acceptodds
Under review as a conference paper at ICLR 2027

Physics-Aware Constraints for Inverse Shape Design

Abstract

Neural surrogates now carry computer aided engineering (CAE) into inverse shape design, where optimization descends the gradients of a frozen model. However, the optimizer and the judge are the same network: the search freely leaves the region where the surrogate has evidence, claiming gains that no data can verify. Today this trust gap is bridged by hand, with engineers specifying for every task where a shape may deform and how far it may drift. We present Physics-Aware Constraints (PAC), the first supervision layer to read both constraints from the frozen GeoPT, pretrained on geometry lifted with synthetic dynamics. The core insight is that a model pretrained to perceive geometry through synthetic physical response should supervise the search rather than drive it. The model’s weights bound how far optimization may drift off the family response manifold, and its distance field bounds where deformation is admissible, both calibrated on the family without labels. Across four CAE families, PAC consistently keeps optimization inside the trust band at little cost to legitimate gains, delivering on ship hulls a 64% wave drag reduction verified by an analytic oracle. The results separate the two sources of the constraint: its scale selectivity comes from the distance field representation and its detection sensitivity from physics lifted pretraining. PAC performs the judgment of a delivered design that autonomous design agents in industry still leave to an engineer, moving their full automation a step closer.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Ai4CAE, AI4Physics, Optimization

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_SrzpOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Physics-Aware Constraints, a framework for preventing neural surrogate models from exploiting inaccurate predictions during inverse shape optimization. The authors argue that previous methods tend to push geometries far outside the training distribution, producing visually unrealistic designs. PAC addresses this using a frozen GeoPT to introduce two complementary constraints, a response manifold constraint and a deformability field. Both constraints are calibrated without GT labels and can be integrated into different optimization frameworks. Experiments across four engineering design categories show that PAC reduces out-of-distribution optimization and preserves meaningful performance improvements.

Critical strengths
  • The authors highlight the imbalance issue in previous shape optimization methods, where the surrogate models significantly optimize the physical properties while leading to visually unrealistic designs.
  • The authors propose to separate optimization and supervision by assigning them to two different frozen models, a task-specific physics surrogate and a pretrained physics-geometry model.
  • The two introduced constraints operate at different levels. The response fingerprint measures global distance from the object, and the deformability field controls local modifications. This makes PAC applicable to multiple shape parameterizations.
  • Experiments show that PAC yields visually realistic shape designs and also improves the physical attributes.
Critical weaknesses
  1. My major concern is the paper's positioning relative to prior work. Several existing methods [1] have already explored incorporating physical constraints into shape generation and optimization. For instance, PhysGen regularizes the flow-matching process using physical constraints, such as drag and surface pressure, to generate physically plausible and visually realistic shapes. The advantages of the proposed PAC framework over such methods are unclear. A more comprehensive discussion of related approaches, along with experimental comparisons, would help better establish the novelty and benefits of the proposed method.
  2. The proposed method introduces additional computational overhead, but the paper does not provide a corresponding efficiency analysis. PAC requires additional GeoPT forward passes, backward computations, fingerprint evaluations, etc. A detailed runtime and computational cost analysis, including comparisons with unconstrained optimization, would be important for assessing the practicality and scalability of the proposed framework.
  3. PAC heavily relies on the frozen pretrained GeoPT model. However, GeoPT may produce unreliable representations in some scenarios. In such cases, PAC could become ineffective or overly restrictive, potentially undermining its reliability. The paper currently lacks an evaluation of the framework's robustness to such failures.

[1] Yingxuan You, Chen Zhao, Hantao Zhang, Ming Xu, and Pascal Fua. "Physgen: Physically grounded 3d shape generation for industrial design." In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 27209-27218. 2026.

Questions for authors

Please refer to the weaknesses I listed

4 comments
AuthorsOpenReview ↗

Response to Reviewer Srzp

We sincerely thank the reviewer for the detailed and constructive feedback, and for recognizing the separation of optimization from supervision and the two levels at which our constraints operate. We address each point below. New and changed text in the revised manuscript is shown in blue.

