be replaced by a differentiable CFD solver to avoid this source of error from the learned predictor; we have not experimentally evaluated this replacement in the present work.


3. Direct optimization and the value of the diffusion prior

We conducted additional damped least-squares optimization experiments using the same FFD representation and aerodynamic surrogate. With either dataset wings or diffusion-generated wings as initialization, direct optimization achieved surrogate target errors below 0.1% on most tasks.

The value of the diffusion prior primarily concerns shape control. In engineering practice, direct optimization typically incorporates shape parameterization, geometric constraints, or regularization to control geometric changes, with manual inspection and correction when necessary. The diffusion process instead provides a learned geometric prior that continues to influence shape evolution alongside aerodynamic guidance.

The motivation for incorporating diffusion is therefore to guide geometric evolution using a learned shape prior and reduce reliance on manual initialization selection and shape correction. AMG addresses how to coordinate multiple aerodynamic targets adaptively within this generative process.


Thank you again for your suggestions. If any concerns remain or further clarification is needed, we welcome follow-up questions and will do our best to address them.

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts to address my concerns.

However, the new CFD results do not support AMG's claimed performance advantage in actual aerodynamic design, and the added value of the diffusion prior over direct optimization remains unclear.

Given these unresolved concerns, I will maintain my original scores.

Reviewer_jp7iOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The aim of the paper is to use a diffusion model to perform multi-point aerodynamic design. To achieve this, the authors propose an inference-time guidance method which balances the multi-point objective. The reward function is given by a surrogate model evaluating the lift/drag coefficients of the FFD realized from the current Tweedie estimate. The whole process is differentiable, allowing them to use standard reward guidance techniques. The authors argue that existing approaches weight the multi-point objectives using heuristic weighting schemes, and instead proposes to use the jacobian to solve for the contribution of each part. The proposed method performs well on a benchmark dataset compared to relevant baselines.

Critical strengths

Method is simple, tuning-free and shows an improvement over the baselines. The experiments are controlled and, to the best of my knowledge, relevant to the task.

Critical weaknesses

The proposed guidance method is very similar to previously proposed methods such as ΠGDM [1] and MMPS [2]. All three compute Δxt=J⊤G−1r\Delta x_t = J^\top G^{-1}r where J is the jacobian, r is the residual and G given by:

GG: HH⊤HH^\top (ΠGDM), σt2HDtH⊤+Σy\sigma_t^2HD_tH^\top + \Sigma_y (MMPS) and HDtDt⊤H⊤HD_tD_t^\top H^\top

Multi-point targets are simply a vector-valued observations, which both methods already handle. Neither work is cited or compared against. The heuristic scaling in AMG is replaced in MMPS with the observation noise which gives it a principled bounded target.

[1] Song, Jiaming, et al. "Pseudoinverse-guided diffusion models for inverse problems." International conference on learning representations. 2023.

[2] Rozet, François, et al. "Learning diffusion priors from observations by expectation maximization." Advances in Neural Information Processing Systems 37 (2024): 87647-87682.

Questions for authors
  • State how AMG relates to ΠGDM and MMPS, and why the choice of GG is preferred over the alternatives
  • Add MMPS as a baseline, with Σy\Sigma_y chosen from the surrogate error (or some other relevant number)
2 comments

Thank you for pointing out these important related works. We acknowledge that our literature review overlooked relevant research in the image domain, leaving the related-work discussion and baseline comparisons incomplete. We have added comparisons with ΠGDM and MMPS and, in response to requests raised in the other reviews, further included PCFM and CLAMP. Below, we clarify their relationship to AMG and present the additional experiments.


1. Differences from ΠGDM and MMPS and the choice of G

AMG aims to determine multi-target guidance adaptively: it jointly computes each target's contribution and the overall correction magnitude from the current residuals and local sensitivities, without retuning target weights or an additional global guidance strength for different target combinations.

ΠGDM and MMPS also coordinate multiple targets through a pseudoinverse or a related linear-system solve. However, the resulting guidance signal does not directly specify how much the targets should change in the current step; its effective strength still depends on prescribed sampling steps, time schedules, or noise-scale settings. AMG instead first specifies the desired target change and then solves for the corresponding latent displacement. This gives the correction magnitude an explicit target-based interpretation under the local linear model, without a task-specific search for an additional strength coefficient.

The different forms of G follow from these different objectives and modeling assumptions. ΠGDM adopts an isotropic approximation to denoising uncertainty, and its G describes how this uncertainty propagates through the observation mapping. MMPS further incorporates the denoising covariance and observation noise to approximate the conditional observation distribution. AMG directly takes the pseudoinverse of the full aerodynamic Jacobian, naturally yielding G=JtJt⊤G=J_tJ_t^\top, which corresponds to the minimum-norm latent correction for the specified local target change.

