acceptodds
Under review as a conference paper at ICLR 2027

Diffusion-Guided Elite Path Inpainting for 3D Electromagnetic Coil Initialization

Abstract

Electromagnetic coil design is a representative physical inverse-design problem: finding conductor geometries that achieve desired physical performance under geometric constraints. It underpins applications including power transfer, imaging, and induction heating. Although automatic synthesis is established for prescribed supports and parameterizations, many optimization workflows still require a manually supplied initial coil. We address this initialization gap by generating geometrically feasible candidate routes with high performance potential under a limited simulation budget. Our framework combines dual-view action diffusion with performance-preference-guided elite path reconstruction. Sequence and spatial views capture route order and three-dimensional occupancy, respectively. Quality-weighted residual learning from evaluated paths guides reconstruction while constraining deviations from a frozen geometry base. Hard depth-first search filters invalid actions and backtracks from dead ends; complete-path validation precedes evaluation and synchronous elite-archive updates. The archive maintains a pool of promising parents so that later rounds can build on evaluated candidates. We define a benchmark spanning symmetric enclosing, asymmetric enclosing, and non-enclosing scenes, three label budgets, and five repetitions. Averaged over five runs, our method leads the four learning baselines in Top-5@100 and Mean@100 across all nine settings, and in Best@100 in seven.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
diffusion models, generative design, simulation-based optimization, constrained path generation, electromagnetic coil design

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_YHpJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a new generation framework for initializing a design of 3D electromagnetic coil. This paper introduced a tailored structure of diffusion model for the problem, and sampling method (sec 4.5). Training data is composed by random (bad data) and simulations (good data). The model is pretrained with the former, and further trained with the latter. Experimental results show that the proposed method consistently outperforms all four learning-based baselines.

Critical strengths

This paper addresses a niche yet potentially high-impact problem.

Critical weaknesses

A primary concern is that, is this really a deep-learning problem? I am somewhat not familiar with this domain (3D electromagnetic coil design), but the voxel grids used in the study (e.g., 11 x 10 x 6) seem small enough to be explored via non deep-learning methods. This paper lacks a rigorous comparison to show why a complex diffusion model should be used, in terms of runtime, computational cost, or generalizability. I also do not see novelties on methodologies that can be useful in other problems, although each modules are reasonably employed. Moreover, the training data themselves are generated by simulations. Then, why do we need NN to mimic the dataset? The paper fails to provide a compelling reason for it. Can model generate a new sample that is effectively not in the dataset?

Questions for authors
  1. Why do you need to build NN even though you have a training dataset? 1-1. Can the method generalize in a different configuration (from training dataset)? 1-2. Is computational cost significantly smaller than traditional methods? 1-3. Can you compare the computational costs of generating training dataset with sampling with your method? 1-4. Can model generate a new sample that is effectively not in the dataset?

  2. Is there any possibilities or necessities of scaling-up (so that traditional / handcraft methods can not easily be applied)?

6 comments
AuthorsOpenReview ↗

Responses to Reviewer YHpJ [1/2]

We appreciate your thoughtful questions about the role of learning in coil design. They helped us explain the motivation for our approach more clearly. We would like to address your questions below.


Q1. A primary concern is that, is this really a deep-learning problem? I am somewhat not familiar with this domain (3D electromagnetic coil design), but the voxel grids used in the study (e.g., 11 x 10 x 6) seem small enough to be explored via non deep-learning methods.

A1. Yes, non-deep-learning methods can explore these voxel grids, but the enormous search space makes efficient optimization challenging. Even an 11×10×611 \times 10 \times 6 grid has approximately 108210^{82} possible paths.

To evaluate the benefit of learned guidance under limited evaluation budgets, we compared against non-deep-learning methods GA [A] and PSO [B] using the same 100 labeled paths and 100 additional physical evaluations per scene. Ours outperforms both across all scenes and metrics, demonstrating the effectiveness of learned guidance under equal evaluation budgets. We have clarified in Section 1 that this count refers to unconstrained action sequences and reported the equal-budget GA and PSO comparison in Section 5.1 and Appendix G.2 (Table 22).