W1: Positioning relative to prior work such as PhysGen, and experimental comparison.

We thank the reviewer for this valuable suggestion. We have added PhysGen and related physics guided generators to Sec. 2, and a positioning table to Appendix G.1.

PhysGen and PAC address different stages of design, and they are complementary rather than competing. PhysGen is a generator. Its flow matching sampling alternates with a refinement step that moves the latent code along the gradients of a drag decoder and of forces integrated from a pressure decoder, both trained on CFD labels. In the terms of our paper, this refinement is an instance of the implicit path (Sec. 4.3), in which the design is the latent code of a learned shape model.

PAC is not another design path. It is a general supervisor for the loop that produces a design, whether that loop is explicit, implicit or generative. It measures whether the design still lies where the predictor that steers the loop has evidence, using a separate frozen model and no labels. Our goal is to automate the step of AI for CAE that still depends on an engineer, namely checking whether an optimized design can be trusted before it is sent to simulation, so that the design loop can run without manual supervision.

PhysGen (You et al., 2026)PAC (Ours)
Rolegenerates shapessupervises an optimizer or a generator
Physics signallearned drag and pressure decoders, guiding the latentnever guides; only bounds the search
Plausibility fromthe flow matching priorthe family response manifold and deformability field
LabelsCFD drag and pressure, for trainingnone
Support of the guiding predictorchecked afterwards by CFD on samplesmeasured at every design, without labels
Parametrizationsits own latent spaceFFD, neural implicit, design vector, generated population

Two of our results relate directly to this distinction:

  • A plausible shape does not guarantee a reliable prediction. On the blended wing body, every iterate of the unconstrained neural implicit path is decoded by a neural implicit model trained on the family and is moved by the gradient of a learned predictor through the latent code. This is the same mechanism as the latent refinement of a physics guided generator. Under independent CFD (Sec. 5.1, Table 4), the iterates beyond twice our boundary are predicted to have 20 to 27% lower drag, while the body volume nearly doubles and the measured drag is 14% higher. The fence identifies these iterates, although each of them is still a valid decoded shape.
  • PAC can be combined with a generator. On the ship hull, PAC filters a population sampled from a diffusion generator trained on the family (Figure 7). With 8 oracle calls it reaches a 56.9× gain in wave drag over the reference, compared with 7.9 to 9.7× for distances in the design parameter space (Sec. 5.1, Table 5).

Since PhysGen's latent refinement is itself a differentiable inverse loop, PAC's two constraints can be added to it as penalties on each latent update, in the same way as on our neural implicit path. We discuss this combination in Appendix G.1.

AuthorsOpenReview ↗

Response to Reviewer Srzp

W2: Computational overhead and runtime analysis, compared with unconstrained optimization.

We thank the reviewer for raising this important practical point. We measured the wall clock time and peak memory per optimization step with and without PAC, on one RTX 5090 with the same seeds and settings and with all diagnostics excluded.

PathUnconstrained (s/step)PAC (s/step)Added per stepAdded per runPeak memory (GiB)
Car, FFD, 120 steps0.0310.131+0.10 s+12.8 s0.38 → 1.21
Blended wing body, neural implicit, 180 steps6.236.38+0.15 s+18.2 s0.61 → 1.59
Ship hull, scoring a generated hull0.47 s per hull
One time calibration per family12 s (car), 73 s (BWB), 80 s (ship)

PAC adds one forward and one backward pass through the frozen model to each step. This is 0.10 s per step on the car and 0.15 s on the blended wing body, and less than 20 s over a full run. On the neural implicit path, where decoding the shape dominates the step, the overhead is 2.4%. Calibration is performed once per family and is reused by every task on that family. For comparison, a single CFD verification costs 170 to 262 core hours for a car on our 6.4M-cell mesh, 1,030 to 1,130 on the 24.8M-cell mesh, and 45 to 110 for a blended wing body. These results are summarized in Sec. 5.2 and detailed in Appendix G.2, Table 16 (in blue).

