acceptodds
Under review as a conference paper at ICLR 2027

Simulator-Refined Diffusion for Radio-Frequency Inverse Design

Abstract

Diffusion models have shown potential in inverse design of printed circuit boards (PCBs), enabling the generation of layouts conditioned on target S-parameters. Despite this promise, applying diffusion models to PCB layout generation remains challenging due to their difficulty in meeting the quantitative electromagnetic specifications. A common approach is gradient-based guidance, which biases the diffusion sampling process with the gradient of an objective used for evaluation. However, full-wave electromagnetic simulators are accurate but expensive and typically non-differentiable, whereas differentiable surrogates are informative but not always reliable. To address these limitations, this paper proposes Simulator-Refined Diffusion (SRD), a novel combination of a low-fidelity differentiable surrogate and a high-fidelity non-differentiable simulator within the diffusion sampling process. Unlike standard zeroth-order optimization, which requires a great number of random perturbations, our approach uses the surrogate's gradient to propose the perturbation direction while the simulator then searches based on this direction to identify an effective design update. Experimental results across different settings show that this method consistently outperforms current state-of-the-art methods, producing layouts whose simulated S-parameters match the target specifications up to 21.2% closer for in-distribution targets and up to 19.8% closer for out-of-distribution targets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
printed circuits, diffusion models, inverse design, guided diffusion, inference-time scaling, electromagnetic simulation, printed circuit boards, surrogate models, RF design

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

In this paper, a new method called simulator refined diffusion (SRD) is proposed to solve the RF PCB inverse design. In view of the contradiction between full wave simulator (accurate but expensive and non differentiable) and differentiable proxy model (cheap but with distribution offset), SRD combines the two in the final stage of diffusion sampling. Specifically, SRD uses CenterNet to extract the diffusion prediction as a low dimensional structured rectangle parameter, proposes the editing direction using the gradient of the proxy model, and then uses the limited simulator budget to evaluate a small number of candidates in this direction (or reverse/random direction), selects the geometry with the least simulation loss, and injects it into the diffusion trajectory through the "Write back & Renoise" mechanism.

Critical strengths

1.In this paper, the gradient of surrogate is used to generate candidate directions, and then the palm simulator is used to select them. In view of the contradiction between full wave simulator (accurate but expensive and non differentiable) and differentiable proxy model (cheap but with distribution offset), SRD skillfully combines the two in the final stage of diffusion sampling.

Critical weaknesses

Weakness 1: Insufficient Justification for Using a Single Surrogate Gradient Direction

The proposed method computes a single surrogate gradient at the current geometry and generates multiple candidates by exploring different step sizes along the resulting direction. However, this restricts the surrogate-guided search to a fixed direction, even though the surrogate is differentiable and inexpensive to evaluate. A natural alternative would be to iteratively recompute the surrogate gradient and update the rectangle parameters, while respecting the same geometric admissibility requirements. Importantly, rather than relying solely on the final optimized geometry, several intermediate geometries along the optimization trajectory could be collected as candidates and evaluated by the full-wave simulator. This would allow the search to explore changing gradient directions while retaining simulator-based selection to mitigate surrogate inaccuracies. The paper does not compare against such an iterative surrogate-guided candidate generation strategy under the same simulator-call budget. Consequently, it remains unclear whether restricting candidate generation to a single surrogate gradient direction is preferable to exploring a sequence of surrogate-guided geometric updates.

Weakness 2: Unclear Necessity of Performing Refinement During Diffusion Sampling

The proposed method performs simulator-based refinement during the final denoising steps. However, it remains unclear whether integrating refinement into diffusion sampling provides an advantage over applying the same refinement procedure after diffusion has completed. Since the proposed method operates on extracted geometric parameters, its candidate generation and simulator-based selection strategies could also be applied directly to the final generated PCB layout.

A more appropriate comparison would be single-step (what the authors do in the maintext) and multi-step post-hoc refinement using exactly the same candidate search strategies (directional, adaptive, or hybrid) as SRD, under the same simulator-call budget. In the multi-step setting, each refinement round would start from the previously selected geometry, recompute the surrogate gradient, and generate a new set of candidates using the original SRD search procedure. This would allow the optimization direction to change across rounds without involving additional diffusion steps.

Such a comparison is particularly important because subsequent denoising may modify the simulator-selected geometry and potentially diminish the improvements obtained from refinement. Furthermore, the RENOISE operation inherently injects noise back into the simulator-selected geometry, which risks introducing artifacts or pushing the state off the natural data manifold. Without these controlled post-hoc baselines, it is difficult to determine whether the reported improvements arise from performing refinement within diffusion sampling or simply from repeated surrogate-guided candidate search and simulator-based selection.

Weakness 3: Unclear Noise Magnitude and Reliability of Geometry Extraction at Intermediate Diffusion Steps

The diffusion process introduces Gaussian noise into clean PCB layouts during training, causing the noisy states to deviate from the original clean-layout distribution. Although SRD applies CenterNet to the predicted clean layouts rather than directly to the noisy states, these intermediate predictions may still contain reconstruction errors. However, the paper does not quantify how much noise remains at the final denoising steps where refinement is performed. The authors should report the Gaussian noise schedule used during diffusion training, the corresponding noise standard deviation or signal-to-noise ratio at each refinement timestep, and whether CenterNet was exposed to similarly corrupted inputs during its training.