Table 1: Comparison with non-deep-learning baselines (five-run mean ± sample SD; higher is better ↑).

SceneMethodBest@100 ↑Top-5@100 ↑Mean@100 ↑
Our proposed scene RhombusGA−58.29±13.91−69.70±6.66−80.51±0.57
PSO−43.89±3.17−46.88±2.79−80.12±2.26
Ours−37.10±2.37−38.28±1.02−52.49±1.86
Our proposed scene StaggeredGA−39.50±21.64−67.04±6.09−84.05±0.36
PSO−22.82±4.94−39.40±3.67−86.29±2.32
Ours−15.44±1.32−16.69±2.03−34.09±0.80
Our proposed scene YGA−70.93±2.59−76.33±1.95−83.37±0.21
PSO−62.28±3.09−68.86±2.57−83.83±0.11
Ours−47.66±3.86−50.94±3.00−66.43±1.07

[A] Adaptation in Natural and Artificial Systems. MIT Press, 1992.

[B] Particle swarm optimization. ICNN, 1995.


Q2. I also do not see novelties on methodologies that can be useful in other problems, although each modules are reasonably employed.

A2. Our methodological contribution extends beyond coil initialization by integrating action-level inpainting, performance-preference guidance, and hard-constrained search. This framework can potentially be adapted to other constrained design tasks, such as cable or pipe routing, by modifying the action representation, geometric constraints, and evaluation objectives. We have clarified the method contribution in Section 1 and discussed the cable- and pipe-routing examples in Section 6 under “Generalization.”


Q3. Why do you need to build NN even though you have a training dataset?

A3. Our neural network aims to generate higher-quality initializations beyond training samples to improve downstream optimization efficiency, rather than simply reuse existing training samples. We compared Best-of-Training and Ours using the same downstream optimizer CMA-ES[A], with 300 labeled paths and 200 additional physical evaluations per scene (200 for refinement vs. 100 for generation + 100 for refinement). Ours achieves higher mean final scores across all three scenes, demonstrating the benefit of learned generation over directly refining the best training sample under the same evaluation budget. We have reported the two allocations of the 200-evaluation budget and their final scores in Section 5.3 (Table 4).

Table 2. Post-refinement best scores (five-run mean ± SD; higher is better).

MethodOur proposed scene RhombusOur proposed scene StaggeredOur proposed scene Y
Best-of-Training + CMA-ES−38.45 ± 0.01−7.27 ± 1.37−54.37 ± 0.46
Ours + CMA-ES−9.47 ± 3.64−5.33 ± 2.20−38.21 ± 5.97

[A] Completely Derandomized Self-Adaptation in Evolution Strategies. Evolutionary Computation, 2001.

AuthorsOpenReview ↗

Responses to Reviewer YHpJ [2/2]

Q4. Can the method generalize in a different configuration (from training dataset)?

A4. Yes. Our framework can be adapted to new coil configurations and broader constrained path-design tasks. For a new scene, feasible paths can be generated through geometric checks without physical simulation, supporting geometry pretraining, while a small labeled subset enables preference guidance and elite initialization. The core architecture and constrained search procedure remain unchanged. We have described the data needed to adapt the framework to a new coil configuration in Section 6 under “Generalization.”


Q5. 1-2. Is computational cost significantly smaller than traditional methods? 1-3. Can you compare the computational costs of generating training dataset with sampling with your method?

A5. Our method achieves runtime comparable to GDP and substantially lower than 3D-CDM and predictor-based baselines, though higher than Random-DFS, which we use to generate unlabeled training paths. We measured the time to generate 100 valid, unique candidate paths on our proposed scene Staggered using a single NVIDIA RTX 3090 GPU. Results are reported as the mean ± sample standard deviation over five runs. Timing covers the full online pipeline, including loading, generation, failed DFS attempts, backtracking, retries, deduplication, screening, elite updates, and output processing. For fairness, physical evaluation time was excluded for all methods. We have clarified how the training paths and labels are obtained in Section 3 under “Data and budget,” and reported the online sampling comparison in Section 5.1 and Appendix G.1 (Table 21).

