acceptodds
Under review as a conference paper at ICLR 2027

AdaSolver: Efficient Neural Operator Inference via Test-Time Adaptive Sampling

Abstract

Neural operators provide efficient surrogates for partial differential equations (PDEs), yet their inference performance is typically tied to uniform spatial discretizations or sampling schemes. Many PDE solutions exhibit localized structures that require dense sampling for accurate prediction, while smooth regions can be represented with substantially fewer points. Naively changing the sampling density, however, shifts the empirical input measure and can introduce sampling-dependent bias into the learned neural operator. We introduce **AdaSolver**, a measure-aware framework for adaptive spatial computation at test time. We characterize sampling-induced bias and its propagation through physics-attention, and derive a sampling-density-consistent Transolver using self-normalized importance weighting. The resulting measure-consistent formulation preserves consistency with the reference measure, enabling adaptive test-time sampling across different physical systems and geometric designs. Building on this formulation, we explore three refinement strategies based on physics-informed priors, coarse-prediction guidance, and cache-aware learned refinement. Across five standard PDE benchmarks and three industrial aerodynamic tasks, AdaSolver improves prediction accuracy under matched sampling budgets and accelerates surrogate-based airfoil and wing optimization. These results establish adaptive spatial sampling as a principled axis of test-time scaling for neural operators.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operator, Test-Time Scaling, Partial Differential Equations, Scientific Machine Learning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_YVkAOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper attempts to introduce a new direction of test-time scaling for solving PDEs that enables efficient inference of trained neural operators through adaptive spatial sampling. Furthermore, the paper proposes three dynamic test-time sampling methods for efficient inference. Experiments on eight benchmarks show that introducing dynamic test-time sampling can substantially reduce the inference cost while largely maintaining the performance of the trained Transolver.

The reviewer appreciates the motivation of this work and believes that the problem studied in this paper has the potential to become a worthwhile research direction. However, the current experiments do not yet sufficiently demonstrate the necessity and importance of this direction. In addition, several important inference details are missing. Therefore, the reviewer does not believe that the paper has reached the acceptance standard in its current form.

Critical strengths
  1. The paper is easy to follow.
  2. The problem is well motivated.
  3. The experiments cover eight benchmarks and are relatively comprehensive.
Critical weaknesses
  1. The paper makes a key implicit assumption that test-time sampling is an important and worthwhile direction to study. However, beyond the intuitive motivation, the paper lacks sufficient experimental evidence to justify this premise. In particular, if the goal is simply to show that introducing dynamic test-time sampling can improve the inference efficiency of Transolver, then more efficient Transolver-based methods, such as LinearNO, which not only significantly improve efficiency but also achieve better performance, should be considered as important baselines. On the other hand, if the more fundamental claim is that reasonable test-time sampling methods provide an effective way to improve the inference efficiency of already-trained neural operators, then the evaluation should be conducted on different trained backbones, similar to how pruning methods are usually validated across multiple model architectures, rather than only on Transolver. Considering that the community has not yet established a common understanding of test-time sampling, while the authors claim “the first systematic exploration of adaptive spatial sampling as a new axis of test-time scaling for neural operators,” the reviewer considers this a very important issue. Therefore, the motivation should be supported not only by intuitive arguments in the introduction, but also by sufficient experiments.
  2. The paper lacks experimental validation of another important question: how does AdaSolver behave when the same problem is represented with different observation densities, and how does its final performance change accordingly? Answering this question would provide a more continuous evaluation of the authors’ important motivation that different physical regions may require finer or fewer observations depending on their physical properties. The reviewer believes that a series of experiments on Darcy (85x85, 141x141, 211x211, and other resolutions as in Table 10 of the original paper of Transovler) with different observation densities could partially answer this question and further reveal whether AdaSolver exhibits some consistent behavior across observation densities. For example, does the sampling density consistently remain low in smooth regions, while the number of sampled points near boundaries or other critical regions continues to increase as the observation density increases?
  3. The authors explicitly mention full-field evaluation in multiple places. However, it is unclear how AdaSolver reconstructs the full-field prediction from the retained points. If the reviewer has missed the corresponding description, the authors are encouraged to point it out. Otherwise, the reconstruction method should be described in sufficient detail so that the actual inference pipeline of AdaSolver can be clearly understood and reproduced.
  4. Minor weakness: the formatting before equations is not fully consistent. Most descriptions before equations end with a blank space, while some use a colon, such as Equation 12, and some use a comma, such as Equation 4. The reviewer suggests making the formatting consistent throughout the paper.
Questions for authors

Please refer to the weaknesses above. If the authors can sufficiently address the weaknesses, the reviewer would consider raising the score.

5 comments

We thank the reviewer for highlighting the need for stronger experimental support for test-time sampling. Below, we provide detailed responses to the concerns raised:

Our density-design approach to test-time adaptive sampling is a general strategy for neural operators with spatial aggregation. To evaluate its applicability beyond Transolver, we extend AdaSolver to LinearNO [1] as an additional backbone.

Measure correction in LinearNO. In LinearNO, the output is computed as ϕ(Q)(ψ⊤(K)VN)\phi(Q)(\psi^{\top}(K)V_N), where VN=Linear⁡(V)(HN)V_N=\operatorname{Linear}^{(V)}(H_N). The aggregation ψ⊤(K)VN\psi^{\top}(K)V_N uses a softmax over the spatial points. Writing ℓij=[Linear⁡(K)(hi)]j\ell_{ij}=[\operatorname{Linear}^{(K)}(h_i)]_j, each aggregated representation is

sj=∑i=1Nexp⁡(ℓij)vi∑i=1Nexp⁡(ℓij).s_j=\frac{\sum_{i=1}^{N}\exp(\ell_{ij})v_i}{\sum_{i=1}^{N}\exp(\ell_{ij})}.

Changing the spatial sampling measure from ρ\rho to ρ~\widetilde{\rho} changes the measure underlying this weighted average. Consequently, the aggregated representation can change even for the same physical input and model parameters.

We correct this aggregation using ci=(dρ/dρ~)(xi)c_i=(\mathrm{d}\rho/\mathrm{d}\widetilde{\rho})(x_i):

s^j=∑iciexp⁡(ℓij)vi∑iciexp⁡(ℓij)=∑i[Softmax⁡N(ℓ⋅j+log⁡c)]ivi.\widehat{s}_j=\frac{\sum_i c_i\exp(\ell_{ij})v_i}{\sum_i c_i\exp(\ell_{ij})}=\sum_i\left[\operatorname{Softmax}_{N}(\ell_{\cdot j}+\log\mathbf{c})\right]_i v_i.