This leads to two concerns. If the intermediate predictions still differ substantially from clean layouts, the CenterNet extractor, which is trained on rasterized PCB layouts, may suffer from a distribution shift. In this case, why not train a timestep-aware geometry extractor on intermediate diffusion predictions, analogous to the noise-aware classifier used in classifier guidance? If the differences are negligible, it is unclear whether CenterNet can resolve such small geometric variations, particularly if its training images were quantized to 8-bit grayscale. More importantly, if the geometries barely change during the final denoising steps, what advantage does intermediate refinement offer over applying the same geometric refinement directly to the final generated layout?

The authors should quantify the changes in predicted layouts and extracted rectangle parameters across the final denoising steps, clarify CenterNet's training input precision, and evaluate its sensitivity to small geometric variations. These results are necessary to justify both the reliability of geometry extraction and the choice of refinement timesteps.

Questions for authors

See weaknesses

6 comments

Many thanks for your detailed assessment of our manuscript. We are pleased to read of your appreciation for our novel combination of the surrogate model and full-wave simulator, which is integral to the strong empirical performance reported. Below, we have responded to each concern raised in your assessment.

W1: Insufficient Justification for Using a Single Surrogate Gradient Direction

Thank you for this suggestion. To test whether one surrogate direction is enough, we implemented a variant that recomputes the surrogate gradient for each candidate. At each guided denoising step, the predicted board is converted to a rectangle layout. After that, the rectangles are perturbed along the surrogate's descent direction, and recomputing the surrogate gradient. The original board and all perturbed candidates are simulated with Palace, and the best layout is written back before denoising continues. The budget is the same as SRD. The only difference is that SRD computes the gradient once and places its candidates along that single direction.

MethodTest Mean L_sim ↓Test Median L_sim ↓OOD Mean L_sim ↓OOD Median L_sim ↓
SRD (directional)0.26400.19800.43460.4082
SRD (adaptive)0.26080.19770.44440.4338
Multiple Surrogate Gradient0.25080.16830.46470.4937

In the test set, where the surrogate is reliable, Multiple Surrogate Gradient follows the curved descent path more closely and lowers the test loss. Out of distribution, where the surrogate is much less reliable (Figure 1(c)), repeatedly following its gradient compounds errors and spends the simulator budget on poor candidates, so Multiple Surrogate Gradient is clearly worse than SRD (mean 0.4647 vs. 0.4346, median 0.4937 vs. 0.4082). This trade-off supports our design: committing to one direction and letting the simulator probe decide whether to follow it makes SRD more robust when the surrogate cannot be trusted.

W2: Unclear Necessity of Performing Refinement During Diffusion Sampling

We agree that the effect of refinement during the final denoising steps should be compared with post-hoc refinement. Indeed, we did include such an ablation in Figure 5. The left panel illustrates a single guidance step, which is equivalent to the post-hoc refinement baseline you described. We see that the benefit of additional simulator calls levels off quickly. In contrast, applying SRD at multiple diffusion steps results in notably stronger performance.

The unguided sample is too narrow and broken by a small gap. Post-hoc it closes the gap and widens one segment until it reaches the board edge, leaving a line of uneven width with 13 simulator calls. In SRD, the denoising step after each edit redraws the line as a single segment of uniform width, so the next search step widens the whole line, reaching a lower s-params loss with the same 13 calls.

We agree that this could be better emphasized in our paper. To make this clearer, we’ve added a new visualization to the Appendix A.5.

W3a: If the intermediate predictions still differ substantially from clean layouts, the CenterNet extractor, which is trained on rasterized PCB layouts, may suffer from a distribution shift.

Note that CenterNet is never applied to the noisy diffusion state xtx_t. At each guided step it is invoked on clean samples x^0t\hat{x}_0^t, i.e., the model's current estimate of the final design.

In our experiments we did not observe degradation of the rectangle extraction at the guided steps. Had extraction been unreliable, it would surface as malformed geometries and elevated simulation failure rates. SRD's failure rate (0.5% of final designs) matches that of the non-guided methods, and its seed-level variability (±0.009) is comparable to the baselines.

However, we agree that there is no guarantee against distribution shift. CenterNet may degrade on predictions far from its training distribution. In developing the method we also evaluated traditional CV extraction (e.g., watershed segmentation) and CenterNet proved more robust on our layouts. We have added this limitation to Section 7 where we also note our intention to explore more compact representations (e.g., diffusion directly in rectangle space), which would remove the need for CenterNet altogether.

W3b: If the geometries barely change during the final denoising steps, what advantage does intermediate refinement offer over applying the same geometric refinement directly to the final generated layout?

Starting refinement before the final steps gives the remaining denoising steps the chance to act on the edit. The prior can repair artifacts the geometric edit introduces and adapt to it, including changes that no local rectangle perturbation can express.