Table 3. Online sampling time (seconds; lower is better). Predictor-based methods incur additional costs from large-pool generation and screening (10,000 paths); 3D-CDM requires iterative voxel denoising and constrained decoding.

MethodTime (s) ↓
Random-DFS46.17 ± 2.42
GDP88.40 ± 9.82
Pathformer-Predictor3657.53 ± 50.95
Voxel-Predictor3605.89 ± 44.32
3D-CDM202.88 ± 3.27
Ours92.18 ± 2.71

Q6. Can model generate a new sample that is effectively not in the dataset?

A6. Yes, our method explicitly rejects duplicates of labeled and previously accepted paths during inference. To further verify novelty, we compared all 4,500 generated candidates against each scene's 10,000-path catalog and labeled set. No complete-path duplicates were found, confirming that our method generates novel paths beyond the existing dataset. We have clarified the inference-time duplicate checks against labeled paths and previously accepted candidates in Section 4.5.


Q7. Is there any possibilities or necessities of scaling-up (so that traditional / handcraft methods can not easily be applied)?

A7. Yes, larger routing domains, finer voxel grids, and more complex obstacles motivate scaling beyond traditional methods. Our framework can be extended by adapting scene inputs, action representations, and training data while retaining preference-guided reconstruction and hard geometric constraints. Larger-scale applications may require additional training data and computational resources. We have discussed the effects of grid resolution, routing-domain size, and obstacle density in Section 6 under “Scalability.”


We sincerely hope our clarifications above have addressed your questions.

AuthorsOpenReview ↗

Looking Forward to the Response from Reviewer YHpJ

Dear Reviewer YHpJ,

We express our gratitude for your valuable feedback aimed at enhancing our work. We have provided detailed responses addressing your initial concerns.

We look forward to more discussions with you if you have any outstanding concerns or questions.

Best regards,

The Authors

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

This paper proposes a diffusion-guided framework for 3D electromagnetic coil initialization. It combines dual-view diffusion, physics-preference guidance, and elite path reconstruction with hard geometric constraints. Experiments across three scenes and three label budgets show consistent improvements in candidate quality over learning-based baselines.

Critical strengths
  1. Constraint-aware generation. Integrating learned action guidance with hard DFS backtracking ensures geometric feasibility, addressing a key challenge in physical inverse design.

  2. Consistent empirical improvements. The method achieves the best Top-5@100 and Mean@100 across all nine settings among the learning-based baselines, demonstrating improved candidate-pool quality.

Critical weaknesses
  1. Incomplete attribution of the gains to the learning components. Table 3 provides evidence in favor of the learned reconstruction components jointly, since Full outperforms Random Reconstruction. However, Random Reconstruction removes geometry pretraining and physics-preference guidance together, while w/o Elite Search removes both initial elite selection and subsequent parent-pool updates. These comparisons therefore leave important component effects entangled. A geometry-pretrained model without preference guidance, together with a simpler non-diffusion reconstruction policy evaluated within the same elite-DFS framework, would more directly establish the incremental value of the proposed learning design. This matters because the current results support the overall framework more clearly than they isolate its central learning contributions.

  2. Limited evidence for downstream initialization utility. The paper explicitly limits its evaluation to initialization potential, rather than claiming downstream refinement gains. Nevertheless, its motivation is to provide useful starting points for subsequent optimization. The current pre-refinement metrics do not establish whether higher-scoring initial candidates lead to better final designs, a higher rate of satisfying the physical thresholds, or lower refinement cost under a common optimizer and budget.

Questions for authors
  1. Were the non-ablated training and search settings matched between Full and the Table 3 variants, particularly the window size, DFS budget, and guidance strength?
  2. Can the authors separately ablate geometry pretraining and physics-preference guidance, and compare the diffusion policy with a simpler non-diffusion policy under the same elite-DFS framework and budget?
  3. Under the same downstream optimizer and refinement budget, do the proposed initializers achieve better final scores or higher physical-threshold satisfaction rates than the baselines?
  4. What is the wall-clock cost of generating 100 valid unique candidates, including failed DFS attempts, compared with the key baselines?
5 comments
AuthorsOpenReview ↗

Responses to Reviewer fgET [1/2]

Thank you for your careful reading of our experiments. Your suggestions helped us identify where additional comparisons were needed. We would like to address your questions below.