W3: GeoPT may produce unreliable representations, making PAC ineffective or overly restrictive.

We thank the reviewer for this insightful question. We agree that the supervisor should fail safely, and we tested this directly. We degraded GeoPT in 12 ways: sparse, noisy and partially missing input surfaces, randomly initialized weights, and shuffled probes. Each degradation was evaluated with the calibration frozen and with the boundary recalibrated on equally degraded members. On the blended wing body, 20 of the 40 held out members set the boundary and the other 20 test it. We report the rejection of held out members and the admission of anomalous designs, namely the unconstrained deliveries that fail a geometry audit.

GeoPT input or weightsFrozen: members rejectedFrozen: anomalies admittedRecalibrated: members rejectedRecalibrated: anomalies admitted
Released (Ours)5.0%0 of 48.5%0%
25 to 50% of points kept5.0 to 10.0%0 of 47.8 to 9.3%0%
Noise, 0.1 to 1% of the diagonal12.5 to 100%0 of 410.8 to 11.8%0 to 63.8% †
70% of one side removed, 5 planes7.5 to 100%0 of 48.0 to 11.3%0%
Random weights100%0 of 410.0%0%
Shuffled probes100%0 of 49.8%100% †

† The two nonzero values (1% noise and shuffled probes) are detected before optimization by the label free check described below.

Our findings are twofold:

  • With the calibration frozen, PAC fails conservatively. No degradation admits an anomalous design, and at least 99.1% of the designs rejected by the released boundary remain rejected. Instead, a degraded fingerprint moves the family's own members out of the band, so PAC becomes stricter. Because held out members need no labels, this loss of coverage is visible before any search is run.
  • After recalibration, the remaining failures are detectable without labels. Recalibration brings the rejection of held out members back to between 7.8 and 11.8%. Anomalies are then admitted in only two conditions, and in exactly these two the detection of a 2% bump on held out members (the measure of Table 3) drops to 0.54 or below, compared with 0.77 to 0.97 in all other conditions.

The paper also contains the extreme case of a randomly initialized GeoPT (Table 3). It still works as a fence, and after recalibration it admits no anomaly here either, but it loses detection (0.05 to 0.25 AUC) and selection power (17.4× to 7.4×). These results are summarized in Sec. 5.2 and detailed in Appendix G.3, Table 17 (in blue). We hope these additions address the reviewer's concerns, and we would be glad to discuss any remaining questions.

I thank the authors for their feedback. My major concerns have been addressed. It seems that the method leads to a bit of computational overhead, but I consider this trade-off acceptable given the improved performance. Therefore, I will keep my original positive rating.

AuthorsOpenReview ↗

Response to Reviewer Srzp

Thank you for taking the time to read our response. We appreciate your positive assessment and your constructive feedback throughout the review process.

Reviewer_cZEPOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Physics-Aware Constraints (PAC), a supervision layer that can help prevent an optimizer from disregarding relevant physical constraints. The authors demonstrate that across four CAE families, PAC consistently keeps optimization inside the trust band.

Critical strengths
  1. Throughout the paper, the authors are clear about predictions made by the model versus that are are made through independent hull calculations. This is particularly important because an optimizer can find a shape that looks better to the prediction model without actually improving the hull.
  2. Comparisons with an untrained GeoPT model and ShapeVAE, along with tests of the separate PAC components, address several possible reasons for the results.
Critical weaknesses
  1. I think this paper misues hte idea of a "trust" region, to some degree. The guarantee covers new shapes drawn like the calibration examples but it doesnt necessarily show that the physics model is accurate on shapes selected during optimization. I think some degree of independent verification is necessary.
  2. The hull results do not isolate PAC’s contribution. After the feasibility rules are applied, PAC and simpler boundaries produce the same median improvement. This supports the final design, but its not clear why PAC is better than a simpler search rule.
  3. Engagement with related work is limited.
Questions for authors
  1. Does PAC identify regions where the physics model predicts accurately, including for shapes chosen during optimization? How does it compare with error-based approaches such as Zhang et al. (10.1115/1.4026150)?
  2. With the same number of independent hull calculations, does PAC find better feasible designs than no boundary or simpler boundaries? Ablation to demonstrate what each part of PAC actually contributes would be valuable, but perhaps out of scope.
5 comments
AuthorsOpenReview ↗