Thus, the correction is implemented simply by adding the log correction weights to the key logits before the spatial softmax. Specifically, we add log⁡ci\log c_i to the key logits at each spatial point ii. Writing the corrected spatial weights as AcA_c, we obtain

Ac=Softmax⁡N(Linear⁡(K)(HN)⊤+1M(log⁡c)⊤).A_c=\operatorname{Softmax}_{N}\left(\operatorname{Linear}^{(K)}(H_N)^{\top}+\mathbf{1}_M(\log\mathbf{c})^{\top}\right).

The resulting module is

LinearNO⁡c(HN)=ϕ(Q)[AcVN].\operatorname{LinearNO}_{c}(H_N)=\phi(Q)\left[A_c V_N\right].

We apply this correction at each spatial aggregation layer, leaving the learned parameters and query-side normalization unchanged. Under the usual support, integrability, and positive-denominator assumptions, the corrected aggregation converges to its reference-measure counterpart.

Results on Airfoil2D and ShapeNet-Car. We evaluate this extension on Airfoil2D and ShapeNet-Car, with the results shown below.

Airfoil2D: Field

BackbonePICPGCA
Transolver0.6166±0.01970.6166 \pm 0.01970.5831±0.02190.5831 \pm 0.02190.5579±0.01390.5579 \pm 0.0139
LinearNO0.680622±0.0308190.680622 \pm 0.0308190.559296±0.0210600.559296 \pm 0.0210600.589981±0.0259450.589981 \pm 0.025945

ShapeNet-Car: Field

BackbonePICPGCA
Transolver0.49792±0.001630.49792 \pm 0.001630.50064±0.001570.50064 \pm 0.001570.50063±0.001590.50063 \pm 0.00159
LinearNO0.574891±0.0024690.574891 \pm 0.0024690.531376±0.0025790.531376 \pm 0.0025790.531345±0.0026400.531345 \pm 0.002640

ShapeNet-Car: CDC_D

BackbonePICPGCA
Transolver0.63346±0.003280.63346 \pm 0.003280.63530±0.003960.63530 \pm 0.003960.63540±0.004400.63540 \pm 0.00440
LinearNO0.596994±0.0257660.596994 \pm 0.0257660.604585±0.0251920.604585 \pm 0.0251920.603715±0.0244370.603715 \pm 0.024437

Entries report mean ± standard deviation. Transolver results are taken from Tables 1 and 2 of the manuscript. Uniform is normalized to 1 for each backbone and metric. Lower is better.

All three policies improve on Uniform for both backbones. These results show that the accuracy–point-budget benefits of adaptive sampling extend to LinearNO.

2. Practical motivation for test-time sampling

Our motivation comes from repeated surrogate evaluations in inverse design, particularly for large 3D geometries. In this setting, training and inference may already operate on subsampled point clouds because processing every point incurs substantial computation and memory costs. Uniform subsampling provides a simple baseline, but it does not explicitly prioritize regions that are informative for the predicted field or design objective. The practical question is therefore how to allocate a limited spatial budget to improve prediction quality and the resulting design process.

Our existing inverse-design experiments provide application-level evidence. For 2D airfoil optimization, AdaSolver retains 18.75% of the points and achieves a 2.18× speedup for the complete design process, with essentially the same CFD-validated aerodynamic performance as full-mesh optimization. For 3D SuperWing optimization, it retains 37.5% of the points and achieves a 1.86× speedup, while attaining a slightly lower CFD-validated CD/CLC_D/C_L. These results connect adaptive sampling to a downstream computational benefit beyond forward-prediction accuracy.

We expect this question to become increasingly relevant as neural operators move toward engineering deployment. We will revise the Introduction and experimental discussion to make this application setting explicit, incorporate the additional comparisons, and state the scope of our contribution more precisely.

3. Effect of grid resolution on Darcy

We evaluated Darcy at resolutions of 85285^2, 1412141^2, and 2112211^2. All three policies improve upon Uniform at every resolution. However, the relative gains diminish: CPG and CA reduce rAUEC by approximately 36%, 11%, and 6.3%, respectively, across these resolutions.

MethodOriginal paper (85285^2)1412141^22112211^2
Uniform111
PI0.9965 ± 0.01770.9715 ± 0.00330.9699 ± 0.0127
CPG0.6419 ± 0.00490.8906 ± 0.01000.9368 ± 0.0046
CA0.6400 ± 0.00620.8904 ± 0.01000.9369 ± 0.0046

Darcy offers limited geometric cues, and denser grids may allow uniform sampling to capture most relevant information, reducing AdaSolver’s relative gains. This remains a challenging setting, while our CFD benchmarks with richer geometric structure show substantial benefits.

4. Full-field reconstruction

We agree that the reconstruction procedure was insufficiently described. During inference, each method predicts physical quantities at its selected points. To compare their accuracy fairly, we interpolate these predictions onto the same full reference mesh. The reconstructed fields are used to compute field errors or integrated aerodynamic quantities.

The reconstruction methods are:

  • Airfoil, Pipe, Plasticity, SuperWing, Darcy, and Elasticity: piecewise-linear interpolation within the convex hull of the retained points, with nearest-neighbor assignment outside the convex hull.
  • ShapeNet-Car: four-neighbor inverse-distance interpolation, using a distance-weighted average of the four nearest retained predictions.
  • DrivAerML: nearest-neighbor assignment, using the prediction at the nearest retained point.

Within each benchmark, all sampling methods use the same reconstruction procedure. We have added these details to the appendix.

5. Manuscript revisions

We have improved the manuscript’s writing and standardized the punctuation before equations throughout the paper.

References

[1] Hu, W., Liu, S., Qiao, P., Sun, Z., and Dou, Y. Transolver Is a Linear Transformer: Revisiting Physics-Attention Through the Lens of Linear Attention. arXiv:2511.06294.

Thank you for the response. Despite the considerable effort, the authors did not directly address the questions. In addition, the reviewer has further concerns about this work. The writing issues remain.

  1. Thank you for clarifying that the core motivation is to speed up inference. Therefore, it is necessary to compare AdaSolver with methods that are more efficient than Transolver at inference. However, the authors only added a comparison of LinearNO performance under the setting of this paper. There is neither a direct comparison of inference speed nor any performance result beyond rAUEC.
  2. Thank you for the additional experiments on Darcy. However, the authors only reported the results without answering the question that the reviewer is most concerned about: whether AdaSolver exhibits some consistent behavior across observation densities. The reviewer suggests that the authors describe and compare the behavior of AdaSolver across resolutions, rather than only comparing performance.
  3. Thank you for adding the description of reconstruction. However, the response is confusing. Why are different full-field reconstruction methods used on different benchmarks? This further leads to two additional questions: (1) Why is interpolation still applicable when the sampling ratio is only 37.5%? (2) Why do the authors not comprehensively report full-field prediction results on the benchmarks, but instead use indirect metrics such as rAUEC and relative improvements? Providing the original relative L2 errors would help address this issue.
  4. Thank you for only fixing Equation 4 and Equation 12. The reviewer is unclear why the authors did not consistently revise Equations 2, 3, 6, and 14, as well as the other equations. In addition, the reviewer is also unclear whether the outer box added to Equation G.4 has any special meaning. Given these issues, the reviewer’s confidence in this work has substantially decreased. Please manually check whether all equations and statements are correct.
Reviewer_FzuJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper describes a method for adaptive sampling for neural operators using a method they call AdaSolver. The core issue is that you'd like to be able to get finer resolution at some parts of the domain than others, but if you train for one sampling density and then change it at test time, the ratio between these densities can induce a bias. AdaSolver is supposed to prevent this.