Q1. Were the non-ablated training and search settings matched between Full and the Table 3 variants, particularly the window size, DFS budget, and guidance strength?

A1. Yes. Full and the Table 3 variants share W=4, up to two window attempts and 250 DFS expansions per attempt, as well as data splits, seeds, applicable training hyperparameters, and decoding settings. Guidance is κ=1, except Random Reconstruction (κ=0); w/o Elite Search alters parent selection and archive updates. We have specified the shared search budget in Section 5.2 and detailed the guidance and parent-selection settings in Appendix E.1, including κ=0 for the new Geometry-only variant.


Q2. Can the authors separately ablate geometry pretraining and physics-preference guidance, and compare the diffusion policy with a simpler non-diffusion policy under the same elite-DFS framework and budget?

A2. Yes, we conducted four additional ablations to isolate the contributions of individual components under the same data, search, and evaluation settings on our proposed scene Staggered:

  • E1, Initial Elite Only: Fixes the initial elite parent pool; no updates.
  • E2, Geometry-only: Uses the pretrained geometry base without preference guidance (κ=0).
  • E3, w/o Geometry Pretraining: Trains residual heads on a frozen, randomly initialized diffusion base.
  • E4, AR-Transformer: Uses autoregressive action prediction with the same dual-view inputs, guidance, and elite-DFS.

The Full method achieves the highest mean across all three metrics over five runs, supporting the contributions of geometry pretraining, physics-preference guidance, diffusion, and dynamic elite updates in this setting. We have added the four variants and their results to Section 5.2 (Table 3) and reported their definitions and per-seed results in Appendix E.1 (Table 11).

Table 1. Ablation results on our proposed scene Staggered (higher is better). E1–E4 are newly added variants.

MethodBest@100Top-5@100Mean@100
Full−15.44 ± 1.32−16.69 ± 2.03−34.09 ± 0.80
Random Reconstruction−21.51 ± 1.07−22.59 ± 0.93−41.66 ± 0.65
Sequence-only−17.35 ± 3.60−19.39 ± 3.11−37.73 ± 3.16
Spatial-only−16.07 ± 2.05−17.21 ± 2.24−37.34 ± 2.03
w/o Elite Search−34.36 ± 5.72−45.73 ± 4.16−90.20 ± 3.38
E1: Initial Elite Only−23.87 ± 1.04−27.33 ± 2.35−54.44 ± 0.74
E2: Geometry-only−17.36 ± 2.60−18.63 ± 2.60−36.38 ± 1.24
E3: w/o Geometry Pretraining−18.18 ± 4.35−21.03 ± 3.08−39.51 ± 1.17
E4: AR-Transformer−16.47 ± 2.93−18.94 ± 3.79−37.56 ± 2.95

Q3. Under the same downstream optimizer and refinement budget, do the proposed initializers achieve better final scores or higher physical-threshold satisfaction rates than the baselines?

A3. Yes. Ours achieves the highest mean post-refinement score in all three scenes. We applied CMA-ES[A] to each method's best candidate from its n=300 pool with five paired seeds and 100 physical evaluations per run. All methods share the objective, population size (16), and normalized initial step size (0.2). Constraint-aware updates preserve endpoints, segment directions, obstacle avoidance, and wire spacing. We have reported the common refinement protocol and the final scores for all six initializers in Section 5.3 (Table 4).

Table 2. Post-refinement best scores across three scenes (higher is better).

