一种频域自注意力机制引导的多尺度逆光刻技术

罗斌玲; 王盈; 蔡述庭

doi:10.11999/JEIT251382

一种频域自注意力机制引导的多尺度逆光刻技术

doi: 10.11999/JEIT251382 cstr: 32379.14.JEIT251382

罗斌玲^{1, 2},
王盈¹,
蔡述庭^1, ,

1.
广东工业大学集成电路学院广州 510006
2.
广西水利电力职业技术学院自动化工程学院南宁 530013

基金项目: 广东省科技计划(2022B0701180001)

详细信息

作者简介:
罗斌玲：女，博士生，研究方向为电子设计自动化

王盈：女，博士生，研究方向为电子设计自动化

蔡述庭：男，教授，研究方向为电子设计自动化

通讯作者:
蔡述庭　shutingcai@gdut.edu.cn

中图分类号: TN47; TP301
计量
- 文章访问数: 12
- HTML全文浏览量: 8
- PDF下载量: 0
- 被引次数: 0
出版历程
- 收稿日期: 2025-12-30
- 修回日期: 2026-05-14
- 录用日期: 2026-05-14
- 网络出版日期: 2026-06-03

A Frequency Domain Self-Attention Guided Multi-Scale Inverse Lithography Technology

LUO Binling^{1, 2},
WANG Ying¹,
CAI Shuting^{1
, ,}

1.
School of Integrated Circuits, Guangdong University of Technology, Guangzhou 510006, China
2.
Guangxi Vocational College of Water Resources and Electric Power, NanNing 530013, China

Funds: Guangdong S&T Programme (2022B0701180001)

摘要

摘要: 光刻制程中的光学邻近效应(OPE)会导致晶圆上的打印图像偏离设计版图，需要对光刻掩膜进行光学邻近校正(OPC)。该文提出一种频域自注意力机制引导的多尺度逆光刻技术(FMS-ILT)用于掩膜优化。FMS-ILT采用基于残差卷积的多尺度编码器-解码器结构，通过残差连接融合局部细节与全局特征，并在编码器末端引入频域自注意力机制，将动态注意力与频域全局建模能力相结合，自适应关注影响打印图像质量的关键信息。在预训练阶段，模型通过掩膜生成与目标图像重构增强物理一致性；在主训练阶段，进一步结合光刻仿真对掩膜进行精细优化。实验结果表明，FMS-ILT的平均$ \mathcal{L}2 $损失较基准模型降低2%–107%，平均边缘放置误差(EPE)降低47%–1115%，显示了FMS-ILT方法在优化掩膜方面的显著优势。
- 掩膜优化 /
- 逆光刻 /
- 频域自注意力 /
- 多尺度
Abstract: Objective Optical Proximity Effects (OPE) in lithographic processes cause printed patterns on wafers to deviate from target layouts, necessitating Optical Proximity Correction (OPC) through mask optimization prior to exposure. Traditional rule-based OPC methods suffer from significant accuracy degradation when handling complex layouts, while model-based OPC approaches incur high computational cost. In recent years, deep learning--based methods have been introduced to accelerate mask generation; however, their limited receptive fields hinder effective modeling of long-range optical interference effects, thereby constraining optimization accuracy. To address these challenges, this work proposes a Frequency Domain Self-Attention Guided Multi-Scale Inverse Lithography Technology (FMS-ILT), which jointly models local geometric details and global optical interactions, leading to improved printed image fidelity, edge placement accuracy, and process robustness. Methods FMS-ILT adopts a residual convolution--based multi-scale encoder--decoder architecture, where shallow layers extract fine-grained geometric features such as edges and corners, while deeper layers capture large-scale layout context. Residual blocks and multi-level skip connections are employed to preserve high-frequency information and stabilize training. To overcome the limited receptive field of spatial convolutions, a Frequency Domain Self-Attention Mechanism (FSAM) is introduced at the encoder output. Global feature interactions are enabled via the Fourier transform, and the resulting attention responses are mapped back to the spatial domain through the inverse Fourier transform