We’d like to, again, express our appreciation for your time reviewing our work. We believe that our responses have addressed each concern raised, but we are happy to engage in further discussion if any questions remain. Thank you.

Issues Addressed by the Revision

W1: Insufficient Justification for Using Only a Single Surrogate Gradient Direction

The authors have addressed this concern by introducing a multiple-surrogate-gradient variant in their rebuttal. The additional results show that repeatedly recomputing surrogate gradients improves performance on in-distribution targets but leads to worse performance on OOD targets, suggesting that repeatedly following potentially inaccurate surrogate gradients can accumulate errors under distribution shift.

This provides useful empirical evidence supporting the proposed strategy of using a single surrogate-proposed direction and relying on simulator feedback to determine whether and how to continue the search. I consider this concern largely addressed. However, I encourage the authors to incorporate the additional experiment and its implementation details into the revised manuscript for completeness and reproducibility.

Remaining Weaknesses

W2: Unclear Necessity of Performing Refinement During Diffusion Sampling

The revised manuscript has improved the discussion of diffusion-time refinement and added a comparison with post-hoc refinement in Figure 5 and Appendix A.5. In particular, the authors provide a qualitative example showing that subsequent denoising can repair artifacts introduced by local geometric edits, such as gaps and non-uniform conductor widths. This offers a reasonable explanation of how the diffusion prior can complement local geometric optimization.

However, the current post-hoc baseline performs refinement only once after diffusion sampling. It remains unclear whether the advantage of SRD comes specifically from integrating refinement into the diffusion process or simply from repeatedly applying surrogate-guided geometric optimization with simulator-based candidate selection.

A more conclusive comparison would involve multi-step post-hoc refinement using exactly the same candidate generation and selection strategies as SRD (directional, adaptive, or hybrid). After each refinement round, the surrogate gradient should be recomputed at the previously selected geometry, and a new candidate set should be generated using the original SRD procedure. The total number of simulator calls should be matched.

This comparison is important because the authors attribute the improvement of in-loop refinement to the diffusion prior's ability to repair and adapt local geometric edits. However, it remains unclear whether this advantage is specific to diffusion-time refinement or could also be achieved through multi-step post-hoc geometric optimization. Such a controlled comparison would better isolate the contribution of the diffusion prior.

W3: Unclear Noise Magnitude and Geometric Stability of Intermediate Diffusion Predictions

The proposed method applies CenterNet to predicted clean layouts at intermediate denoising steps to extract rectangle geometry. However, the paper does not specify the actual Gaussian noise schedule used during diffusion training or quantify the noise levels associated with the refinement timesteps.

The authors should report the noise schedule used during diffusion training, including the corresponding beta values, cumulative noise levels, and noise standard deviations or signal-to-noise ratios (SNRs). In particular, the specific values at the final two denoising steps on the test set (steps 19–20) and the final four steps on the OOD set (steps 17–20) should be provided.

Furthermore, the authors should quantify how the predicted clean layouts and their extracted rectangle parameters evolve across denoising timesteps along ordinary diffusion sampling trajectories without any SRD refinement. In particular, the analysis should cover steps 17–20 for the OOD set and steps 19–20 for the test set. Specifically, reporting the changes in extracted rectangle centers, widths, and heights across these timesteps, as well as the geometric differences between intermediate predicted layouts and the final generated layouts, would help establish the geometric stability of ordinary diffusion trajectories and assess whether CenterNet can reliably extract geometric information from intermediate predictions.

These analyses would help clarify the noise characteristics of the selected refinement timesteps, assess the reliability of geometry extraction from intermediate diffusion predictions, and justify the choice of performing refinement at these particular stages.

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

This paper proposes a diffusion-based inverse design method for radio-frequency PCB layout design. The PCB is modeled in a 64x64 image-like representation and transferred into a rectangle-based representation using a neural network. The diffusion model operates on the image-like representation, while a simulator and a neural surrogate model use as input the rectangle-based representation. The main idea of the method is to use the differentiable surrogate model for guiding the diffusion. The guidance by the surrogate model is not directly applied. Instead a computationally expensive simulator is used to evaluate either steps in the guidance direction or random perturbations at the expected clean sample of the denoising process. In the experimental results, the approach demonstrates improvements over previous methods which either use only the surrogate model for guidance or simulator based sampling.

Critical strengths
  • The approach to combine surrogate guidance with simulator-based evaluation of perturbations in an adaptive or hybrid approach is interesting and appears novel.
  • The paper is well written and mostly clear.
Critical weaknesses
  • Why not train the diffusion model on the much more compact rectangle representation which also seems to be continuous (up to the number of rectangles). This might be a more natural choice, since the surrogate loss is also defined on the rectangles. It might be possible to use an overparametrization of rectangles by adding an existance variable per rectangle to the state of the diffusion model. How does the approach perform on this representation instead of denoising images and going through the rectangle detection network ?
  • The related work section should also discuss related works on inverse design in other similar application domains such as material design (e.g. [*1,*2]), which are also based on guided diffusion. How do the underlying methods differ?
  • The hybrid SRD method in Table 1 requires 29 Palace calls instead of the usual budget of 17 calls. Why was the number of calls altered? What is the average run-time of the Palace calls in the experiments ?
  • What is the number of denoising steps for the diffusion model?
  • What is the overall run-time of the different methods in Table 1 to yield a denoised PCB board sample?
  • Sec 6.2. provides results for different guidance steps. When are these guidance steps applied? Once for the whole denoising step or at every denoising step?
  • Sec 7 claims "results are significant", but statistical significance is not evaluated. Please adjust the claim.