Response to Reviewer cZEP

We sincerely thank the reviewer for the careful reading and for recognizing that we keep the model's predictions apart from independent calculations. We address each point below. New and changed text in the revised manuscript is shown in blue.

W1 / Q1.1: The guarantee covers new shapes drawn like the calibration examples, but does the physics model predict accurately on shapes selected during optimization? Some independent verification is necessary.

We completely agree with the reviewer that this distinction is important. The coverage bound in Appendix A.1 holds for new family members, as the reviewer notes. PAC's band therefore certifies support, meaning that a design lies where the family gives the surrogate evidence (Sec. 3). Whether the surrogate is accurate on the optimized designs is an empirical question.

To address this, we re-simulated designs selected during optimization with our own OpenFOAM setup, which is independent of the surrogates, on two families and two parametrizations. Together with the runs that align our solver with the dataset labels, this verification used about 14,000 core hours of RANS simulation.

(a) Car, free form deformation (DrivAerNet++). We re-ran PAC from the starting car of Table 1 and delivered the best in band iterate that also meshes cleanly (watertight, wheel clearance, no new self-intersection). We simulated the iterates at 22%, 41% and 100% of the claimed improvement (steady RANS, k-ω SST, 30 m/s, 6.4M cells).

Car designStepCellsShare of claimCFD C_dCFD ΔC_d
Initial06.4M0%0.2828
PAC26.4M22%0.2680−5.2%
PAC56.4M41%0.2621−7.3%
PAC, delivery236.4M100%0.2464−12.9%
6.4M, remeshed100%0.2444−13.6%
24.8M100%0.2483−9.6%
Unconstrained, final119not meshable, body passes through the wheels

(b) Blended wing body, neural implicit (BlendedNet). This comparison is not one to one, because the labels of this dataset come from a different solver (FUN3D). On 41 dataset pairs that differ mainly in geometry, our OpenFOAM setup reproduces the direction of the labelled drag change in 37. The designs are decoded from the PAC and unconstrained trajectories (compressible RANS, Spalart–Allmaras). PAC's deliveries are averaged over three meshes, with the range in brackets.

StartDesignd/τClaimed ΔC_DCFD ΔC_DVolume
1PAC delivery0.73−10.5%−1.6% [−2.6, −0.5]+6%
1Unconstrained, first past 2τ2.81−19.9%+14.3%+105%
2PAC delivery0.71−8.0%+1.1% [−0.7, +3.4]+2%
2Unconstrained, first past 2τ2.55−27.1%+14.0%+90%

The results show three things:

  • Car. The car labels come from the same solver (Pearson 0.89 with the labels on 10 dataset cars). CFD drag decreases at each simulated PAC iterate in the order the surrogate predicts (Spearman 1.0). The delivered reduction is 12.8 times the 2σ remeshing noise and is also observed on a 24.8M-cell mesh.
  • Blended wing body. Beyond twice the boundary, the surrogate predicts 20 to 27% lower drag for a body whose volume nearly doubles, while CFD measures 14% higher drag. This is far above the variation of drag over the averaging window, which is below 2%. The prediction error is about 9 points inside the band and 34 to 41 points beyond 2τ.
  • Unconstrained endpoints. On the car, the final unconstrained design passes the body through the wheels (11,133 new self-intersections) and cannot be meshed. Applying the same delivery rule returns its first step, because the search leaves the band at step 2 and does not return. On the blended wing body, the unconstrained endpoints do not reach a steady state from any of four starting points.