Critical strengths

This paper addresses a key challenge of needing to vary the resolution in a way that depends on the system being emulated and its geometry. This adaptivity seems like a reasonable and desirable goal. The authors correctly note that there is bias associated with changing the sampling distribution without any correction using the new-to-old density ratio. The team assesses their method on a range of different problems, some at a large scale.

Critical weaknesses

Overall, the paper is poorly written, with notation not clearly defined, key concepts not clearly articulated, and generally it is unnecessarily difficult to read.

Figure 1 is hard to interpret. What should I get out of Figure 1a besides "transolver bad red X, AdaSolver good green check"? 1b introduces notation that is not used elsewhere in the paper, like W, T, t_j's. Some careful editing would make this paper more accessible to a wider audience.

Section 3.1 describes the Transolver as an integral operator before really describing the transolver, what it's supposed to do or how it's supposed to work. This makes the very notation- and equation-heavy section very difficult to parse. What is the role of C (dimension of feature encodings) vis-a-vis M (dimension of slice weights)? (1) introduces slice weights, only there are no slices involved. (2) introduces the s subscript to indicate slice s, but s only appears on the left hand side and not the right hand side. Where is the slice here?

Section 3.2 introduces slide domains \Omega_s which are never explained, and (5) now has a weight that depends on slice \xi_s which is never defined. Section 3.2 introduces a proposition and theorem that are neither proved nor sketched in the main body, but which the authors say points to the fact that not correcting for different sampling distributions at train and test times leads to bias. However, it's not clear why this isn't self evident. Doesn't the entire literature on distribution drift tell us this is a problem? Is it necessary to show that attention does not remove the bias? The significance of the theoretical results is unclear to me. The authors use reweighting via the density ratio (a common method in the face of distribution drift) to correct this bias, which is the right thing to do but not a major contribution.

Section 3.3 then moves to ways to select the test-time sampling distribution and three methods are proposed. It would have been helpful to see some discussion on what's happening here, what is the intuition behind the refinement score, and how it would be used. (This is important later when it comes to the empirical results.) Rather, the paper introduces a "refinement score" and some additional notation, then proposes three methods for setting the refinement score. The methods are pretty vague, like coarse-prediction-guided, which starts with a coarse estimate, then uses an unspecified "task dependent combination of coarse predictive features" as a score -- it's so vague that I have no sense of what was actually done in the empirics.

The empirical results are hard to parse. It would be great to have an overview of the strategy. The overall metric is based on the ratio of "test errors" for different methods, but since the different methods collect different samples at test time, what does "test error" refer to? Is there a common mesh used for assessing test error, or is "test error" for their sampling method assessed on their test samples and "test error" for uniform sampling assessed on the uniform test samples? This isn't clear, so I cannot assess the empirical results.

In addition, the overall setup is unclear. Why would one perform uniform subsampling at test time? I understand the basic premise of, for instance, wanting more samples near edges or singularities at test time, but I don't see why someone would uniformly SUBsample. Why is this an appropriate baseline? Is it to keep to a standard compute budget? If so, I don't fully understand the test time compute requirements and why subsampling is necessary once a neural operator is fully trained at the ful resolution.

Figure 3 is also hard to interpret. It would be great if the figures could be connected to the notation introduced in the paper. Is the leftmost column a, and the second column u? Figure 3 is never discussed in the body of the paper. It would be good to have an explanation of the plot where "optimization time" appears in the horizontal axis. Does this account for data generation and training time, or only inference time? It's not described.

Questions for authors

See questions integrated with weaknesses. All are clarification requests and not suggestions for future work.

2 comments

We thank the reviewer for the constructive feedback and are pleased that the reviewer considers our objective a “reasonable and desirable goal.”

1. Revisions to Figure 1

We have replaced “Transolver” and “AdaSolver” in Figure 1(a) with “Original aggregation” and “Corrected aggregation,” respectively. This makes our central finding more explicit: adapting spatial sampling density together with the corresponding aggregation weights can reduce the cost of neural operator evaluation while improving prediction accuracy. We have also revised the notation in Figure 1(b) to match the definitions and equations in the main text.

2. Architecture and notation in Section 3.1

We have rewritten Section 3.1 into two parts: “Revisiting Transolver” and “Transolver as an Integral Operator.”

The first part introduces the input embedding, stacked Transformer blocks with Physics-Attention, and output readout before presenting the discrete computations. We explicitly define slices as learned soft groups of points and distinguish CC, the feature dimension, from MM, the number of slices. We also clarify that ss denotes slice space, while jj indexes an individual slice through wi,jw_{i,j}.

The second part defines the continuous slice-space notation before use and explains how aggregation, token attention, and deslicing together approximate an integral operator.

3. Contribution of the analysis

We agree that density-ratio reweighting is a standard tool. Our analysis identifies and quantifies a specific failure mode within Physics-Attention: changing only the spatial sampling measure can change the limiting point-to-token aggregation, even for the same physical input and frozen network. The token-bias identity quantifies this discrepancy, while the attention analysis characterizes its propagation through attention weights and values. These results pinpoint the source of sampling dependence and justify a minimal correction to the aggregation step. The purpose is to enable adaptive spatial sampling while preserving the reference operator, rather than to introduce a new general theory of distribution shift.

4. Design principles for the refinement scores

We agree that the intuition behind the refinement score should be clearer. Appendix C provides the task-specific constructions. Under a fixed point budget, the scores prioritize three types of regions:

  • Geometric or input structure, such as walls, contact boundaries, coefficient interfaces, and changes in surface normals.
  • Rapid solution variation, indicated by gradients, local differences, or curvature in predicted pressure, velocity, stress, or displacement.
  • Contributions to the target quantity: for aerodynamic metrics, regions that strongly influence drag, lift, or their ratio.

This follows an established inverse-design workflow: a neural operator is constructed offline and then reused as a fast surrogate for repeated evaluations during optimization. Published examples include [4-6]. The offline cost can be shared across many subsequent design evaluations, making faster inference particularly valuable for repeated optimization tasks. Our timing comparison therefore measures the online benefit of adaptive sampling, with both methods using the same pretrained surrogate. We add figures in appendix D for illustration.