Minor comments:

  • Fig. 1 should be referenced from the text.

[*1] Kreber et al. Guided Diffusion by Optimized Loss Functions on Relaxed Parameters for Inverse Material Design. TMLR 2026. [*2] Yang et al. Guided Diffusion for Fast Inverse Design of Density-based Mechanical Metamaterials. arXiv:2401.13570, 2024.

Questions for authors

Please address questions in "Weaknesses", especially about the state representation for the diffusion (rectangles vs images) and the runtime of the approaches.

2 comments

Thank you for your time and efforts in reviewing our work. We appreciate your recognition of the novelty of our work and are glad that you found the paper to be clear and well-written. We have reviewed each of the points raised in your assessment, and we provide responses to each below.

W1: Why not train the diffusion model on the much more compact rectangle representation which also seems to be continuous (up to the number of rectangles). This might be a more natural choice, since the surrogate loss is also defined on the rectangles. It might be possible to use an overparametrization of rectangles by adding an existance variable per rectangle to the state of the diffusion model. How does the approach perform on this representation instead of denoising images and going through the rectangle detection network ?

We agree that this is a natural extension of the work, and, indeed, we flagged this in Section 7 as future work motivated by our paper. However, the goal of this paper was to leverage an existing pretrained surrogate model, formulated to work with the pixel representation adopted. Existing RF design literature has exploited similar pixel representation [1], and our existing surrogate model was designed under those assumptions.

W2: The related work section should also discuss related works on inverse design in other similar application domains such as material design (e.g. [*1,*2]), which are also based on guided diffusion. How do the underlying methods differ?

Thank you for the suggestion. We have added this discussion to the Section 2, highlighting the differences. Kreber et al. (2026) design composite materials with a linear FEM model on a fixed grid, enabling exact gradients via implicit differentiation. While RF PCB design requires solving full-wave 3D Maxwell equations, and our pipeline is non-differentiable. SRD therefore uses the surrogate only to propose directions and the simulator to select edits. Yang et al. (2024) design metamaterials with guidance from a learned predictor, similar to our GGD baseline (marginal gains in Table 1).

W3: The hybrid SRD method in Table 1 requires 29 Palace calls instead of the usual budget of 17 calls. Why was the number of calls altered? What is the average run-time of the Palace calls in the experiments ?

Hybrid uses the same guided steps but evaluates more boards per step: the other variants evaluate kk per step (the current design and k−1k − 1 candidates), giving 4×4+1=174 \times 4 + 1 = 17 calls on OOD, while hybrid evaluates the union of the directional and k−1k − 1 random candidates (i.e. 2k−1=72k − 1 = 7 per step), giving 4×7+1=294 \times 7 + 1 = 29 (11 vs. 7 on test).

The average run-time of the Palace call is 28.7 s (median 26.3 s) over 102,246 solves on one NVIDIA A10G GPU. We hope these details help clarify the runtime requirements and budgets.

We have added these description to the Appendix A.3.

W4: What is the number of denoising steps for the diffusion model?

We conduct 20 denoising steps per board. We’ve updated the Appendix A.2 to include this description alongside other hyperparameter choices.

W5: What is the overall run-time of the different methods in Table 1 to yield a denoised PCB board sample?

Thank you for the question. We are happy to clarify, and we’ve added these details to the Appendix A.3.

MethodTest: Palace callsTest: time per design (s)OOD: Palace callsOOD: time per design (s)
CD121.9125.7
GGD126.4130.4
DPO779.417150.5
SRD (adaptive)7128.217259.4
SRD (directional)7128.217261.6
SRD (hybrid)11128.129249.6

Time per design includes diffusion sampling, CenterNet extraction, surrogate gradients and Palace calls. Calls that do not depend on each other run in parallel:

  • DPO: its random perturbations are independent, so all palace calls run in parallel.
  • SRD: each guided step needs that the probe runs first, and the remaining candidates are then evaluated in parallel based on its outcome.

W6: Sec 6.2. provides results for different guidance steps. When are these guidance steps applied? Once for the whole denoising step or at every denoising step?

Guidance is applied only at the last S denoising steps, earlier steps are unguided. "1 guidance step" is post-processing: a single refinement of the final CD layout after sampling has finished,. "2" and "4" refine at each of the last 2 or last 4 denoising steps, writing the best candidate back before denoising continues. Each refinement evaluates k boards with Palace, so post-processing costs k calls, and the in-loop settings cost S·k + 1.