In summary, PAC's delivered car keeps its improvement under an independent solver, while predictions beyond the band can be wrong in sign. These results are now in the main text (Sec. 5.1, Table 4), with details in Appendix F.6 (in blue).

AuthorsOpenReview ↗

Response to Reviewer cZEP

Q1.2: How does PAC compare with error-based approaches such as Zhang et al. (10.1115/1.4026150)?

We thank the reviewer for pointing us to this work, which we now discuss in Sec. 2 and Appendix G.1. DSUS (Zhang et al., 2014) estimates the surrogate's error at its training designs by leave-one-out cross-validation and trains a classifier that divides the design space into classes of error level. It is well suited to problems where labels are available over a fixed design vector. The two methods are complementary:

DSUS (Zhang et al., 2014)PAC (Ours)
Signalcross-validated surrogate errorresponse fingerprint distance to the family
Labels neededat every training designnone
Design representationfixed design vectorany surface: FFD, neural implicit, vector, generated
Source of the estimatemodels fitted to the surrogate's own dataa separate frozen model, never fitted to the family
Use in the loopregion map over the vectordifferentiable penalty or gate
Guaranteeempirical error classesdistribution free coverage (Appendix A.1)

Two differences are relevant for offline inverse design:

  • Labels. DSUS learns its error classes from labels at the training designs and is defined over the design vector on which they were collected. PAC needs no labels, and it applies when the representation changes from FFD to a neural implicit or a generated one.
  • Independence. DSUS estimates the error with models fitted to the surrogate's own labelled data. Appendix F.5 evaluates the closest label free signal of this kind, an ensemble of five members and a Monte Carlo dropout model trained on the bank geometries. They rank their own error at ρ = +0.55 but detect a 2% deformation at a mean AUC of only 0.65, compared with 0.89 for the frozen pretrained readout (paired gaps +0.087 to +0.344, all excluding zero). As the deformation grows, the predicted changes of the members become more correlated (0.72 to 0.93), so their disagreement becomes less informative away from the data.

The two methods can also be combined. When a family has labels over a fixed vector, a DSUS map can grade the designs inside PAC's band into error classes. We note this extension in Appendix G.1.

W3: Engagement with related work is limited.

We thank the reviewer for this suggestion, and we agree that the paper should make these connections explicit. PAC lies at the intersection of several communities. The question it asks is whether the surrogate that steers an inverse loop can still be trusted at the design the loop produces, when no label can be obtained there. Each of the following lines of work addresses part of this question, and to our knowledge none addresses it directly. We now relate PAC to each of them (Sec. 2, Appendix G.1):

  1. Trust region model management (Alexandrov et al., 1998; Forrester & Keane, 2009) is the closest in purpose, since it limits how far a search may rely on a surrogate. It sizes the region by solver evaluations, which an offline loop does not have.
  2. Error based characterization of surrogate uncertainty (Zhang et al., 2014) is the closest in output, a map of where the surrogate can be relied on. It is built from labels over a fixed design vector.
  3. Physics guided generation (Wu et al., 2024; You et al., 2026) is the closest in setting, where a learned physics predictor steers shape synthesis. It keeps the shape plausible and leaves the reliability of the steering predictor to later simulation.
  4. Offline model based optimization and out of distribution detection (Trabucco et al., 2021; Ben Braiek et al., 2022) is the closest in the failure it studies, a search that exploits a learned score outside its data. It works on labelled designs over vector spaces without spatial structure, and does not indicate where a shape may change.

The new table in Appendix G.1 compares these lines on the labels they need, the parametrizations they accept, and the question they answer. We hope this clarifies what PAC inherits from this literature and where it differs.

AuthorsOpenReview ↗

Response to Reviewer cZEP

W2 / Q2: With the same number of independent hull calculations, does PAC find better feasible designs than no boundary or simpler boundaries?