to adaptively reweight feature representations. A two-stage training strategy is adopted. During pretraining, a dual-branch structure is used to jointly learn mask geometry and imaging consistency, providing physically meaningful initialization. During main training, lithography simulation is applied under nominal, maximum, and minimum process conditions to further refine mask optimization under physical constraints. Results and Discussions The comparison results with baseline models are summarized in Tables 2 and 3. Our method is set as the reference (Ratio = 1), and all experiments are conducted on the LithoBench dataset. In terms of overall imaging $ \mathcal{L}2 $ error, our method achieves the lowest value of 19,998, outperforming baseline models by 2%–107%. For the process robustness metric Process Variation Band (PVB), GAN-OPC obtains the best result of 19,156, which is 31% lower than ours; however, its $ \mathcal{L}2 $ error and EPE are 107% and 1115% higher, respectively, indicating an imbalance between imaging fidelity and edge accuracy. The remaining baseline models exhibit PVB performance comparable to ours. Regarding Edge Placement Error (EPE), our method also demonstrates a significant advantage, achieving an average EPE of 1.95, which is 47%–1115% lower than the baselines. These improvements can be attributed to three key factors: (1) a multi-scale encoder–decoder fusion mechanism that effectively integrates local and global features, (2) the combination of attention mechanisms and frequency-domain operations to guide the model toward critical regions, and (3) a dual-branch pretraining strategy that injects physical priors into the network. With these modules jointly contributing, FMS-ILT achieves more balanced and superior performance in imaging fidelity, process stability, and edge accuracy. Conclusions This work proposes a Frequency Domain Self-Attention Guided Multi-Scale Inverse Lithography Technology (FMS-ILT). The model adopts a residual convolution--based multi-scale encoder--decoder architecture to extract rich spatial features and incorporates a frequency-domain self-attention mechanism to jointly model local geometric details and global optical interference characteristics. A two-stage training strategy is employed. In the pretraining stage, a dual-branch task of mask generation and target image reconstruction is used to enhance the physical consistency between the mask and the printed image. In the main training stage, lithography simulation is introduced to further improve imaging accuracy and process robustness. Experimental results on the public LithoBench dataset demonstrate that FMS-ILT achieves superior performance in terms of $ \mathcal{L}2 $, PVB, and EPE metrics, effectively improving printed image quality and providing a feasible and efficient solution for computational lithography.
- Mask optimization /
- Inverse lithography /
- Frequency Domain Self-Attention /
- Multi-Scale