W7: Sec 7 claims "results are significant", but statistical significance is not evaluated. Please adjust the claim.

Thank you for bringing this to our attention. During the rebuttal period, we repeated the Table 1 validation experiment with five independent diffusion seeds.

Methodmean ± sd
CD0.3194 ± 0.0073
GGD0.3243 ± 0.0072
DPO0.2819 ± 0.0085
SRD(directional)0.2618 ± 0.0092
SRD(adaptive)0.2592 ± 0.0086
SRD(hybrid)0.2511 ± 0.0089

We hope these additional results help clarify statistical significance!


Thank you, again, for your thoughtful questions. We hope that these responses have helped address your remaining concerns and that the additional clarifications would reaffirm your positive assessment of our work. Please let us know if there is any remaining concerns or questions. Many thanks.

[1] Dreossi, Tommaso, et al. "Inverse Design of Multi-Layer Sub-Pixel-Resolution RF Passives Through Grayscale Diffusion with Flexible S-Parameter Conditioning." arXiv preprint arXiv:2605.08233 (2026).

Reviewer_ELN9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Simulator-Refined Diffusion uses surrogate-model gradients to propose geometric modification directions at the final stage of diffusion-based PCB layout generation, followed by full-wave electromagnetic simulations to select candidates. This approach aims to mitigate the inaccuracies of surrogate models.

Critical strengths
  1. Well-motivated approach. RF inverse design indeed faces a trade-off between expensive accurate simulations and imperfect differentiable surrogates. Using the surrogate to propose search directions and simulations to validate candidates is a reasonable and practical design choice.

  2. Clear improvements under matched simulation budgets. SRD adaptive consistently outperforms DPO on both the test and OOD sets while using the same number of simulation calls.

Critical weaknesses
  1. Most importantly, the independent value of integrating search into the diffusion process has not been established. The current experiments lack a direct comparison against a simple alternative: first completing standard diffusion generation, then applying the same rectangle-parameter search, surrogate gradients, and simulation-based candidate selection to the final layouts under an identical simulation budget.

    If this straightforward post-processing approach achieves comparable results, the primary contribution may lie in surrogate-assisted search in the geometric space, rather than refinement within the diffusion process itself. Previous work has also explored using surrogate gradients to accelerate black-box search (Maheswaranathan, et al. "Guided evolutionary strategies: Augmenting random search with surrogate gradients." ICML, 2019.). Therefore, the authors needs to articulate its incremental contribution more precisely.

  2. Insufficient implementation details hinder reproducibility. Eq. 5 computes gradients with respect to rectangle parameters, but the paper describes the surrogate as a ViT operating on layouts without adequately explaining how rectangle parameters are differentiably incorporated into the surrogate model. Algorithm 1 also omits important implementation details, including the actual step size, improvement threshold, random perturbation distribution, and the precise definitions of WRITEBACK and RENOISE.

  3. The OOD experimental setting is insufficiently defined, making the generalization claims somewhat overstated. Page 6 only mentions 50 "constructed S-parameter targets" without explaining how these targets are generated, which variables are shifted, or whether the resulting targets are physically realizable.

  4. Statistical uncertainty and the treatment of failed samples are insufficiently addressed. Tables 1–2 and Figure 5 do not clearly report multiple independent runs, confidence intervals, or paired-difference analyses. The simulation failure rates also vary slightly across methods, but the paper does not explain how failed samples are handled when computing average errors.

Questions for authors
  1. Under matched initial noise, edit spaces, and simulator-call budgets, does SRD outperform the same refinement applied only after sampling?

  2. How were the 50 OOD targets constructed, and how were their distribution shift and physical realizability assessed?

  3. Does SRD retain its gains when targets are shuffled across compatible designs, with gradient norms, candidate counts, and step sizes matched?

  4. Given Eq. 6’s per-step guarantee, how often do write-back and denoising undo refinement gains, and why not return the best feasible design evaluated during sampling?

5 comments

Thank you for your thoughtful evaluation of our work. Thank you also for recognizing the strong motivation underscoring our contribution and your acknowledgement of the clear improvements our approach provides over the baselines. We have now prepared responses to address each remaining concern:

W1 & Q1: [..] the independent value of integrating search into the diffusion process has not been established. The current experiments lack a direct comparison against a simple alternative: first completing standard diffusion generation, then applying the same rectangle-parameter search, surrogate gradients, and simulation-based candidate selection to the final layouts under an identical simulation budget. If this straightforward post-processing approach achieves comparable results, the primary contribution may lie in surrogate-assisted search in the geometric space, rather than refinement within the diffusion process itself. Previous work has also explored using surrogate gradients to accelerate black-box search (Maheswaranathan, et al. "Guided evolutionary strategies: Augmenting random search with surrogate gradients." ICML, 2019.). Therefore, the authors needs to articulate its incremental contribution more precisely.