References

[1] Zhou, H., et al. Transolver-3: Scaling Up Transformer Solvers to Industrial-Scale Geometries. arXiv:2602.04940, 2026.

[2] Elrefaie, M., Morar, F., Dai, A., and Ahmed, F. DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks. NeurIPS Datasets and Benchmarks Track, 2024.

[3] Alkin, B., et al. Universal Physics Transformers: A Framework for Efficiently Scaling Neural Operators. NeurIPS, 2024.

[4] Li, Z., Huang, D. Z., Liu, B., and Anandkumar, A. Fourier Neural Operator with Learned Deformations for PDEs on General Geometries. Journal of Machine Learning Research, 24(388):1–26, 2023.

[5] Shukla, K., et al. Deep neural operators as accurate surrogates for shape optimization. Engineering Applications of Artificial Intelligence, 2024.

[6] Zhou, T., et al. AI-aided geometric design of anti-infection catheters. Science Advances, 10(1):eadj1741, 2024.

5. Test error on the reference mesh

AdaSolver and Uniform are evaluated on the same full reference mesh for each test case. The methods use different point sets for inference, but their predictions are reconstructed on the same full mesh before computing the error.

Specifically, let XrefX_{\mathrm{ref}} be the reference mesh and u^full\widehat u_{\mathrm{full}} the reconstructed prediction. For each evaluated field, we compute

e=∥u^full−uGT∥2,Xref∥uGT∥2,Xref.e = \frac{\lVert \widehat{u}_{\mathrm{full}} - u_{\mathrm{GT}} \rVert_{2,X_{\mathrm{ref}}}}{\lVert u_{\mathrm{GT}} \rVert_{2,X_{\mathrm{ref}}}}.

We average the per-case errors over the test set at each point budget. The reported rAUEC divides the area under the method’s mean error–budget curve by that of Uniform, using trapezoidal integration over the 37.5%, 50%, and 75% budgets. Thus, the ratio summarizes errors evaluated on a common reference mesh.

The appendix gives the reconstruction procedure for each dataset.

DatasetReconstruction method
Airfoil, Pipe, Plasticity, SuperWing, Darcy, ElasticityPiecewise-linear interpolation, with nearest-neighbor assignment outside the convex hull of the retained points.
ShapeNet-CarFour-neighbor inverse-distance interpolation.
DrivAerMLNearest-neighbor assignment.

6. Motivation and choice of baseline

Our motivation is fast, repeated evaluation during 3D shape design. High-resolution meshes can contain millions of points; our DrivAerML cases contain approximately 8.8 million surface points. Techniques such as caching in [1] enable large-mesh inference, but repeatedly evaluating full meshes during design optimization remains costly. AdaSolver studies how to reduce this cost by allocating a smaller point budget to informative regions while maintaining prediction quality.

Uniform sampling is an established choice for reduced-point representations and evaluation, as used in [2] and [3]. In our experiments, Uniform gives each reference mesh point the same selection probability. Comparing uniform and adaptive sampling with the same model and retained point count isolates the benefit of where points are allocated. Actual runtimes are reported separately because sampling-policy overheads differ, while full-mesh inference remains the 100% reference.

Our results show that adaptive allocation reduces prediction errors relative to Uniform at matched point budgets (Tables 1 and 2). In SuperWing optimization, AdaSolver uses 37.5% of the points and achieves a 1.86× speedup of the online optimization loop, while attaining a slightly lower CFD-validated CD/CLC_D/C_L than full-mesh optimization (Figure 4).

7. Revisions to Figures 3 and 4

We have improved the layouts and panel titles of Figures 3 and 4 and added explicit discussion in the main text. In Figure 3, the first two columns show CFD pressure fields for the initial and optimized geometries, corresponding to uCFD(a(0))u_{\mathrm{CFD}}(a^{(0)}) and uCFD(a∗)u_{\mathrm{CFD}}(a^{\ast}), respectively. Thus, both columns show solution fields for different geometries.

8. Definition of optimization time

“Optimization time” measures the online optimization process using the pretrained neural operator, including surrogate evaluations, geometry updates, sampling updates, and optimizer steps. It excludes data generation, model training, and independent CFD validation. We have clarified this definition in the captions. The CFD results assess design quality separately from the timed optimization process.

References

[1] Zhou, H., et al. Transolver-3: Scaling Up Transformer Solvers to Industrial-Scale Geometries. arXiv:2602.04940, 2026.

[2] Elrefaie, M., Morar, F., Dai, A., and Ahmed, F. DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks. NeurIPS Datasets and Benchmarks Track, 2024.

[3] Alkin, B., et al. Universal Physics Transformers: A Framework for Efficiently Scaling Neural Operators. NeurIPS, 2024.

[4] Li, Z., Huang, D. Z., Liu, B., and Anandkumar, A. Fourier Neural Operator with Learned Deformations for PDEs on General Geometries. Journal of Machine Learning Research, 24(388):1–26, 2023.

[5] Shukla, K., et al. Deep neural operators as accurate surrogates for shape optimization. Engineering Applications of Artificial Intelligence, 2024.

[6] Zhou, T., et al. AI-aided geometric design of anti-infection catheters. Science Advances, 10(1):eadj1741, 2024.

Reviewer_ww5qOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether a frozen Transolver-type neural operator can be evaluated on a non-uniform, test-time-chosen point set (a fixed budget concentrated near shocks, boundary layers, stress concentrations) without changing the operator it has learned. The key observation is that Physics-Attention forms each physics token as a slice-weighted average over the sampled points, so drawing points from a density $ \tilde\rho $ instead of the training density $ \rho $ changes the limit of this aggregation. Proposition 3.1 gives the token shift as $ Cov_{\pi}(\lambda,h)/E_{\pi}[\lambda] $ (with an operator-norm bound $ \sqrt{Var_{\pi}(\lambda)}/E_{\pi}[\lambda] $ ), and Theorem 3.2 describes how it propagates through continuous Physics-Attention. The proposed fix is a self-normalized inverse-density weighting in the point-to-slice aggregation ( $ \hat u_j = \sum_i c_i w_{ij} h_i / \sum_i c_i w_{ij} $ ), applied at every layer, with a tunable exponent $ \beta $ ( $ \beta = 0 $ recovers the original Transolver).