InitializerOur proposed scene RhombusOur proposed scene StaggeredOur proposed scene Y
Ours−9.47 ± 3.64−5.33 ± 2.20−38.21 ± 5.97
3D-CDM−24.33 ± 13.68−13.80 ± 1.77−44.71 ± 5.77
GDP−38.03 ± 1.08−14.15 ± 3.07−50.72 ± 5.96
Pathformer−34.28 ± 1.64−15.00 ± 3.13−39.82 ± 8.23
Random-DFS−38.55 ± 0.06−11.47 ± 1.78−53.61 ± 5.96
Voxel Predictor−31.46 ± 7.57−15.21 ± 2.92−42.49 ± 2.06

[A] Completely Derandomized Self-Adaptation in Evolution Strategies. Evolutionary Computation, 2001.

AuthorsOpenReview ↗

Responses to Reviewer fgET [2/2]

Q4. What is the wall-clock cost of generating 100 valid unique candidates, including failed DFS attempts, compared with the key baselines?

A4. Our method generates 100 valid, unique new candidates in 92.18 ± 2.71 seconds, comparable to GDP and faster than 3D-CDM and predictor baselines. We measured the inference time on our proposed scene Staggered using a single NVIDIA RTX 3090 GPU. Results are reported as the mean ± sample standard deviation over five runs. Timing covers the full online pipeline, including model loading, path generation, failed DFS attempts, backtracking, retries, deduplication, screening, elite updates, and output processing. For a fair comparison, physical evaluation time was excluded for all methods. We have added the timing comparison in Section 5.1 and reported the full online timing protocol and results in Appendix G.1 (Table 21).

Table 3. Online inference time for 100 valid, unique candidates (seconds). Predictor-based methods incur additional costs from large-pool generation and screening (10,000 paths); 3D-CDM requires iterative voxel denoising and constrained decoding.

MethodTime (s)
Random-DFS46.17 ± 2.42
GDP88.40 ± 9.82
Pathformer-Predictor3657.53 ± 50.95
Voxel-Predictor3605.89 ± 44.32
3D-CDM202.88 ± 3.27
Ours92.18 ± 2.71

We sincerely hope our clarifications above have addressed your questions.

AuthorsOpenReview ↗

Looking Forward to the Response from Reviewer fgET

Dear Reviewer fgET,

We express our gratitude for your valuable feedback aimed at enhancing our work. We have provided detailed responses addressing your initial concerns.

We look forward to more discussions with you if you have any outstanding concerns or questions.

Best regards,

The Authors

Reviewer_1MoNOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an action-level diffusion-inpainting based approach for automating electromagnetic coil initialization. It does so in two stages. In first stage, it learns a diffusion transformer model to predict routing actions for reconstructing the missing segment of coil design by using dual (sequential and spatial) views of the partially observed coil geometry; this stage is used to generate geometrically valid coil initializations. The second stage then performs quality-weighted residual learning to learn residuals which adjust the predictions from (frozen, first stage) geometry base to generate coil initializations that give better physical performance. The inference process then uses these, along with depth-first search which filters invalid actions and performs backtracking, to generate a pool of candidate coil initializations that are geometrically valid as well as physically performant. The proposed method is shown to be better than the baselines considering three coil design scenes, three simulation-label budgets, and three metrics (Top-5@100, Mean@100, Best@100).

Critical strengths

It is interesting how the concept of diffusion-inpainting is utilized to solve a practical engineering problem: automate the initialization for electromagnetic inverse design and reduce human effort.

Critical weaknesses
  1. In Section 4.5, the authors mention “The diffusion policy scores and orders this hard-legal set”. This is confusing because from Section 4.3, we see that z_base provides logits for 4x|A| i.e. the entire action set. This also gives rise to another question: was the hard-legal set concept not applied during training? Currently, it is not clear whether it was applied. Given that we know most geometric constraints during training itself (e.g., stay within voxel grid, obstacles), can the action set not be constrained accordingly or the information be made available to the diffusion model in some way e.g. through channels in S_j? Could this potentially help improve the prediction of the diffusion model and reduce the inference time? The latter is doubtful given some constraints may only be known during inference? A discussion on these aspects would help remove such confusion.

  2. Given that the proposed method includes multiple steps in inference (e.g. diffusion guided DFS, backtracking, multiple reconstruction attempts, physical evaluation to update elite set), having a computational efficiency comparison with the baselines (e.g. overall time taken for inference of 100 random paths) would be useful.