We thank the reviewer for this excellent question. To address it, we conducted an equal budget study. We drew 20 independent pools of 1,400 hulls from a diffusion generator trained on the family (28,000 hulls in total) and screened them with the analytic naval feasibility rules, which require no oracle call. Each rule then spends B Michell calls on its top B hulls. We report the best verified feasible gain over a fixed reference hull (the median hull of the pool in Figure 7), defined as the ratio of wave drag averaged geometrically over 32 operating conditions, with means and paired bootstrap intervals over pools.

Selection ruleB=1B=4B=8B=16
No boundary, random6.9×11.8×16.9×23.8×
Parameter distance, standardized, 5 nearest2.3×7.4×9.3×14.3×
Parameter distance, whitened to mean2.1×5.5×8.2×14.2×
Parameter distance, whitened, 5 nearest2.5×6.5×9.7×16.3×
Parameter distance, PCA 95%1.8×5.5×7.9×14.2×
PAC fingerprint, label free (Ours)27.0×43.5×56.9×69.1×
Surrogate ranking, with or without the PAC band50.4×68.2×73.4×73.4×

The results show three things:

  • Among the rules that use neither labels nor the surrogate, PAC performs best. At B = 8 it is 5.9 to 7.2 times better than the parameter-space boundaries and 3.4 times better than random selection. The paired 95% intervals exclude zero at every budget from 1 to 32.
  • PAC approaches the surrogate's own ranking without using the surrogate. The gap decreases from 56.9× versus 73.4× at B = 8 to 69.1× versus 73.4× at B = 16 (difference −4.4, 95% CI [−10.5, 0.0]). PAC also spends fewer calls on hulls the oracle cannot complete (65.9% valid calls versus 58.8%, +7.2 points, CI [+2.8, +11.9]).
  • Used as a fence on the surrogate's ranking, PAC loses no gain. When hulls inside each band are ranked by the surrogate, the PAC band returns the same best hull as no boundary in all 20 pools (73.4×). A parameter band at 0.4 r90, in contrast, admits no candidate (1.0×).

This also explains the observation in W2. On the ship hull, the analytic feasibility rules already remove the region where the surrogate's descent fails (Appendix F.3), so they act as a fence themselves, and the benefit of PAC appears as label free ranking. On the families without such rules, the fence is necessary: the unconstrained car design cannot be meshed, and the blended wing body predictions beyond the band are wrong in sign (see W1). These results are now in the main text (Sec. 5.1, Table 5), with details in Appendix F.7 (in blue).

Q2 (ablation): What does each part of PAC contribute?

We appreciate this suggestion. The paper isolates each component, and the revision adds two further studies:

Component removed or replacedEffectLocation
Both constraintssearch ends outside the band in 6 of 6 runs (3.2 to 10.3 times the boundary)App. D
Fence (field only)search ends outside the band in 6 of 6 runs (1.0 to 9.5 times)Table 2, App. D
Field (fence only)delivers in band, but moves the regions the family holds fixed as much as the free onesTable 2, Table 12
Pretraining (random GeoPT)detection AUC lower by 0.05 to 0.25; oracle gain 17.4× to 7.4×Table 3
Response metric (voxel)harmless noise consumes 40% of the band, compared with 19%Table 3, App. F.4
Family read budget (fixed fraction)removes every feasible iterate on CRMSec. 5.2
GeoPT degraded (new)no anomaly admitted under 12 degradations (frozen); the two failures after recalibration are detected in advanceApp. G.3
Family reduced (new)rates within 5 points of the full bank on average with 8 of 113 members; a removed mode is treated more conservativelyApp. G.4

With both constraints, the median share of the search spent inside the band increases from 11% to 67%. We hope these additions address the reviewer's concerns, and we would be glad to discuss any remaining questions.