Three refinement policies decide where to put the points: PI (geometric/physical priors, no forward pass), CPG (a coarse forward pass on a small set $ S_0 $ , then gradient/Laplacian-based scores), and CA (a light head on first-block features that reuses cached $ S_0 $ features). Evaluation covers five standard benchmarks (Airfoil, Pipe, Darcy, Elasticity, Plasticity) and three industrial ones (ShapeNet-Car, DrivAerML, SuperWing) at 37.5/50/75% budgets, against uniform subsampling, using the relative area under the error–budget curve (rAUEC, Uniform = 1) over 3 seeds. Reported rAUECs for CPG/CA include 0.47–0.64 on Elasticity/Darcy, 0.56–0.58 on Airfoil fields, and 0.67 (CPG) on SuperWing fields; design quantities such as $ C_D $ improve less ( $ \approx $ 0.65–0.92). A fixed-points ablation over $ \beta $ (Table 6) shows that dropping the correction can make results much worse than Uniform (e.g. Pipe CPG 0.67 $ \to $ 11.1). Latency is reported in Table 7 (e.g. CA is $ 1.95\times $ / $ 2.61\times $ / $ 2.04\times $ faster than full mesh at 37.5% on ShapeNet-Car/DrivAerML/SuperWing), and two surrogate-based shape optimizations (2D airfoil with SLSQP, 3D wing with CMA-ES) are $ 2.18\times $ and $ 1.86\times $ faster with CFD-verified $ C_D/C_L $ close to the full-mesh result.

Critical strengths
  1. The paper identifies exactly where sampling density enters a Transolver (the point-to-token aggregation; slice assignment, token attention and deslicing are pointwise or token-level), derives the resulting bias, and fixes only that step with no new parameters. The most persuasive evidence is the controlled experiment in Appendix D.1 / Table 6, where checkpoints and sampled points are held fixed and only $ \beta $ is changed: on Darcy the CPG and CA rAUECs go from 0.642/0.640 to 4.28/4.27 without correction, and on SuperWing the correction reduces the field error–budget area by 66–69%. This directly supports the central claim that naive non-uniform sampling harms a frozen operator, and that the harm is repairable. This is what makes the idea useful beyond the specific policies.
  2. Beyond standard PDE benchmarks, the paper evaluates design-relevant quantities ( $ C_D $ , $ C_L $ , $ C_D/C_L $ ) on million-point industrial data (including DrivAerML), reports measured latency per budget for every policy (Table 7, including cases where the method is not faster), and checks two surrogate-based optimizations with independent CFD. Reporting the unfavourable timing cases (e.g. Elasticity) is appreciated and makes the efficiency claims more credible.
Critical weaknesses
  1. All main results compare against Uniform subsampling only (plus the $ \beta = 0 $ ablation), which is a natural but weak reference: it does not show that the proposed policies beat other reasonable ways to place points or other ways to compensate for the measure shift. More importantly, the headline "matches or surpasses full-mesh accuracy with ~one third of the points" statements (Section 4.3; Appendix D.2: Plasticity 5.02% vs 6.40%, DrivAerML $ C_D $ 0.948% vs 1.05%) are measured against a separately trained full-mesh Transolver reference, whereas AdaSolver uses a backbone trained with variable point counts ( $ N_t \in [0.4N, N] $ ) for four of the five standard benchmarks, and a different pressure checkpoint for DrivAerML. The gap may therefore partly reflect the training protocol (sub-sampling as augmentation) rather than adaptive test-time sampling. Figure 6 normalizes by the shared backbone's full-mesh error, but the text never states whether the adaptive policies beat that reference. Finally, the "3 seeds" in Tables 1, 2 and 6 appear to be sampling seeds with a single trained backbone and refinement head, so the reported variance does not include training variability; some entries are very noisy (Airfoil $ C_D/C_L $ : PI 0.2174 $ \pm $ 0.1688), and several differences are tiny (Plasticity 0.889/0.889/0.890).
  2. The correction is exactly consistent only at $ \beta = 1 $ , but the deployed policies use policy-specific exponents $ \beta^\star $ "specified during policy design" (e.g. 0.75 for ShapeNet-Car $ C_D $ ), and the paper does not explain how $ \beta^\star $ is chosen, on what data, or how much per-task tuning it requires; the ablation itself shows that the best result occurs when $ \beta $ matches the policy's own density construction. On efficiency, most of the speed-up comes from evaluating fewer points rather than from adaptivity: at equal budgets, CPG and CA are slower than Uniform (Elasticity at 37.5%: Uniform 0.99 ms, CPG 2.09 ms, CA 1.52 ms vs 1.26 ms full mesh, so CA is slower than the full mesh there), and the Elasticity timing uses a special configuration (25% coarse set, two-block head). Latency excludes full-field reconstruction, while accuracy (as I understand the protocol) is measured on the reconstructed field. The inverse-design experiments, the main end-to-end evidence, are single runs, compare only with the full mesh (not with Uniform at the same budget), and report CFD-verified differences that are very small (0.05897 vs 0.05879; 0.04275 vs 0.04283), so they show "no loss" rather than a benefit of adaptivity. The 2D case also uses an 18.75% budget, below the 37.5–75% range analysed in the main tables.

Further points that are not decisive: the use of inverse-density / self-normalized weighting is established (Monte Carlo convolution, QuadConv), so the novelty lies in the application to physics-token aggregation and the bias analysis; the conclusions are tied to the structure of Transolver-type models and generalization to other operators is stated but not tested; PI requires hand-designed priors per task, CA requires a task-specific trained head, and Appendix C contains many task-specific density constructions.

Questions for authors
  1. How is $ \beta^\star $ chosen for each policy and task (which split, which search, how many values tried)? How much does the final accuracy depend on this choice if $ \beta^\star $ is selected on validation data and then fixed across tasks (e.g. $ \beta = 1 $ or a single default)?
  2. At 37.5% and 50% budgets, do CPG/CA outperform the shared variable-point backbone evaluated on the full mesh (the horizontal line in Fig. 6)? Please report this explicitly, since it separates the effect of adaptive sampling from that of the training protocol.
  3. For the inverse-design experiments, how do Uniform sampling at the same budget (18.75% / 37.5%) and multiple initializations/runs compare? Also, what accuracy do the policies reach at 18.75% on the airfoil task?
  4. Does the end-to-end latency in Table 7 include the reconstruction onto the full mesh that is used in the error computation? What is the total cost when it is included?
4 comments

We sincerely thank the reviewer for their valuable feedback and constructive suggestions, which have helped us improve the clarity, accuracy, and completeness of the paper.

1. (W1 and Q1) Effect of the training strategy

We thank the reviewer for highlighting this confounding factor. The full-mesh-trained Transolver reference was included to illustrate sensitivity to changing point counts at inference. We agree that this comparison alone does not isolate the benefit of adaptive sampling.

To clarify, we report full-mesh inference errors for both training protocols:

BenchmarkMetricOriginal TransolverVariable-point-count Transolver
Airfoil2DPressure0.003870320.00511753
PipeField0.004315660.00483350
DarcyField0.004879370.00532525
ElasticityField0.006739200.00668239
PlasticityField0.063991770.05243651
ShapeNet-CarField0.062474090.06170939
DrivAerMLPressure0.059879840.05522662
SuperWingField0.017161190.01762329
Airfoil2DCD/CLC_D/C_L absolute error0.028133940.01521042
ShapeNet-CarCDC_D relative error0.012850360.01398623
DrivAerMLCDC_D relative error0.010495470.00950459
SuperWingCD/CLC_D/C_L relative error0.005228240.00540848

Both models are evaluated using the full mesh. Bold indicates the lower error in each row.

