acceptodds
Under review as a conference paper at ICLR 2027

LithoGDV: General Visual Understanding Meets Domain-Specific Visual Priors for Robust Lithography Hotspot Detection

Abstract

Lithography hotspot detection (LHD) identifies local integrated circuit layout patterns susceptible to manufacturing defects. As technology nodes shrink, denser layouts and narrower process windows make detection increasingly sensitive to subtle geometric variations. To address this challenge, mainstream LHD methods rasterize polygonal layout clips into images and apply vision models for hotspot detection. These detectors, however, are trained on limited domain data and rely on learned parametric decision boundaries without explicitly reusing validated layout cases. Consequently, they may generalize poorly to unseen designs and miss hotspot patterns that differ only subtly from non-hotspots. We propose LithoGDV, built on the insight that robust LHD requires complementary knowledge: general visual understanding of layout geometry and domain-specific visual priors from comparable cases. LithoGDV fine-tunes a pretrained large vision model to learn fine-grained layout representations and constructs a knowledge base of process-verified cases. Retrieved hotspot and non-hotspot cases provide explicit evidence that complements the model's parametric prediction. On OpenHSDBench, LithoGDV achieves state-of-the-art detection accuracy against the leading open-source and non-open-source baselines across full-data, low-resource, and cross-design settings. We will publicly release the complete LithoGDV codebase, including all training and evaluation configurations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Lithography Hotspot Detection, Large Vision Model, Case‑Based Retrieval, Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_m6qSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LithoGDV, a lithography hotspot detection framework that integrates two complementary sources of visual knowledge: general visual understanding of layout geometry and domain-specific visual priors derived from comparable cases. LithoGDV achieves state-of-the-art hotspot detection accuracy compared with existing baselines, but exhibits relatively high false-alarm rates.

Critical strengths
  1. The paper presents a reasonable motivation for integrating general visual understanding with domain-specific visual priors for lithography hotspot detection. The combination of an adapted large vision model and a case-based knowledge base provides a meaningful way to leverage both learned representations and historical layout evidence. The experimental results support the effectiveness of this design.
  2. The paper evaluates the proposed method under full-data, low-resource, and cross-design settings. LithoGDV achieves the highest hotspot detection accuracy across all six benchmark datasets in the full-data setting. The ablation studies also provide empirical evidence for the contributions of the main components, strengthening the experimental support for the proposed framework.
Critical weaknesses
  1. Although LithoGDV achieves state-of-the-art hotspot detection accuracy, it exhibits substantially higher false-alarm rates than several baselines in the full-data setting. While the authors acknowledge this trade-off, the paper lacks a systematic evaluation across different operating points, making it difficult to determine whether the improvement in detection accuracy reflects better discrimination or a more aggressive operating point.
  2. Although the paper demonstrates the efficiency of graph-based retrieval, it does not provide a systematic comparison with other learning-based methods in terms of end-to-end inference latency, GPU memory consumption, and overall training cost. This limits the assessment of the computational overhead and practical deployment efficiency of the proposed framework.
  3. The authors state that GDSII textual semantics can be readily incorporated into the proposed framework and may further improve detection performance, but provide no implementation or experimental evidence to substantiate this claim. Given the trade-off between detection accuracy and false-alarm rates observed in the full-data setting, it remains unclear whether incorporating textual information could improve this trade-off while preserving the current detection accuracy advantage. Although multimodal integration is presented as future work, the claimed ease of integration and potential benefits remain unverified.
Questions for authors
  1. Could the authors provide a failure-case analysis, particularly focusing on high-confidence false positives in the full-data setting? Additionally, could they report detection accuracy at fixed false-alarm rates, or provide ROC curves, to better characterize the trade-off between AccHS and FAR?
  2. Table 8 demonstrates the efficiency of the graph-based retrieval module. Could the authors additionally report end-to-end inference latency, training time, and GPU memory consumption, preferably with comparisons against representative learning-based baselines?
  3. The paper states that GDSII textual semantics can be readily integrated into LithoGDV and may further improve detection performance. Is there any preliminary implementation or experimental evidence supporting this claim? In particular, could textual information help reduce false alarms without sacrificing the current advantage in hotspot detection accuracy?
  4. The proposed retrieval mechanism relies on visual representations adapted through supervised contrastive learning. To what extent does similarity in this learned feature space correspond to similarity in lithography-related manufacturing risk? Could the authors discuss or provide evidence for the reliability of this relationship, particularly under different process conditions or technology nodes?
Reviewer_RCWjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed LithoGDV, an image-based LHD framework that integrates general visual understanding with domain-specific visual priors. The framework adapts a pretrained large vision model using low-rank adaptation and supervised contrastive learning to extract fine-grained layout representations. Experiments on OpenHSDBench evaluate the framework under full-data, low-resource, and cross-design settings and demonstrated relatively good performance.

Critical strengths
  1. The low-resource and cross-design results are a good strength of this paper, as demonstrated in both the main contributions and the experimental tables.
  2. The application of pretrained vision models in IC layout hotspot detection is a good and rising research direction, and this paper studied this promising direction.
  3. The evaluation performance of LithoGDV seems to be good in comparison other learning-based LHD baselines.
Critical weaknesses
  1. As a paper that is closely related to modern CV technicals, this paper provides limited visual examples showing how retrieved cases help detect hotspots. Examples of query layouts, retrieved cases, and correct or incorrect predictions could be very helpful throughout the paper to help us understand geometric properties.
  2. The false-alarm control is a problem. In Table 2, LithoGDV produces more false alarms on several datasets, especially on Via-4. More discussion and comparisons are needed here.
  3. The technical novelty from the computer vision side seems limited. The framework mainly combines already existing methods, like pretrained vision models, LoRA, and some supervised contrastive learning.
Questions for authors
  1. Minor comment: the caption of figure 1 and 2 could be improved to be more informative about the detailed components of the figure. I think this is important because these are the main method demonstration figure.
  2. Minor comment: Table 2 is extremely hard to read. And the number font seems to be different than the other tables as well. I would recommend fixing that.
Reviewer_CK5eOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This work proposed LithoGDV, which is used to detect errors during manufacturing. The authors first use a pre-trained VLM to extract features, and then build a library of hotspot and non-hotspot examples. For a new layout, this model will search similar historical cases and judge the potential errors by combining these features. The results show that this method achieves great results.

Critical strengths
  1. This work aims to address an important problem that can reduce errors before manufacturing.
  2. Comprehensive experiments and good performance.
Critical weaknesses
  1. The technical novelty is limited. This work utilizes a lot of modules and the contributions of these modules. A recent work, "Streamlining Computational Lithography With Efficient Pattern Database," should be included as a reference. It is better to clarify the differences and compare the results.

  2. The performance is not stable. On dataset Via-1, LithoGDV improves recall from 95% to 97%, but the FAR increases from 28% to 51%. The same trend is observed on Via-4 and the average. A clearer analysis is needed to explain.

Questions for authors

For each added module, could you please do an ablation study to confirm the effectiveness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.