  3. The current Ablation shows “w/o Elite Search” as removing both initial elite selection and subsequent parent-pool updates. An additional ablation with “only initial elite selection” will be useful, mainly because if there isn’t much difference between “only initial elite selection” vs “with both initial elite selection and subsequent parent-pool updates”, users of this method can choose to leave out subsequent parent-pool updates for potentially better computational efficiency.

Questions for authors
  1. Scope: how common is it to need a new physical metric to judge performance or are the current three exhaustive enough? In case we need to include a new performance metric, can we just add a new residual head, and train just the new head and the official score head while using pretrained networks for other components? Including a discussion on this will help increase the strength of the method.

  2. Why 14 channels for S_j and shape of (50x19) in Q_j is not clear.

4 comments
AuthorsOpenReview ↗

Responses to Reviewer 1MoN [1/2]

Thank you for your thoughtful comments. We would like to address your questions below.


Q1. In Section 4.5, the authors mention “The diffusion policy scores and orders this hard-legal set”. This is confusing because from Section 4.3, we see that z_base provides logits for 4x|A| i.e. the entire action set. This also gives rise to another question: was the hard-legal set concept not applied during training? Currently, it is not clear whether it was applied. Given that we know most geometric constraints during training itself (e.g., stay within voxel grid, obstacles), can the action set not be constrained accordingly or the information be made available to the diffusion model in some way e.g. through channels in S_j? Could this potentially help improve the prediction of the diffusion model and reduce the inference time? The latter is doubtful given some constraints may only be known during inference? A discussion on these aspects would help remove such confusion.

A1. No, the hard-legal set is not applied during training. Our design intentionally separates distribution learning from constraint enforcement: unsupervised geometry pretraining aims to learn the distribution of feasible paths, while Depth-First Search(DFS) ensures geometric validity by ranking only legal actions using the predicted logits. Incorporating constraints during training is a possible extension to this design. We have clarified the training vocabulary in Section 4.3, the inference-time legal-action checks in Section 4.5, and the rationale for this design in Section 6 under “Geometric constraints during training and inference.”


Q2. Given that the proposed method includes multiple steps in inference (e.g. diffusion guided DFS, backtracking, multiple reconstruction attempts, physical evaluation to update elite set), having a computational efficiency comparison with the baselines (e.g. overall time taken for inference of 100 random paths) would be useful.

A2. Our method achieves runtime comparable to GDP and substantially lower than 3D-CDM and the predictor-based baselines. We measured the inference time required to generate 100 candidate paths in our proposed scene Staggered using a single NVIDIA RTX 3090 GPU. All results are reported as the mean ± sample standard deviation over five runs. The reported time covers the entire inference pipeline, including path generation, unsuccessful DFS attempts, backtracking, retries, deduplication, screening, elite updates, and output processing. For a fair comparison, physical evaluation time was excluded for all methods. We have added the timing comparison in Section 5.1 and reported the timing protocol and all six results in Appendix G.1 (Table 21).

Table 1. Online inference time (seconds). Predictor-based methods incur additional costs from large-pool generation and screening (10,000 paths), while 3D-CDM requires iterative voxel denoising and constrained decoding.

MethodTime (s)
Random-DFS46.17 ± 2.42
GDP88.40 ± 9.82
Pathformer-Predictor3657.53 ± 50.95
Voxel-Predictor3605.89 ± 44.32
3D-CDM202.88 ± 3.27
Ours92.18 ± 2.71
AuthorsOpenReview ↗

Responses to Reviewer 1MoN [2/2]

Q3. The current Ablation shows “w/o Elite Search” as removing both initial elite selection and subsequent parent-pool updates. An additional ablation with “only initial elite selection” will be useful, mainly because if there isn’t much difference between “only initial elite selection” vs “with both initial elite selection and subsequent parent-pool updates”, users of this method can choose to leave out subsequent parent-pool updates for potentially better computational efficiency.

