- 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
- 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.
- 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.
- 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
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.
| Path | Unconstrained (s/step) | PAC (s/step) | Added per step | Added per run | Peak memory (GiB) |
|---|---|---|---|---|---|
| Car, FFD, 120 steps | 0.031 | 0.131 | +0.10 s | +12.8 s | 0.38 → 1.21 |
| Blended wing body, neural implicit, 180 steps | 6.23 | 6.38 | +0.15 s | +18.2 s | 0.61 → 1.59 |
| Ship hull, scoring a generated hull | 0.47 s per hull | ||||
| One time calibration per family | 12 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 weights | Frozen: members rejected | Frozen: anomalies admitted | Recalibrated: members rejected | Recalibrated: anomalies admitted |
|---|---|---|---|---|
| Released (Ours) | 5.0% | 0 of 4 | 8.5% | 0% |
| 25 to 50% of points kept | 5.0 to 10.0% | 0 of 4 | 7.8 to 9.3% | 0% |
| Noise, 0.1 to 1% of the diagonal | 12.5 to 100% | 0 of 4 | 10.8 to 11.8% | 0 to 63.8% † |
| 70% of one side removed, 5 planes | 7.5 to 100% | 0 of 4 | 8.0 to 11.3% | 0% |
| Random weights | 100% | 0 of 4 | 10.0% | 0% |
| Shuffled probes | 100% | 0 of 4 | 9.8% | 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.
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.
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.
Two of our results relate directly to this distinction:
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.