Thank you for giving us the opportunity to clarify. We indeed have made this comparison in Figure 5, where the left panel illustrates applying refinement only at the final step, which is equivalent to applying complete diffusion followed by refinement. The figure illustrates that the benefit of additional simulator calls degrade within a single step. Conversely, applying it for multiple diffusion steps, as in the middle and right panels, allows us to reach a lower S-parameter loss than is achievable with post-hoc refinement.

The unguided sample is too narrow and broken by a small gap. Post-hoc it closes the gap and widens one segment until it reaches the board edge, leaving a line of uneven width with 13 simulator calls. In SRD, the denoising step after each edit redraws the line as a single segment of uniform width, so the next search step widens the whole line, reaching a lower s-params loss with the same 13 calls.

We agree that this could be better emphasized in our paper. To make this clearer, we’ve added a new visualization to the Appendix A.5.

W2: Insufficient implementation details hinder reproducibility. Eq. 5 computes gradients with respect to rectangle parameters, but the paper describes the surrogate as a ViT operating on layouts without adequately explaining how rectangle parameters are differentiably incorporated into the surrogate model. Algorithm 1 also omits important implementation details, including the actual step size, improvement threshold, random perturbation distribution, and the precise definitions of WRITEBACK and RENOISE.

Thank you for calling this to our attention. We have added all hyperparameter details to the Appendix A.2., and we clarify the specific points mentioned below:

NameValue
Rectangle parameters4 edge coordinates per rectangle. The edges carrying a feed are frozen
Differentiable rasterizationAnalytic per-pixel coverage on the 64×64 grid
Surrogate inputRasterized metal channel + fixed via channel
Gradient (Eq. 5)Surrogate S-parameter error, back-propagated to the edge coordinates
Step sizes1, 2, 3, 5, 7 px along u (1 px = 125 µm); first probe +1 px
Improvement thresholdτ\tau= 0.001 on the simulated loss
Random perturbationIsotropic Gaussian, σ\sigma = 2 px.
WRITEBACKExact raster of the chosen rectangles written into x^0\hat{x} _ {0}, only inside the moved rectangles' bounding boxes (+1 px)
RENOISExt←αˉt x^0+1−αˉt ϵ^x _ {t} \leftarrow \sqrt{\bar{\alpha} _ {t}}\,\hat{x} _ {0} + \sqrt{1-\bar{\alpha} _ {t}}\ \hat{\epsilon}, with ϵ^\hat{\epsilon} the model's prediction at this step (no fresh noise)

W3 & Q2: The OOD experimental setting is insufficiently defined, making the generalization claims somewhat overstated. Page 6 only mentions 50 "constructed S-parameter targets" without explaining how these targets are generated, which variables are shifted, or whether the resulting targets are physically realizable.

The out-of-distribution targets are composed of 50 S-parameter curves (51 points, 1–20 GHz). These are considered out-of-distribution as each has a nearest-neighbour RMSE (the Lsim\mathcal{L} _ {\mathrm{sim}} metric) to the 2,250 validation set boards at or above the median of the 250 test targets used for the in-distribution evaluation. This provides a median Lsim=0.23\mathcal{L} _ {\mathrm{sim}}= 0.23 for the out-of-distribution targets versus a median Lsim=0.11\mathcal{L} _ {\mathrm{sim}} = 0.11 for the in-distribution targets.

In practice, these response are created by shifting the S-parameter curves from the in-distribution targets. These targets are indeed realizable (e.g., reciprocal, passive and non-negatively dissipative at all frequencies), and no lossier than comparable real boards. The construction of this out-of-distribution set was overseen by domain experts, further reinforcing the applicability of the targeted responses. The paper has been updated in Section 6 to reflect these specific details.

W4: Statistical uncertainty and the treatment of failed samples are insufficiently addressed. Tables 1–2 and Figure 5 do not clearly report multiple independent runs, confidence intervals, or paired-difference analyses. The simulation failure rates also vary slightly across methods, but the paper does not explain how failed samples are handled when computing average errors.

We appreciate the suggestion. During the rebuttal period, we repeated the Table 1 validation experiment with five independent diffusion seeds:

Methodmean ± sd
CD0.3194 ± 0.0073
GGD0.3243 ± 0.0072
DPO0.2819 ± 0.0085
SRD(directional)0.2618 ± 0.0092
SRD(adaptive)0.2592 ± 0.0086
SRD(hybrid)0.2511 ± 0.0089

Simulation failures are rare. The EM solver failed to return a finite result for 0.5% of final designs for most methods (6–7 of 1,250) and 1.3% for GGD (16 of 1,250). Failed designs are excluded from the mean Lsim\mathcal{L}_{sim} and counted as unsuccessful in the success rate. Failures are caused by the generated layout rather than by solver instability. In most cases copper overlaps a port, or slivers and overlaps between rectangles prevent the mesher from producing a valid mesh.

Q3: Does SRD retain its gains when targets are shuffled across compatible designs, with gradient norms, candidate counts, and step sizes matched?

We are confused by the question; could you please clarify? It’s unclear what is meant here by “shuffled” targets. The target S-parameter response is the design specification, so the generated designs are evaluated on the same response that SRD optimizes the sample toward.