A3. We conducted the suggested Initial Elite Only ablation to isolate the contribution of subsequent parent-pool updates. This variant retains initial elite selection, guided diffusion, and hard DFS, but keeps the initial parent pool fixed throughout the search, preventing newly evaluated candidates from becoming parents. We followed the same ablation settings on our proposed scene Staggered, evaluating 100 new candidates per run. Results are reported as the mean ± sample standard deviation over five runs. The Full method consistently outperforms Initial Elite Only across all three metrics, demonstrating the benefit of dynamic elite updates beyond initial selection. We have added Initial Elite Only to Section 5.2 (Table 3) and reported the variant definition and per-seed results in Appendix E.1 (Table 11).

Table 2. Ablation comparison on our proposed scene Staggered; higher is better.

ConfigurationBest@100 ↑Top-5@100 ↑Mean@100 ↑
Full−15.44 ± 1.32−16.69 ± 2.03−34.09 ± 0.80
Random Reconstruction−21.51 ± 1.07−22.59 ± 0.93−41.66 ± 0.65
Sequence-only−17.35 ± 3.60−19.39 ± 3.11−37.73 ± 3.16
Spatial-only−16.07 ± 2.05−17.21 ± 2.24−37.34 ± 2.03
w/o Elite Search−34.36 ± 5.72−45.73 ± 4.16−90.20 ± 3.38
Initial Elite Only (new)−23.87 ± 1.04−27.33 ± 2.35−54.44 ± 0.74

Q4. Scope: how common is it to need a new physical metric to judge performance or are the current three exhaustive enough? In case we need to include a new performance metric, can we just add a new residual head, and train just the new head and the official score head while using pretrained networks for other components? Including a discussion on this will help increase the strength of the method.

A4. The three physical metrics are not exhaustive, as additional objectives may be needed depending on the application. Our modular framework allows incorporating a new metric by adding a residual head and retraining the official-score head, while keeping the pretrained geometry base and existing metric heads frozen. This extension would require labels for the new metric, along with corresponding updates to the quality weights, head fusion, and elite selection strategy. We have described how to incorporate a new metric in Section 6 under “Adapting to new physical objectives.”


Q5. Why 14 channels for S_j and shape of (50x19) in Q_j is not clear.

A5. The input dimensions follow from the feature breakdown in the tables below: (S_j) has 14 spatial channels, while (Q_j) has 19 features per row, zero-padded to 50 rows. We have clarified the input dimensions in Section 4.2 and listed the spatial channels, per-waypoint features, and padding rule in Appendix A.3 (Tables 6 and 7).

Table 3. Spatial input (S_j) (14 channels).

FeatureChannels
Prefix/suffix occupancy2
Signed x/y/z directions6
Endpoint/anchor indicators2
Normalized coordinates3
Endpoint–anchor bounding box1
Total14

Table 4. Sequence input (Q_j) (50 × 19).

FeatureDimensions
Coordinates3
Incoming direction6
Run length1
Waypoint counts, progress, observed dense-edge count4
Prefix/anchor flags2
Waypoint index1
Window-normalized remaining actions1
Capacity-normalized remaining actions1
Total per row19

We sincerely hope our clarifications above have addressed your questions.

Thank you for the revisions and additional experiments. I acknowledge the clarification that the hard-legal action set is applied only during inference and that the incorporation of such constraints during training is considered as a possible extension to the design. The additional runtime comparison, the new Initial Elite Only ablation, and the discussions on extending the framework to new physical objectives help address my concerns.

AuthorsOpenReview ↗

Thank you for appreciating our Responses!

Dear Reviewer 1MoN,

We are glad to know that our clarifications, additional experiments, and discussions have helped address your concerns. We sincerely appreciate your constructive feedback and would like to thank you again for your time and recognition of our efforts to improve the paper!

Best,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Responses

Dear ACs and Reviewers,

Thank you for your constructive suggestions and thoughtful questions. To complement our replies to each reviewer, we would like to express our appreciation for your recognition of our work and summarize the main revisions we made in response to your feedback.

