acceptodds
Withdrawn from ICLR 2027

TopoBalance: Diverse Legal Layout Generation via Support Expansion and Distribution Balancing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generating large, diverse, and distributionally balanced libraries of legal layout patterns is important for design-for-manufacturability applications, yet remains challenging with limited legal training data. The combinatorial topology space leads to sparse coverage of complexity labels, while topology filtering and physical-feasibility constraints introduce acceptance-induced distribution shifts that concentrate the final legal library on a limited subset of labels. To address these challenges, we propose TopoBalance, a distribution-aware layout-pattern generation framework that integrates training-stage support expansion and generation-stage distribution balancing. TopoBalance employs Topology-Valid RegSLERP (TV-RegSLERP) for training-stage support expansion and Distribution-Balanced Adaptive Sampling and Generation (DBAS-Gen) for closed-loop distribution balancing during generation. Experiments on ICCAD 2016 show that TopoBalance substantially improves complexity-label diversity and distributional balance while maintaining a high legality. It markedly reduces concentration on dominant complexity regions, yielding a broader and more balanced legal-layout library. The code and datasets are available at https://anonymous.4open.science/r/RegSLERP-Hybrid-5BB1/.

Primary area:
generative models
Keywords:
Layout pattern generation, Diffusion models, Topology-aware augmentation, Distribution-aware sampling, Physical legalization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_TMYrOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This manuscript proposes a framework for generating VLSI layout patterns, using TV-RegSLERP and DBAS-Gen for training and balancing. The experimental results demonstrate that the framework substantially reduces the concentration of generated patterns in dominant complexity regions, producing a more diverse and balanced library of legal layouts. However, its practical utility remains unclear. Industry already has extensive collections of layout patterns that reflect the specific characteristics and constraints of lithographic processes. There is no assurance that the artificially generated patterns preserve these characteristics or satisfy the relevant manufacturing requirements. Therefore, industrial validation is needed to substantiate the framework’s practical applicability.

Critical strengths

None.

Critical weaknesses

Although the framework generates a more diverse and balanced library of legal layout patterns, the manuscript does not demonstrate that these patterns preserve the process-specific characteristics and satisfy the lithographic constraints required in industrial applications. Layout legality alone does not establish manufacturability or practical usefulness. Moreover, the added value of the generated patterns over the extensive pattern collections already available in industry remains unclear. Validation using industrial datasets and lithography-aware evaluation is therefore needed to substantiate the framework’s practical applicability.

Questions for authors

Do you have industrial validation to substantiate the framework’s practical applicability?

Reviewer_h1CiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduce a distribution-aware framework for legal layout-pattern generation under limited training data. Two techniques are: 1) expanding complexity-label support through topology-valid cross-complexity latent interpolation, 2) adaptively coordinating complexity-label sampling, layout generation and legalization, and feedback update to balance the evolving legal-layout distribution.

Critical strengths
  1. The paper is on an important problem with distribution aware layout-pattern generation, which is important to evaluate the quality of generation.
  2. Generally, the paper is well written and easy to follow, The two techniques TV-RegSLERP and DBAS-Gen is well motivated and the experiments is extensive.
Critical weaknesses
  1. The two proposed TV-RegSLERP and DBAS-Gen appear to be primarily engineering refinements, and their relatively shallow integration limits the paper’s methodological novelty.
  2. In Table 4, the w/o TV-RegSLERP variant performs only slightly worse than the full model, suggesting that TV-RegSLERP makes a limited contribution.
Questions for authors
  1. What constitutes a “good” generated layout: one that matches the training distribution, approaches a chosen target distribution, or meets application-specific needs? Could the authors explain how the evaluation metrics reflect this objective?
  2. What role does this type of layout generation play in practical applications, and how do the paper’s contributions address the problems encountered in those applications?
  3. Does DBAS-Gen substantially increase inference time? Please report its computational overhead relative to generation without DBAS-Gen.
Reviewer_ZNhyOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TopoBalance, which extends the complexity label coverage through topologically effective latent space interpolation during the training phase, and adaptively adjusts the sampling based on the distribution of valid samples during the generation phase, significantly improving the distribution balance of the layout sample library while maintaining a high compliance rate.

Critical strengths

1.The problem defined in this paper is innovative; 2.The experimental results shows some kind of effectiveness.

Critical weaknesses

1.The conducted evaluation can only can prove 'more balanced coverage', but cannot alone prove 'wider coverage, richer topology, or more effective downstream'. 2.The tested benchmarks and used dataset is old and limited; 3.Novelty in the proposed methodology is limited

Questions for authors

what is the different betwe the two 'overall' and 'DBAS-GEN' items in table 4?

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

Related papers

Discussion (0)

Sign in to comment.