Q4: Given Eq. 6’s per-step guarantee, how often do write-back and denoising undo refinement gains, and why not return the best feasible design evaluated during sampling?

Methodtest(250 targets)
SRD (adaptive)215 / 250 = 86.0 %
SRD (directional)211 / 250 = 84.4 %
SRD (hybrid)208 / 250 = 83.2 %
DPO209 / 250 = 83.6 %

On the test set, the final sample is the best design simulated during sampling for most targets (83–86% across methods). We’ve added a these results to theAppendix A.6.


Many thanks for your consideration. We have provided additional clarifications and results to support the discussion, and we believe our responses have addressed the points raised in your assessment. If any questions remain, we would be pleased to continue the discussion during the remaining period.

Q1 Supplement

Why does SRD use surrogate gradients rather than established black-box optimization methods, (e.g., recent BO approaches) ? Could the authors include comparisons with these methods under matched simulation budgets? Also, is it necessary to integrate refinement into the diffusion process? Would applying the same gradient-based or gradient-free refinement after diffusion sampling achieve similar gains?

Q3 Supplement

That is, although the ablation study shows that gradient-guided perturbations outperform random, it does not sufficiently isolate the contribution of target-specific information in the gradient direction. If we compute the surrogate gradient for design A using the target S-parameter response of another compatible design B, while keeping A as the actual optimization target, would SRD still achieve comparable performance gains?

Addressing these concerns would strengthen my confidence of the proposed approach.

Reviewer_FsqfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SRD for RF PCB inverse design. During late denoising steps, a surrogate gradient proposes edits to rectangle-based geometry, while a full-wave simulator evaluates the candidates and selects the best edit. Three variants follow, reverse, or supplement the proposed direction with random perturbations. The experiment shows that simulated response on the designed targets is improved.

Critical strengths

The proposed method makes use of an imperfect surrogate. It uses the surrogate to focus a small search and lets the full simulator decide the retained edit, which can address the practical failure of surrogate-only guidance without requiring gradients through the simulator or model training.

Critical weaknesses
  • The OOD results are not convincing without a clear account of how the 50 targets are constructed. The paper does not specify which factors differ from training or whether the requested S-parameters are physically realizable.

  • The switching rule lacks a theoretical justification. An unsuccessful probe can result from a large stepsize even when the surrogate direction is useful. SRD nevertheless uses this outcome to reverse the direction or switch to random search, discarding smaller beneficial steps.

  • It is unclear whether DPO used the same rectangle edit space as SRD.

  • The experiment part should be further improved. For example, 1) some key baselines are missing, such as best-of-n CD samples. 2) Some key hyperparameters are also missing. 3) No wall-clock cost is reported.

Questions for authors

The authors are advised to explain how the OOD targets are constructed, which factors differ from training, and whether the requested responses are physically realizable.

Please clarify whether any OOD targets were used to select the refinement settings and explain how these settings were chosen.

The authors are advised to explain under what conditions a single simulator probe can reliably guide the choice between directional and random search.

Please explain how the switching rule distinguishes an inaccurate surrogate direction from an excessive step size or an improvement below the acceptance threshold.

5 comments

Thank you for your time and review. We are glad you also agree regarding the use of imperfect surrogate guidance approaches and thank you for praising our method’s ability to address this known failure mode through gradient-free simulator guidance. Below, we respond to the points raised in your assessment:

W1 & Q1: The OOD results are not convincing without a clear account of how the 50 targets are constructed. The paper does not specify which factors differ from training or whether the requested S-parameters are physically realizable.

The out-of-distribution targets are composed of 50 S-parameter curves (51 points, 1–20 GHz). These are considered out-of-distribution as each has a nearest-neighbour RMSE (the Lsim\mathcal{L} _ {\mathrm{sim}} metric) to the 2,250 validation set boards at or above the median of the 250 test targets used for the in-distribution evaluation. This provides a median Lsim=0.23\mathcal{L} _ {\mathrm{sim}} = 0.23 for the out-of-distribution targets versus a median Lsim=0.11\mathcal{L} _ {\mathrm{sim}} = 0.11 for the in-distribution targets, which is substantial in the PCB context.

In practice, these response are created by shifting the S-parameter curves from the in-distribution targets, so we do have an interpretable understanding of why these are OOD.

Note also that these targets are indeed realizable (e.g., they are reciprocal, passive and non-negatively dissipative at all frequencies) and no lossier than comparable real boards. We take great deal of care in this aspect. The construction of this out-of-distribution set was overseen by domain experts, who are interested in constructing these boards, further reinforcing the applicability of the targeted responses. The paper has been updated to reflect these specific details in Section 6.

W2 & Q4: The switching rule lacks a theoretical justification. An unsuccessful probe can result from a large stepsize even when the surrogate direction is useful. SRD nevertheless uses this outcome to reverse the direction or switch to random search, discarding smaller beneficial steps.