Variable-point-count training improves robustness to subsampling but does not consistently improve full-mesh accuracy. Preserving full-mesh accuracy while improving robustness across point counts therefore remains worth investigating.

Nevertheless, on Plasticity, CPG achieves 5.02% relative L2L_2 error using 37.5% of the points, outperforming both the separately trained full-mesh reference (6.40%) and full-mesh inference of its own backbone (5.244%). On DrivAerML, CPG likewise achieves a lower CDC_D error (0.948%) than both full-mesh references (1.05% and 0.9505%), using 37.5% of the points.

2. (W1 and Q2) Effect of the training seed

We conducted additional evaluations using three independent training seeds. Due to time constraints, we have not yet completed all experiments. The newly available standard-benchmark results are reported below:

BenchmarkMethodOriginal Table 1New evaluation (3 training seeds)
ElasticityPI0.7280±0.00830.7280 \pm 0.00830.726406±0.0021680.726406 \pm 0.002168
ElasticityCPG0.4681±0.00410.4681 \pm 0.00410.467020±0.0011750.467020 \pm 0.001175
ElasticityCA0.5469±0.00210.5469 \pm 0.00210.560134±0.0106470.560134 \pm 0.010647
PipePI0.8711±0.00610.8711 \pm 0.00610.855711±0.0126980.855711 \pm 0.012698
PipeCPG0.6719±0.00980.6719 \pm 0.00980.654441±0.0101130.654441 \pm 0.010113
PipeCA0.7051±0.01160.7051 \pm 0.01160.692265±0.0093940.692265 \pm 0.009394
DarcyPI0.9965±0.01770.9965 \pm 0.01770.995973±0.0028410.995973 \pm 0.002841
DarcyCPG0.6419±0.00490.6419 \pm 0.00490.639290±0.0158710.639290 \pm 0.015871
DarcyCA0.6400±0.00620.6400 \pm 0.00620.639412±0.0164050.639412 \pm 0.016405

Entries report mean ± standard deviation.

The results show limited variation across training seeds for the evaluated benchmarks and policies, supporting the robustness of the improved error–budget trade-off achieved by test-time adaptive sampling. We will incorporate the additional multi-seed results into the camera-ready version.

3. (W2 and Q4) Cost of full-field reconstruction

Evaluation on a common reference mesh. We recognize that our evaluation protocol differs from evaluating field errors directly at the points supplied to the model, as in the ShapeNet-Car evaluation of Transolver [1]. DrivAerNet++ also distinguishes prediction errors at sampled points from errors after interpolation onto the full mesh [2]. In our test-time adaptive sampling setting, AdaSolver and Uniform generally retain different point sets. Evaluating each method only at its retained points would therefore measure errors on different spatial supports. We introduce full-field reconstruction into our evaluation to compare all sampling policies on the same reference mesh. Using the same reconstruction rule within each benchmark ensures a consistent comparison.

Post-processing for inverse design. Inverse-design objectives do not necessarily require full-field reconstruction. Our 2D airfoil objective requires only surface pressure; similarly, car-shape objectives such as CDC_D can be computed from surface pressure and shear, without reconstructing the full 3D volume field. In both cases, the required post-processing takes less than 1% of the model forward-pass time.

Our inverse-design approach also exploits the gradual evolution of geometry to reduce the frequency of sampling updates. For example, SuperWing updates the sampling density every 20 generations. Its reconstruction uses a fixed index lattice, so the interpolation map can be cached while retained indices remain unchanged, even as the physical geometry evolves. Each prediction then requires only applying the cached weights, which takes 1% of the model forward-pass time.

Reconstruction procedures. For full-field evaluation, we use:

  • Airfoil, Pipe, Plasticity, SuperWing, Darcy, and Elasticity: piecewise-linear interpolation, with nearest-neighbor assignment outside the convex hull of the retained points.
  • ShapeNet-Car: four-neighbor inverse-distance interpolation.
  • DrivAerML: nearest-neighbor assignment.

These procedures require neither ground-truth fields nor an additional learned reconstruction model.

4. (W2 and Q3) Statistics of inverse-design experiments

We added a multi-seed evaluation of the 2D airfoil inverse-design experiment in Appendix F.

Initial CFD CD/CLC_D/C_LFull CFD CD/CLC_D/C_LUniform CFD CD/CLC_D/C_LAdaSolver CFD CD/CLC_D/C_L
0.6053670.0587520.060725±0.0018140.060725 \pm 0.001814 (+3.36%)0.058621±0.0000740.058621 \pm 0.000074 (−0.22%)
0.2466920.0587920.061417±0.0006140.061417 \pm 0.000614 (+4.46%)0.058696±0.0001640.058696 \pm 0.000164 (−0.16%)
0.1465870.0587480.061688±0.0001170.061688 \pm 0.000117 (+5.00%)0.058601±0.0000340.058601 \pm 0.000034 (−0.25%)

Uniform and AdaSolver entries report mean ± standard deviation. Percentages in parentheses indicate relative changes from the corresponding Full CFD value. Lower is better.

Across multiple initial geometries, AdaSolver achieves better aerodynamic performance on average and shorter optimization times than full-mesh inference, whereas uniform sampling yields designs with poorer aerodynamic performance.

The 2D airfoil task is a relatively simple benchmark, and we acknowledge that the interpolation used to construct the sampling density incurs non-negligible overhead (see Table 7). Nevertheless, with an appropriate density design, AdaSolver achieves a speedup of more than 2×2\times in inverse design while retaining only 18.75% of the points. This overhead is expected to account for a smaller fraction of the total runtime in larger, more realistic 3D CFD design applications.

5. (W2) Unified configuration

We have also updated the Elasticity results using a unified configuration. The updated results preserve our conclusions: adaptive test-time sampling improves the error–budget trade-off. We have updated the corresponding tables and descriptions in the manuscript.

MethodOriginal (25% coarse)Unified (10% coarse)
Uniform11
PI0.7280±0.00830.7280 \pm 0.00830.7264±0.00220.7264 \pm 0.0022
CPG0.4681±0.00410.4681 \pm 0.00410.4670±0.00120.4670 \pm 0.0012
CA0.5469±0.00210.5469 \pm 0.00210.5601±0.01060.5601 \pm 0.0106

Entries report rAUEC (↓\downarrow) on Elasticity as mean ± standard deviation.