Thus, AMG's G is not an arbitrarily chosen weighting matrix but follows from its adaptive displacement formulation. The distinction we emphasize is the joint determination of multi-target coordination and correction magnitude, rather than the pseudoinverse itself.


2. MMPS noise calibration and additional experiments

Following your suggestion, we randomly selected 64 held-out geometries, comprising 397 operating-condition samples, and used the RMSE of the surrogate prediction residuals to set the MMPS observation noise. The noise scales for CL and CD are 0.003699 and 0.000382, respectively, from which we construct a fixed diagonal covariance matrix.

We adapted public implementations of ΠGDM, MMPS, PCFM, and CLAMP to the aerodynamic observation operator. We also performed independent CFD evaluations using ADflow, with the main physical settings based on SuperWing and computational meshes reconstructed from the generated geometries. Shape distance was measured to quantify the similarity between generated geometries and the training data.

Method groupMethodSurrogate MAPE (%)CFD MAPE (%)Shape distance (%)
Previously includedStandard CG14.0325.791.80
NRCG22.7626.041.75
CcDPM27.6137.271.74
Airfoil-DDPM26.8142.311.91
Dflow-SUR18.6353.603.28
Newly addedΠGDM2.7539.922.97
MMPS8.5760.682.89
PCFM5.9478.485.41
CLAMP10.4350.413.29
OursAMG4.3234.271.92

Shape distance is the RMS distance between corresponding vertices of a generated geometry and its nearest training geometry, normalized by the reference chord length; the corresponding value for the held-out test geometries is 2.14%. This metric measures overall shape similarity rather than providing a sufficient criterion for geometric validity. Although ΠGDM achieves the lowest surrogate MAPE, visual inspection reveals severe geometric distortions, showing that minimizing surrogate target error alone does not ensure plausible designs.

AMG's surrogate advantage is not retained in the current CFD evaluation. Among the methods using surrogate-based guidance, some results exhibit lower surrogate target error but higher CFD target error, suggesting that optimization may exploit surrogate prediction errors. This phenomenon warrants further investigation.

Improving surrogate reliability is outside the main scope of this paper. This issue is not specific to AMG but is a shared risk for methods relying on surrogate models. In practical design, the surrogate can be replaced by a differentiable CFD solver to avoid this source of error from the learned predictor; we have not experimentally evaluated this replacement in the present work.


Thank you again for your suggestions. If any concerns remain or further clarification is needed, we welcome follow-up questions and will do our best to address them.

Thank you for answering my questions and implementing the suggested baselines.

While the new results strengthen the related work section, it unfortunately weakens the case for the paper. The results achieved by ΠGDM look very similar to the no-scaling variant, indicating that AMG is a variant of existing methods rather than a new guidance method.

The current draft also relies heavily on the surrogate error being lowest for their method. With ΠGDM and related guidance methods showing similar/better results, this framing is a bit weaker. More importantly, the additional CFD results contradict the paper's main claim (AMG at 34.3% MAPE vs 25.8% for standard CG) and cannot be treated as out of scope.

Given these unresolved concerns, I'm maintaining my original score.

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

This paper introduces AMG, a method for guiding diffusion models to generate wings that meet aerodynamic targets at multiple flight conditions. Instead of manually weighting the targets, AMG uses their local sensitivities to calculate a joint correction during sampling. It substantially improves target matching in the reported experiments, although the results are evaluated using the same aerodynamic surrogate that guides generation, without independent CFD validation.

Critical strengths

The method provides a clear and practical way to coordinate multiple design targets. Rather than combining target gradients with manually chosen weights, AMG calculates a joint latent correction from the aerodynamic Jacobian. This is a useful contribution because competing requirements are central to multi-condition wing design, and fixed guidance weights can be difficult to choose. The reported improvement is large and persists beyond the best intermediate step. AMG reduces average best-step MAPE from 14.03% for the strongest reported baseline to 4.32%. Appendix G also shows the lowest final-step error, at 9.24%. This makes the result more convincing than an improvement that appears only at a carefully selected intermediate point, although both comparisons still rely on the surrogate.

Critical weaknesses

The evaluation does not establish whether improved surrogate predictions translate into improved aerodynamics. The same surrogate supplies the gradients, selects the best intermediate geometry, and evaluates the final performance. This leaves open the possibility that optimization exploits surrogate errors, particularly on generated shapes outside its training distribution. The reported 4.4% field-prediction error does not resolve this concern because it does not directly establish lift and drag accuracy on the optimized designs. Independent CFD checks would materially strengthen the paper’s central result. An apparent sign error puts the strongest baseline comparison in question. Appendix D.4 defines a positive squared-error loss, but Equation (22) subtracts its gradient from the predicted noise. Substituting this into the reverse update in Equation (8) adds a positive loss-gradient displacement, which appears to increase the loss locally. This may be a manuscript error rather than an implementation error, but standard CG is the baseline used for the headline 69% improvement. Until the implementation is clarified, I cannot confidently interpret that comparison.