This is a reasonable concern, and our probe is designed with this exact challenge in mind. The default probe is 1 px (125 µm), which is the smallest step on the raster (i.e. no smaller step exists to fall back on). Furthermore, when the 1 px probe fails:

  • In 76.6 % of cases, none of the candidates along the line (steps of ±1, 2, 3, 5 and 7 px, in both directions) beat the unedited design;
  • A larger step along uu improves in 14.7 %;
  • The reversed direction improves in 13.1 %.

Hence, the empirical evidence suggests that there is not a strong sensitivity to the step size. Nevertheless, our probe starts at a base case of 1 px, preventing any risk of a step size that is too large.

W3: It is unclear whether DPO used the same rectangle edit space as SRD.

Yes, DPO indeed uses exactly the same rectangle edit space as SRD. The only difference is how candidates are proposed: DPO uses random Gaussian perturbations, while SRD uses steps along the surrogate gradient direction. We have updated the manuscript in Section 6 to make this distinction clearer.

W4: The experiment part should be further improved. For example, 1) some key baselines are missing, such as best-of-n CD samples. 2) Some key hyperparameters are also missing. 3) No wall-clock cost is reported.

Thank you for the suggestions. We provide additional results and details for each point, which have been added to the paper.

  1. Best-of-N

Best-of-n is, in fact, complementary to SRD rather than a competing baseline. Therefore, we added best-of-n baselines. Alongside this baseline, we allowed SRD refine the selected sample’s trajectory from nn samples. We draw nn unguided samples (nn simulator calls), rerun the best one with SRD on the last 2 denoising steps using k=3k = 3 simulator calls per step plus one call for the final board, and compare against CD best-of-BB at the same total budget B=n+2k+1=n+7B = n + 2k + 1 = n + 7.

On test sets the unguided samples are diverse, so best-of-n is a strong use of the first few simulator calls. This advantage, however, comes from a rapidly saturating source. At n=16n = 16, refining the best-of-16 winner achieves 0.1279, compared with 0.1371 for CD best-of-23 at the same 23 calls.

In a word, once independent sampling has exhausted its easy gains, further samples mostly rediscover designs of similar quality, while SRD improves the selected design in a way that resampling cannot. We have added this result to the main paper (Section 6.3, Table 3), and the full sweep over all budgets is in Appendix A.4 (Table 7).

Total budgetnbest-of-n + SRDCD best-of-(n+7)
810.25940.1671
920.21540.1637
1030.18770.1585
1140.17530.1556
1250.16450.1536
1360.15460.1518
1470.15000.1491
1580.14320.1473
1690.14190.1451
17100.13730.1441
18110.13470.1434
19120.13290.1418
20130.13220.1405
21140.13050.1402
22150.13020.1389
23160.12790.1371
  1. Hyperparameters

We have added these descriptions to the Appendix A.2.

DefinitionValue
number of denoising steps20 steps
Guided stepslast 2: steps 19–20 of 20; last 4: steps 17–20 of 20
Step size h1 px = 125 µm
Candidate step sizes±1, 2, 3, 5, 7 px
Improvement threshold ε_imp0.001
Random perturbationGaussian, σ = 2 px
GGD guidance weightw = 0.025, best of {0.025, 0.05, 0.1, 0.25, 0.5}
  1. wall-clock cost

We have added these results to the Appendix A.3.

MethodTest: Palace callsTest: time per design (s)OOD: Palace callsOOD: time per design (s)
CD121.9125.7
GGD126.4130.4
DPO779.417150.5
SRD (adaptive)7128.217259.4
SRD (directional)7128.217261.6
SRD (hybrid)11128.129249.6

Note that time per design includes diffusion sampling, CenterNet extraction, surrogate gradients as well as Palace calls. Calls that do not depend on each other run in parallel:

  • DPO: its random perturbations are independent, so all palace calls run in prarllel.
  • SRD: each guided step needs that the probe runs first, and the remaining candidates are then evaluated in parallel based on its outcome.

We have included all these additional results in the paper. Thanks again for the suggestion!

Q2: Please clarify whether any OOD targets were used to select the refinement settings and explain how these settings were chosen.

The primary change between the in-distribution setting and the out-of-distribution setting is that we extend simulator guidance from the final two steps to the final four steps. We found that this extra budget was more impactful in the out-of-distribution setting, as the conditional diffusion model often created candidates that were much farther from the target response in this case. Hence, strong simulator guidance was necessary to ensure strong generalization to the out-of-distribution cases. This behavior is expected; the model has only learned conditional signals for in-distribution examples, as exhibited by CD’s poor performance on the out-of-distribution set.

Q3: The authors are advised to explain under what conditions a single simulator probe can reliably guide the choice between directional and random search.

In our adaptive SRD variant, the policy for this is a simple heuristic. If the probe reduces the simulator loss by a given threshold, the method continues along uu. Otherwise, we revert back to random search. For our experiments, we selected a threshold of τ=10−3\tau = 10^{-3}: the direction is kept if a 1 px probe along uu lowers the simulator loss by at least τ\tau. We have added these results to the Appendix A.2.


Thank you, again, for your feedback. We hope that these clarifications have resolved the concerns presented in your review. We are happy to follow-up if any questions remain during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.