HTML全文

图 1 EPE畸变示意图

下载: 全尺寸图片幻灯片

图 2 一种频域自注意力机制引导的多尺度逆光刻技术框架

下载: 全尺寸图片幻灯片

图 3 残差卷积编码器

下载: 全尺寸图片幻灯片

图 4 对特征张量作通道平均并投射到掩膜(PM)和投射到目标图像(PT)作可视化，编码层2关注小尺度特征如金属边缘和角点，编码层5关注全局布局与长程光学干涉，而FSAM进一步精准聚焦局部细节与全局关键特征

下载: 全尺寸图片幻灯片

图 5 残差卷积解码器

下载: 全尺寸图片幻灯片

图 6 (a)表明加入目标图重构分支后，掩膜生成损失的收敛趋势依旧成立，(b)中消融实表明移除目标图重构后模型损失升高

下载: 全尺寸图片幻灯片

图 7 频域自注意力机制FSAM结构图

下载: 全尺寸图片幻灯片

图 8 本文方法生成的打印图像和掩膜

下载: 全尺寸图片幻灯片

表 1 LithoBench 掩模优化数据集统计

	MetalSet	ViaSet	StdMetal	StdContact
训练集	14,824	104,733	0	163
测试集	1,648	11,642	271	165

下载: 导出CSV

表 2 掩膜优化对比结果

数据集	GAN-OPC ^[16]				Neural-ILT ^[17]				DAMO^[18]				本文方法
数据集	$ \mathcal{L}2 $	PVB	EPE	时间(s)	$ \mathcal{L}2 $	PVB	EPE	时间(s)	$ \mathcal{L}2 $	PVB	EPE	时间(s)	$ \mathcal{L}2 $	PVB	EPE	时间(s)
Metalset	43414	41200	8.7	0.070	36670	42666	7.3	0.050	32579	41173	5.4	0.055	32384	40142	3.6	0.066
ViaSet	14767	6686	8.3	0.160	12723	8537	6.2	0.059	5081	9962	0	0.162	4526	7936	0	0.062
StdMetal	25929	23715	4.6	0.004	20045	23548	2.4	0.005	16120	23796	0.2	0.006	17886	22738	1.3	0.009
StdContact	81378	4931	73.2	0.003	25422	41537	3.2	0.003	50445	35673	26.7	0.004	25683	40234	2.9	0.007
平均值	41372	19156	23.7	0.059	23715	29072	4.8	0.029	26056	27651	8	0.057	19998	27763	1.95	0.036
比率	2.07	0.69	12.15	1.63	1.19	1.05	2.46	0.81	1.30	1.00	4.10	1.57	1	1	1	1
数据集	CFNO^[19]				MultiILT ^[20]				本文方法
数据集	$ \mathcal{L}2 $	PVB	EPE	时间(s)	$ \mathcal{L}2 $	PVB	EPE	时间(s)	$ \mathcal{L}2 $	PVB	EPE	时间(s)
Metalset	47814	46131	6.2	0.061	27272	43304	2.2	-	32384	40142	3.6	0.066
ViaSet	8949	9890	0.1	0.139	5357	9179	0.6	-	4526	7936	0	0.062
StdMetal	26809	26814	4.2	0.010	13814	24928	0.03	-	17886	22738	1.3	0.009
StdContact	70740	17960	55.1	0.006	34813	39997	8.6	-	25683	40234	2.9	0.007
平均值	38578	25196	18	0.054	20314	29352	2.86	-	19998	27763	1.95	0.036
比率	1.93	0.91	9.23	1.49	1.02	1.06	1.47	-	1	1	1	1

下载: 导出CSV

表 3 消融实验对比结果

模型	MetalSet			StMetal
模型	$ \mathcal{L}2 $	PVB	EPE	$ \mathcal{L}2 $	PVB	EPE
(-)目标图重构模块	46366	47416	9.0	19155	22123	2.4
(-)残差连接	48700	39967	17.4	23874	22974	4.4
(-)频域自注意力模块	36536	41123	6.3	20217	24812	1.9
完整模型	32384	40142	3.6	17396	22738	1.3

下载: 导出CSV

参考文献(25)