Reviewer_zqWWOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses surrogate hacking in inverse shape design, where optimization guided by a learned physics surrogate can exploit prediction errors and produce seemingly improved but physically unreliable designs outside the training distribution. The authors propose Physics-Aware Constraints (PAC), which uses a frozen pretrained geometry–physics model, GeoPT, to constrain the optimization process. PAC introduces two complementary constraints: 1. a response-manifold constraint that limits the deviation of a candidate’s GeoPT fingerprint from those of reference shapes. 2. a deformability field that determines where and how much a shape may change based on spatial variability within the reference design family. The existing surrogate continues to guide performance optimization, while PAC restricts the search to a calibrated region supported by the reference data. Experiments cover blended wing bodies, NASA CRM aircraft, cars, and ship hulls, with ablation studies examining the contributions of individual components and independent analytic verification of the ship-hull results.

Critical strengths
  1. PAC uses an independent, frozen physics–geometry model to construct a response-based trust constraint and combines it with a family-derived deformability field. Together, these address both how far a design may depart from the reference family and where deformation is permitted.
  2. Evaluation across four design families and multiple parameterizations supports the framework’s applicability, while component ablations clarify why both constraints are needed. The independently verified 64.5% reduction in ship-hull wave drag provides a concrete example of retaining meaningful physical improvement while restricting the search.
Critical weaknesses

None

Questions for authors

PAC derives both its trust boundary and deformability field from the reference family. What assumptions about the size and diversity of this family are needed for reliable calibration? In particular, could a region with little observed variation be treated as fixed simply because the reference set underrepresents valid designs?

1 comment
AuthorsOpenReview ↗

Response to Reviewer zqWW

We sincerely thank the reviewer for the positive assessment and for recognizing both the independent frozen supervisor and the verified gain on the ship hull. We address the question below. New text in the revised manuscript is shown in blue.

Q1.1: What assumptions about the size and diversity of the reference family are needed for reliable calibration?

We thank the reviewer for this important question. Each of the two constraints relies on one assumption:

  • Size, for the trust boundary τ. The held out members should be exchangeable with the designs the boundary is meant to admit, and their number determines the coverage that can be certified. By Appendix A.1, n held out members certify a coverage of k/(n+1), so a level of 95% requires n ≥ 19.
  • Diversity, for the deformability field. The family has to vary a region before PAC allows that region to change.

To quantify the first assumption, we subsampled the blended wing body bank of 113 members ten times at each size, using 20 held out members for calibration and 20 for testing.

Bank sizeτHeld out members rejectedOriginal out of band designs rejected
87148.0%96.1%
163767.5%98.1%
322619.5%99.2%
642168.5%99.1%
113 (full)1885.0%99.6%

As the bank becomes smaller, τ increases with the spread of the fingerprints. As a result, both rates remain on average within 5 points of the full bank down to 8 members, and with 8 members the boundary still rejects 96% of the designs that the original boundary rejects.

Q1.2: Could a region with little observed variation be treated as fixed simply because the reference set underrepresents valid designs?

We thank the reviewer for this thoughtful question. This is a conservative choice that PAC makes by design. If the family never makes a certain change, no data can support or refute it, so PAC does not trust it. To measure the cost of this choice, we clustered the latent codes of the family into five groups, removed one group at a time from the bank and the calibration, and scored the members of the removed group. As a control, we removed the same number of random members.

Cluster removedIts members rejectedSame size random removal4 members added back
052.4%18.1%28.6%
171.4%37.6%42.9%
26.5%3.5%3.2%
33.7%0.4%3.7%
415.4%9.2%15.4%

The results show three things:

  • A removed mode is treated more conservatively. In all five clusters, removing the cluster increases the rejection of its members relative to random removal of the same size, by 2.9 to 34.3 points.
  • The deformability field is almost unaffected. With a cluster removed, at most 0.07% of its members' deviation from the starting designs falls in regions of low deformability, compared with 0.02 to 0.04% for the full bank.
  • Adding a few members reduces the cost without retraining. Adding 4 members of the cluster back to the bank reduces the rejection in three of the five clusters, from 52% to 29% and from 71% to 43% in the two most affected ones. Neither GeoPT nor the surrogate is retrained.