1. We appreciate the strengths identified in the reviews.

  • Practical relevance and potential impact

    • “It is interesting how the concept of diffusion-inpainting is utilized to solve a practical engineering problem: automate the initialization for electromagnetic inverse design and reduce human effort.” [Reviewer 1MoN]
    • “This paper addresses a niche yet potentially high-impact problem.” [Reviewer YHpJ]
  • Geometric feasibility and physical performance

    • “Integrating learned action guidance with hard DFS backtracking ensures geometric feasibility, addressing a key challenge in physical inverse design.” [Reviewer fgET]
    • “The inference process then uses these, along with depth-first search which filters invalid actions and performs backtracking, to generate a pool of candidate coil initializations that are geometrically valid as well as physically performant.” [Reviewer 1MoN]
  • Candidate quality across learning-based baselines

    • “The method achieves the best Top-5@100 and Mean@100 across all nine settings among the learning-based baselines, demonstrating improved candidate-pool quality.” [Reviewer fgET]
    • “Experimental results show that the proposed method consistently outperforms all four learning-based baselines.” [Reviewer YHpJ]
  • Joint contribution of the learned reconstruction components

    • “Table 3 provides evidence in favor of the learned reconstruction components jointly, since Full outperforms Random Reconstruction.” [Reviewer fgET]

2. We have made the following revisions, highlighted in blue in the manuscript.

  • Downstream refinement. With 300 labeled paths, we refined the best candidate from each of six reported pools using the same CMA-ES optimizer and 100 refinement evaluations. Our method achieved the highest mean final score in all three scenes over five paired runs. It also led in all three scenes when compared with refining the best labeled path directly. Both approaches received 200 new evaluations. Our method used 100 for generation and 100 for refinement, while the direct baseline used all 200 for refinement. On our proposed scene Staggered, the corresponding mean final scores were −5.33 and −7.27. Section 5.3, Table 4. [Reviewers fgET and YHpJ]

  • Classical search. Given the same 100 labeled paths and 100 new physical evaluations, our method exceeded both a genetic algorithm and particle swarm optimization on all three candidate-pool metrics in every scene, averaged over five runs (Section 5.1 and Appendix G.2, Table 22). Both baselines use constrained DFS. [Reviewer YHpJ]

  • Candidate diversity. Section 4.5 clarifies how we reject duplicates during search. The separate audit reported in our response to Reviewer YHpJ found that none of the 4,500 generated candidates exactly matched a complete path in its scene’s geometry catalog or labeled set. Typical pairs shared 12–20% of their combined grid edges. Power, nonuniformity, and inductance varied within pools (Appendix G.3, Tables 23–24). [Reviewer YHpJ]

  • Contributions of the individual components. On our proposed scene Staggered, four new ablations tested the effects of fixing the initial elite pool, disabling preference guidance, omitting geometry pretraining, and replacing diffusion with an autoregressive Transformer. The full method had the highest five-run mean on all three metrics. In particular, allowing newly evaluated paths to become parents improved Top-5@100 from −27.33 with fixed initial elites to −16.69. Section 5.2, Table 3, and Appendix E.1, Table 11. [Reviewers 1MoN and fgET]

  • Online sampling cost. On our proposed scene Staggered, our method generated 100 valid, unique new candidates in 92.18 seconds on average, compared with 88.40 for GDP, 46.17 for Random-DFS, and 202.88 for 3D-CDM. These five-run means include failed attempts, backtracking, and elite updates where applicable. We excluded physical evaluation time for all six methods to ensure a fair comparison of online sampling times. Section 5.1 and Appendix G.1, Table 21. [Reviewers 1MoN, fgET, and YHpJ]

  • Method and scope. We clarified that training predicts over the full action vocabulary, while DFS ranks only legal actions during inference (Sections 4.3 and 4.5). Section 4.2 and Appendix A.3 detail the two input representations. Section 6 explains how a new physical objective could be learned with a new residual head, what data a new coil configuration would require, and how the approach might extend to larger grids or to cable and pipe routing. [Reviewers 1MoN and YHpJ]

Thank you again for the suggestions that shaped these experiments and clarifications.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.