| 引用本文: | 肖鑫忠, 马瑞君, 徐 辉, 黄文馨, 袁 野.基于改进 Ghost 的半监督光刻热点检测方法(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2026，43（2）：108-115 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 摘要: |
| :目的 鉴于传统的半监督光刻热点检测方法逐渐无法满足集成电路制造对检测精度的要求,且难以解决因
数据集不平衡引起的精度损失问题,提出一种新的半监督光刻热点检测模型 GSSL。 方法 在该模型中,将卷积注意
力模块(Convolutional Block Attention Module, CBAM)引入到 Ghost 模块的线性变化中,设计了 Ghost_CBAM 模块;
将该模块与压缩激励网络(Squeeze-and-Excitation, SE)结合设计了 GhostNeck 模块,实现特征图先降维再升维,建
立各个通道之间的关联性;再通过 GhostNeck 构建整个光刻热点检测模型 GSSL,实现逐步引入无标记数据进入训
练的半监督学习方式;通过集成数据增强方法对数据集中的热点版图进行数据增强,缓解数据不平衡问题;并应用
加权交叉熵损失函数,进一步提升模型对于热点类别的关注度。 结果 在 ICCAD( The International Conference on
Computer-Aided Design) 2012 竞赛基准数据集上进行评估,在标记数据占比为 10% ~ 50%的情况下预测热点的平
均准确率为 91. 73%,平均误报为 680。 结论 与其他传统方法相比,GSSL 可以有效应对数据集不平衡的问题,提升
光刻热点检测精度的同时,显著降低了误报率,在光刻热点检测上具有一定的应用价值。 |
| 关键词: 光刻热点检测 集成电路 半监督学习 改进 Ghost 模块 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Semi-supervised Lithography Hotspot Detection Based on Improved Ghost |
|
XIAO Xinzhong MA Ruijun XU Hui HUANG Wenxin YUAN Ye
|
|
School of Computer Science and Engineering Anhui University of Science and Technology Huainan 232001 Anhui
China
|
| Abstract: |
| Given that traditional semi-supervised lithography hotspot detection methods are gradually unable to
meet the precision requirements for integrated circuit manufacturing and has difficulty addressing accuracy loss caused by
dataset imbalance a new semi-supervised lithography hotspot detection model GSSL is proposed. Methods In this
model the Ghost_CBAM was designed by introducing the convolutional block attention module CBAM into the linear
change of the Ghost module. The GhostNeck module was designed by combining the Ghost _CBAM module with the
squeeze-and-excitation SE network which enabled dimensionality reduction followed by dimensionality expansion of
feature maps and established correlations between channels. Then the whole lithography hotspot detection model GSSL
was constructed by GhostNeck to realize the semi-supervised learning method of gradually introducing unlabeled data into
the training. Data augmentation methods were integrated to augment hotspot layouts in the dataset to alleviate dataset
imbalance. Besides a weighted cross-entropy loss function was applied to further improve the model?? s focus on hotspot
categories. Results The model was evaluated on the benchmark dataset of the 2012 International Conference on
Computer-Aided Design ICCAD competition achieving an average accuracy of 91. 73% and an average false alarm of
680 when the proportion of labeled data ranged from 10% to 50%. Conclusion Compared with other traditional methods GSSL adeptly tackles the challenge of imbalanced datasets improving accuracy in lithography hotspot detection while
significantly reducing false alarms. Therefore this study holds considerable application value in lithography hotspot
detection |
| Key words: lithography hotspot detection integrated circuit semi-supervised learning improved Ghost module |