6. (W3) Results with another Transformer neural operator

Our density-design approach to test-time adaptive sampling is a general strategy for neural operators with spatial aggregation. To evaluate its applicability beyond Transolver, we extend AdaSolver to LinearNO [3] as an additional backbone.

Measure correction in LinearNO. In LinearNO, the output is computed as ϕ(Q)(ψ⊤(K)VN)\phi(Q)(\psi^{\top}(K)V_N), where VN=Linear⁡(V)(HN)V_N=\operatorname{Linear}^{(V)}(H_N). The aggregation ψ⊤(K)VN\psi^{\top}(K)V_N uses a softmax over the spatial points. Writing ℓij=[Linear⁡(K)(hi)]j\ell_{ij}=[\operatorname{Linear}^{(K)}(h_i)]_j, each aggregated representation is

sj=∑i=1Nexp⁡(ℓij)vi∑i=1Nexp⁡(ℓij).s_j=\frac{\sum_{i=1}^{N}\exp(\ell_{ij})v_i}{\sum_{i=1}^{N}\exp(\ell_{ij})}.

Changing the spatial sampling measure from ρ\rho to ρ~\widetilde{\rho} changes the measure underlying this weighted average. Consequently, the aggregated representation can change even for the same physical input and model parameters.

We correct this aggregation using ci=(dρ/dρ~)(xi)c_i=(\mathrm{d}\rho/\mathrm{d}\widetilde{\rho})(x_i):

s^j=∑iciexp⁡(ℓij)vi∑iciexp⁡(ℓij)=∑i[Softmax⁡N(ℓ⋅j+log⁡c)]ivi.\widehat{s}_j=\frac{\sum_i c_i\exp(\ell_{ij})v_i}{\sum_i c_i\exp(\ell_{ij})}=\sum_i\left[\operatorname{Softmax}_{N}(\ell_{\cdot j}+\log\mathbf{c})\right]_i v_i.

Thus, the correction is implemented simply by adding the log correction weights to the key logits before the spatial softmax. Specifically, we add log⁡ci\log c_i to the key logits at each spatial point ii. Writing the corrected spatial weights as AcA_c, we obtain

Ac=Softmax⁡N(Linear⁡(K)(HN)⊤+1M(log⁡c)⊤).A_c=\operatorname{Softmax}_{N}\left(\operatorname{Linear}^{(K)}(H_N)^{\top}+\mathbf{1}_M(\log\mathbf{c})^{\top}\right).

The resulting module is

LinearNO⁡c(HN)=ϕ(Q)[AcVN].\operatorname{LinearNO}_{c}(H_N)=\phi(Q)\left[A_c V_N\right].

We apply this correction at each spatial aggregation layer, leaving the learned parameters and query-side normalization unchanged. Under the usual support, integrability, and positive-denominator assumptions, the corrected aggregation converges to its reference-measure counterpart.

Results on Airfoil2D and ShapeNet-Car. We evaluate this extension on Airfoil2D and ShapeNet-Car, with the results shown below.

Airfoil2D: Field

BackbonePICPGCA
Transolver0.6166±0.01970.6166 \pm 0.01970.5831±0.02190.5831 \pm 0.02190.5579±0.01390.5579 \pm 0.0139
LinearNO0.680622±0.0308190.680622 \pm 0.0308190.559296±0.0210600.559296 \pm 0.0210600.589981±0.0259450.589981 \pm 0.025945

ShapeNet-Car: Field

BackbonePICPGCA
Transolver0.49792±0.001630.49792 \pm 0.001630.50064±0.001570.50064 \pm 0.001570.50063±0.001590.50063 \pm 0.00159
LinearNO0.574891±0.0024690.574891 \pm 0.0024690.531376±0.0025790.531376 \pm 0.0025790.531345±0.0026400.531345 \pm 0.002640

ShapeNet-Car: CDC_D

BackbonePICPGCA
Transolver0.63346±0.003280.63346 \pm 0.003280.63530±0.003960.63530 \pm 0.003960.63540±0.004400.63540 \pm 0.00440
LinearNO0.596994±0.0257660.596994 \pm 0.0257660.604585±0.0251920.604585 \pm 0.0251920.603715±0.0244370.603715 \pm 0.024437

Entries report mean ± standard deviation. Transolver results are taken from Tables 1 and 2 of the manuscript. Uniform is normalized to 1 for each backbone and metric. Lower is better.

All three policies improve on Uniform for both backbones. These results show that the accuracy–point-budget benefits of adaptive sampling extend to LinearNO.

7. (Q1) Design principles of the sampling policies

In the original experiments, the density-design parameters were selected by grid search on the validation set. For β\beta, we considered exactly five values: {0,0.25,0.5,0.75,1}\{0, 0.25, 0.5, 0.75, 1\}. Appendix D reports the corresponding ablation results.

Following Reviewer oHtv’s suggestion, we revised the refinement-score constructions. With the revised scores, β=1\beta=1 was the best validation setting across all evaluated tasks and policies. We therefore use a single default, β=1\beta=1, throughout the revised experiments, fixed before test evaluation. This corresponds to full inverse-density compensation in our measure-consistent formulation. The revised results consequently do not require task-specific tuning of β\beta. We will update Appendix D with the corresponding sensitivity results for the revised score designs.

References

[1] Wu, H., Luo, H., Wang, H., Wang, J., and Long, M. Transolver: A Fast Transformer Solver for PDEs on General Geometries. ICML, 2024.

[2] Elrefaie, M., Morar, F., Dai, A., and Ahmed, F. DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks. NeurIPS Datasets and Benchmarks Track, 2024. Appendix I.1, Table 10.

[3] Hu, W., Liu, S., Qiao, P., Sun, Z., and Dou, Y. Transolver Is a Linear Transformer: Revisiting Physics-Attention Through the Lens of Linear Attention. arXiv:2511.06294.

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

AdaSolver adapts spatial sampling at inference while keeping the neural operator frozen. It corrects Transolver aggregation for sampling-density changes and combines physical priors, coarse predictions, or learned refinement. Experiments span eight PDE/aerodynamic datasets and CFD-validated design examples.

Critical strengths
  1. The sampling-measure analysis identifies an important failure mode of adaptive operator inference. Fixed-point-set ablations provide useful evidence that weighting itself improves predictions.

  2. Evaluation spans diverse tasks and includes independent CFD verification of optimized designs, strengthening the practical motivation.

Critical weaknesses
  1. The reference-measure consistency claim does not directly cover the deployed policies. The derivation uses inverse-density weighting, but several settings use β<1 and approximate coverage weights with farthest-point sampling. Empirical improvements do not establish restoration of the reference operator (Eq. 15; Appendices C and D.1).

  2. Reported timing excludes final full-field reconstruction, and the Elasticity timing uses a different refinement-head configuration from the accuracy setup (Tables 5 and 7). Consequently, the accuracy–speed trade-off is not yet established for the complete output pipeline.