In addition, the field is a map over the surface (Figure 3, black where the family holds the surface fixed), so the fixed regions are known before the search starts and can be reviewed by an engineer. We have added this analysis to Appendix G.4 and to the Limitations (Appendix H), in blue. We hope these additions address the reviewer's question, and we would be glad to discuss it further.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Response Summary and Updated Manuscript to All Reviewers

We sincerely thank all reviewers for their thoughtful and constructive feedback, which has substantially strengthened our work. In particular, we greatly appreciate the acknowledgements from:

  • Reviewers zqWW, Srzp: for recognizing the separation of optimization from supervision, where PAC "uses an independent, frozen physics–geometry model to construct a response-based trust constraint" (zqWW) by "assigning them to two different frozen models" (Srzp).
  • Reviewers zqWW, Srzp: for recognizing that the two constraints "operate at different levels", which "makes PAC applicable to multiple shape parameterizations" (Srzp), and that the "evaluation across four design families and multiple parameterizations supports the framework's applicability" (zqWW).
  • Reviewers cZEP, zqWW: for recognizing the care of our evidence: we are "clear about predictions made by the model versus [those] made through independent hull calculations", and the comparisons "with an untrained GeoPT model and ShapeVAE, along with tests of the separate PAC components, address several possible reasons for the results" (cZEP), while "component ablations clarify why both constraints are needed" (zqWW).
  • Reviewers zqWW, Srzp: for recognizing that PAC keeps real physical gains: "the independently verified 64.5% reduction in ship-hull wave drag provides a concrete example of retaining meaningful physical improvement while restricting the search" (zqWW), and "PAC yields visually realistic shape designs and also improves the physical attributes" (Srzp).

Based on this feedback, we conducted new experiments and analyses. All new and changed text in the revised manuscript is shown in blue, and the appendix now opens with a table of contents in which the new sections are marked.

ConcernReviewerLocation in Manuscript
Independent CFD of designs selected during optimizationcZEPSec. 5.1, Table 4 (main text); Appendix F.6
Trust band versus surrogate accuracy; error based methods (Zhang et al., 2014)cZEPSec. 2; Appendix F.1, F.6, G.1
Equal oracle budget against no boundary and simpler boundariescZEPSec. 5.1, Table 5 (main text); Appendix F.7
Positioning against PhysGen and physics guided generationSrzp, cZEPSec. 2; Appendix G.1
Runtime and computational costSrzpSec. 5.2; Appendix G.2
Robustness to an unreliable GeoPTSrzpSec. 5.2; Appendix G.3
Size and diversity of the reference familyzqWWAppendix A.1, G.4, H

In summary, the new results are:

  1. Independent CFD of optimized designs (cZEP). On the car, PAC's delivered design reduces CFD drag by 12.9%, and drag decreases along the PAC trajectory in the order the surrogate predicts. On the blended wing body, unconstrained designs beyond twice the trust boundary are predicted to reduce drag by 20 to 27% but increase it by 14% in CFD.
  2. Equal oracle budget (cZEP). With 8 oracle calls, ranking by PAC's fingerprint reaches a 56.9× gain in wave drag, compared with 7.9 to 9.7× for boundaries in the design parameter space.
  3. Cost (Srzp). PAC adds 0.10 to 0.15 s per optimization step and less than 20 s per run.
  4. Robustness (Srzp). Across 12 degradations of GeoPT, the frozen boundary admitted none of the anomalous designs on the blended wing body. After recalibration, the two failing cases are detected in advance by a label free check.
  5. Family size and diversity (zqWW). The boundary remains stable on average with as few as 8 bank members. Removing a mode of the family makes PAC more conservative on that mode, and adding 4 of its members back reduces the effect by up to 29 points.

We hope that these updates address the reviewers' concerns and further clarify the value of our work.

Thank you for the detailed response. The author have successfully answered my questions.

AuthorsOpenReview ↗

Response to Reviewer zqWW

Thank you for your kind feedback! We sincerely appreciate the time and care you have devoted to reviewing our manuscript, and we are pleased that our responses have satisfactorily addressed your questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.