Questions for authors

In Appendix D.4, is the minus sign in Equation (22) a typographical error, or was that update used in the experiments? With the positive squared-error loss in Equation (21), it appears to produce an ascent direction when substituted into Equation (8). Please clarify the implemented sign and whether correcting it changes the standard CG results.

Have any AMG-generated designs been evaluated with an independent CFD solver? If so, how closely do their lift and drag coefficients match the surrogate predictions, and does AMG retain its advantage over the baselines? This is particularly relevant for the intermediate geometries selected by minimum surrogate error.

Were the results repeated across independent sampling seeds? Please clarify the variability of the reported gains and whether the sampled two-condition target combinations are known to be jointly achievable by a single wing. These details would help distinguish optimization failures from infeasible design requests.

1 comment

Thank you for recognizing the value of the joint multi-target correction and for your specific suggestions regarding the CG implementation, independent physical validation, and experimental variability. We have checked the CG implementation and added independent CFD evaluations, additional baselines, and sampling-seed experiments.


1. Sign in the CG update

The minus sign in Eq. (22) is a typographical error and should be a plus sign. The implementation uses the correct loss-descent direction. Correcting the equation therefore does not change the reported CG results, including its best-step MAPE of 14.03%.


2. Independent CFD validation and additional baselines

We performed independent CFD evaluations using ADflow, with the main physical settings based on the SuperWing paper and computational meshes reconstructed from the generated geometries.

To address requests raised in the other reviews, we also added ΠGDM, MMPS, PCFM, and CLAMP as strong baselines. We measured the nearest-neighbor distance to the training set to quantify how closely the generated shapes resemble the training geometries. For the methods already included in the submission, the surrogate results below are those reported in the paper.

Method groupMethodSurrogate MAPE (%)CFD MAPE (%)Shape distance (%)
Previously includedStandard CG14.0325.791.80
NRCG22.7626.041.75
CcDPM27.6137.271.74
Airfoil-DDPM26.8142.311.91
Dflow-SUR18.6353.603.28
Newly addedΠGDM2.7539.922.97
MMPS8.5760.682.89
PCFM5.9478.485.41
CLAMP10.4350.413.29
OursAMG4.3234.271.92

Shape distance is the RMS distance between corresponding vertices of a generated geometry and its nearest training geometry, normalized by the reference chord length. The corresponding value for the held-out test geometries is 2.14%. This metric measures overall shape similarity rather than providing a sufficient criterion for geometric validity. Although ΠGDM achieves the lowest surrogate MAPE, visual inspection reveals severe distortions in its generated geometries, showing that minimizing surrogate target error alone does not ensure plausible designs. The reconstructed meshes and some historical geometry-scale settings are not fully uniform, so the CFD results should be interpreted as independent checks rather than a strictly controlled ranking.

AMG's advantage under surrogate evaluation is not retained in the current CFD evaluation. Among the methods using surrogate-based guidance, some results exhibit lower surrogate target error but higher CFD target error, suggesting that optimization may exploit surrogate prediction errors. This phenomenon warrants further investigation.

Improving surrogate reliability is outside the main scope of this paper. This limitation is not specific to AMG but is a shared risk for methods that rely on surrogate models. In practical design, the surrogate can be replaced by a differentiable CFD solver to avoid this source of error from the learned predictor; we have not experimentally evaluated this replacement in the present work.


3. Independent sampling seeds and target feasibility

Due to time constraints, we added two sampling seeds only for the fixed task pair0 and combined them with the existing run to obtain three runs per method. The following table reports the mean and sample standard deviation of the best-step surrogate MAPE.

Method groupMethodMAPE (%)
Previously includedStandard CG4.31 ± 1.07
NRCG5.91 ± 1.72
CcDPM6.74 ± 1.37
Airfoil-DDPM6.81 ± 1.92
Dflow-SUR17.82 ± 10.45
Newly addedΠGDM1.12 ± 0.15
MMPS3.32 ± 1.48
PCFM5.60 ± 1.57
CLAMP4.51 ± 0.21
OursAMG0.91 ± 0.64

These results characterize variability on pair0 only and do not replace repeated experiments across all tasks.

We also acknowledge that the randomly combined two-condition targets may not be jointly achievable by a single physical wing. Consequently, target-matching errors cannot be attributed entirely to optimization failures. We will explicitly clarify this feasibility limitation in the paper.


Thank you again for your careful review. If any concerns remain or further clarification is needed, we welcome follow-up questions and will do our best to address them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.