Questions for authors
  1. Clarification: Is β<1 intended as a biased finite-sample trade-off? How do coverage weights correspond to the actual sampling mechanism?

  2. Assessment-changing evidence: What are the accuracy–latency results with reconstruction included and identical model, sampling, and refinement configurations?

2 comments

We thank the reviewer for the constructive feedback. We are pleased that the reviewer finds our approach practical and acknowledges that our results provide useful evidence that weighting itself improves predictions.

Below, we provide detailed responses to the concerns raised.

1. (W1 and Q1) Consistency of the sampling policy

We agree that the reference-measure consistency argument applies to exact inverse-density correction with β=1\beta=1, under the stated sampling and integrability assumptions.

Our refinement scores use physical or geometric priors, with parameters selected by grid search. At finite budgets, adaptive-sampling accuracy depends on both whether the density allocates samples to informative regions and whether the aggregation appropriately compensates for the density shift.

Specifically, let λ=dρ~/dρ\lambda=\mathrm{d}\widetilde{\rho}/\mathrm{d}\rho denote the actual sampling-density ratio. With exact weights c=λ−1c=\lambda^{-1}, using cβc^\beta leads to the population aggregation

us,j(β)=∫wjh λ1−β dρ∫wj λ1−β dρ.u_{s,j}^{(\beta)}=\frac{\int w_j h\,\lambda^{1-\beta}\,\mathrm{d}\rho}{\int w_j\,\lambda^{1-\beta}\,\mathrm{d}\rho}.

Thus, the effective measure satisfies dρβ∝λ1−β dρ\mathrm{d}\rho_\beta \propto \lambda^{1-\beta}\,\mathrm{d}\rho. Equivalently, its density is proportional to pβq1−βp^\beta q^{1-\beta}, where pp and qq are the reference and sampling densities, respectively. Tempered correction therefore geometrically interpolates between these measures and generally retains residual sampling bias. We will describe it as empirical tempering rather than exact reference-measure correction.

The non-unit optima for some Elasticity and ShapeNet-Car policies motivated us to test less concentrated densities with full correction. Tempering may reduce sensitivity to large weights in sparsely sampled regions, but retains residual measure bias; its benefit alone does not establish that the original densities were overly concentrated.

We have revised our density design accordingly. The revised densities yield improved accuracy, with β=1\beta=1 achieving the lowest mean error among the tested exponents and outperforming the original deployed configurations.

Dataset / targetPolicyOriginal configurationRevised (β=0\beta=0)Revised (β=0.25\beta=0.25)
ElasticityPI0.728043 (β=0.5\beta=0.5)0.7313860.728945
ShapeNet-Car / FieldPI0.589515 (β=0.5\beta=0.5)0.6300770.575146
ShapeNet-Car / FieldCPG0.557696 (β=1\beta=1)0.6205960.570494
ShapeNet-Car / FieldCA0.557738 (β=1\beta=1)0.6206250.570544
ShapeNet-Car / CdC_dPI0.656056 (β=0.5\beta=0.5)1.1398571.029949
ShapeNet-Car / CdC_dCPG0.647430 (β=0.75\beta=0.75)1.1450691.035314
ShapeNet-Car / CdC_dCA0.647856 (β=0.75\beta=0.75)1.1452921.035563

For some cases, we use region-wise constant sampling densities to reduce computational cost and apply farthest-point sampling within each region, which we empirically found to provide stable, approximately uniform spatial coverage. We will clarify in Appendix C that this is a practical sampling heuristic.

References

[1] Wu, H., Luo, H., Wang, H., Wang, J., and Long, M. Transolver: A Fast Transformer Solver for PDEs on General Geometries. ICML, 2024.

[2] Elrefaie, M., Morar, F., Dai, A., and Ahmed, F. DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks. NeurIPS Datasets and Benchmarks Track, 2024. Appendix I.1, Table 10.

2. (W2 and Q2) Cost of full-field reconstruction

Evaluation on a common reference mesh. We recognize that our evaluation protocol differs from evaluating field errors directly at the points supplied to the model, as in the ShapeNet-Car evaluation of Transolver [1]. DrivAerNet++ also distinguishes prediction errors at sampled points from errors after interpolation onto the full mesh [2]. In our test-time adaptive sampling setting, AdaSolver and Uniform generally retain different point sets. Evaluating each method only at its retained points would therefore measure errors on different spatial supports. We introduce full-field reconstruction into our evaluation to compare all sampling policies on the same reference mesh. Using the same reconstruction rule within each benchmark ensures a consistent comparison.

Post-processing for inverse design. Inverse-design objectives do not necessarily require full-field reconstruction. Our 2D airfoil objective requires only surface pressure; similarly, car-shape objectives such as CdC_d can be computed from surface pressure and shear, without reconstructing the full 3D volume field. In both cases, the required post-processing takes less than 1% of the model forward-pass time.

Our inverse-design approach also exploits the gradual evolution of geometry to reduce the frequency of sampling updates. For example, SuperWing updates the sampling density every 20 generations. Its reconstruction uses a fixed index lattice, so the interpolation map can be cached while retained indices remain unchanged, even as the physical geometry evolves. Each prediction then requires only applying the cached weights, which takes 1% of the model forward-pass time.

Reconstruction procedures. For full-field evaluation, we use:

  • Airfoil, Pipe, Plasticity, SuperWing, Darcy, and Elasticity: piecewise-linear interpolation, with nearest-neighbor assignment outside the convex hull of the retained points.
  • ShapeNet-Car: four-neighbor inverse-distance interpolation.
  • DrivAerML: nearest-neighbor assignment.

These procedures require neither ground-truth fields nor an additional learned reconstruction model.

3. (W2) Unified configuration

We have also updated the Elasticity results using a unified configuration. The updated results preserve our conclusions: adaptive test-time sampling improves the error–budget trade-off. We have updated the corresponding tables and descriptions in the manuscript Table 1.

MethodOriginal (25% coarse)Unified (10% coarse)
Uniform11
PI0.7280±0.00830.7280 \pm 0.00830.7264±0.00220.7264 \pm 0.0022
CPG0.4681±0.00410.4681 \pm 0.00410.4670±0.00120.4670 \pm 0.0012
CA0.5469±0.00210.5469 \pm 0.00210.5601±0.01060.5601 \pm 0.0106

Entries report rAUEC (↓\downarrow) on Elasticity as mean ± standard deviation.

References

[1] Wu, H., Luo, H., Wang, H., Wang, J., and Long, M. [Transolver: A Fast Transformer Solver for PDEs on General Geometries]. ICML, 2024.

[2] Elrefaie, M., Morar, F., Dai, A., and Ahmed, F. [DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks]. NeurIPS Datasets and Benchmarks Track, 2024. Appendix I.1, Table 10.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.