[1]	MACK C A. Corner rounding and line-end shortening in optical lithography[C]. Proceedings of SPIE 4226, Microlithographic Techniques in Integrated Circuit Fabrication II, Singapore, Singapore, 2000: 83–92. doi: 10.1117/12.404843.
[2]	CHOU D and MCALLISTER K. Line end optimization through optical proximity correction (OPC): A case study[C]. Proceedings of SPIE 6154, Optical Microlithography XIX, San Jose, USA, 2006: 61543A. doi: 10.1117/12.651455.
[3]	STEWART M D, SCHMID G M, POSTNIKOV S V, et al. Mechanistic understanding of line-end shortening[C]. Proceedings of SPIE 4345, Advances in Resist Technology and Processing XVIII, Santa Clara, USA, 2001: 10–18. doi: 10.1117/12.436844.
[4]	YU Ziyang, ZHENG Su, ZHAO Wenqian, et al. RuleLearner: OPC rule extraction from inverse lithography technique engine[J]. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2025, 44(5): 1915–1927. doi: 10.1109/TCAD.2024.3499909.
[5]	SUZUKI M, KIMURA T, INOUE J, et al. A novel pre-processing approach for OPC using LSMR[C]. Proceedings of SPIE 13425, DTCO and Computational Patterning IV, San Jose, USA, 2025: 1342510. doi: 10.1117/12.3050121.
[6]	CHEN Guojin, YANG Haoyu, REN H M, et al. Differentiable edge-based OPC[C]. Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design (ICCAD), Newark, USA, 2024: 47. doi: 10.1145/3676536.3676764.
[7]	HOPKINS H H. On the diffraction theory of optical images[J]. Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, 1953, 217(1130): 408–432. doi: 10.1098/rspa.1953.0071.
[8]	GOODMAN J W. Introduction to Fourier Optics[M]. 3rd ed. Englewood: Roberts & Company Publishers, 2005: 124–146.
[9]	ZHENG Su, XIAO Gang, YAN Ge, et al. Model-based OPC extension in OpenILT[C]. 2024 2nd International Symposium of Electronics Design Automation (ISEDA), Xi'an, China, 2024: 568–573. doi: 10.1109/ISEDA62518.2024.10617951.
[10]	ZHANG Shengen, MA Xu, HUANG Chaojun, et al. Model-driven optical proximity correction via hypergraph convolutional neural networks and its experimental demonstration[J]. Optics & Laser Technology, 2025, 183: 112199. doi: 10.1016/j.optlastec.2024.112199.
[11]	GIBOU F, FEDKIW R, and OSHER S. A review of level-set methods and some recent applications[J]. Journal of Computational Physics, 2018, 353: 82–109. doi: 10.1016/j.jcp.2017.10.006.
[12]	ZHENG Su, LIANG Xiaoxiao, YU Ziyang, et al. Curvilinear optical proximity correction via cardinal spline[C]. 2025 62nd ACM/IEEE Design Automation Conference (DAC), San Francisco, USA, 2025: 1–7. doi: 10.1109/DAC63849.2025.11133367.
[13]	SHEN Yijiang and ZHANG Zhenrong. Variational level-set formulation for lithographic source and mask optimization[J]. Journal of Microelectronic Manufacturing, 2018, 1(2): 18010203. doi: 10.33079/jomm.18010203.
[14]	HUANG Kai, ZHANG Xiaomeng, ZHENG Dandan, et al. A scalable and adaptable ILP-based approach for task mapping on MPSoC considering load balance and communication optimization[J]. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2019, 38(9): 1744–1757. doi: 10.1109/tcad.2018.2859400.
[15]	ZHENG Su, YANG Haoyu, ZHU Binwu, et al. LithoBench: Benchmarking AI computational lithography for semiconductor manufacturing[C]. Proceedings of the 37th International Conference on Neural Information Processing Systems, New Orleans, USA, 2023: 1315. doi: 10.52202/075280-1315.
[16]	YANG Haoyu, LI Shuhe, DENG Zihao, et al. GAN-OPC: Mask optimization with lithography-guided generative adversarial nets[J]. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2020, 39(10): 2822–2834. doi: 10.1109/TCAD.2019.2939329.
[17]	JIANG Bentian, LIU Lixin, MA Yuzhe, et al. Neural-ILT: Migrating ILT to neural networks for mask printability and complexity co-optimization[C]. 39th IEEE/ACM International Conference on Computer-Aided Design, San Diego, USA, 2020: 20. doi: 10.1145/3400302.3415704.
[18]	CHEN Guojin, CHEN Wanli, MA Yuzhe, et al. DAMO: Deep agile mask optimization for full chip scale[C]. 39th IEEE/ACM International Conference on Computer-Aided Design, San Diego, USA, 2020: 19. doi: 10.1145/3400302.3415705.
[19]	YANG Haoyu and REN Haoxing. Enabling scalable AI computational lithography with physics-inspired models[C]. Proceedings of the 28th Asia and South Pacific Design Automation Conference, Tokyo, Japan, 2023: 715–720. doi: 10.1145/3566097.3568361.
[20]	SUN Shuyuan, YANG Fan, YU Bei, et al. Efficient ILT via multi-level lithography simulation[C]. 2023 60th ACM/IEEE Design Automation Conference (DAC), San Francisco, USA, 2023: 1–6. doi: 10.1109/DAC56929.2023.10247704.
[21]	阮东升, 施哲彬, 王嘉辉, 等. 结合跨模态特征激励与双分支交叉注意力融合的左心房疤痕分割方法[J]. 电子与信息学报, 2025, 47(5): 1596–1608. doi: 10.11999/JEIT240775. RUAN Dongsheng, SHI Zhebin, WANG Jiahui, et al. Left atrial scar segmentation method combining cross-modal feature excitation and dual branch cross attention fusion[J]. Journal of Electronics & Information Technology, 2025, 47(5): 1596–1608. doi: 10.11999/JEIT240775.
[22]	RONNEBERGER O, FISCHER P, and BROX T. U-Net: Convolutional networks for biomedical image segmentation[C]. 18th International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), Munich, Germany, 2015: 234–241. doi: 10.1007/978-3-319-24574-4_28.
[23]	ZHU Binwu, ZHENG Su, MA Yuzhe, et al. Bridging hotspot detection and mask optimization via domain-crossing masked layout modeling[J]. ACM Transactions on Design Automation of Electronic Systems, 2025, 30(4): 54. doi: 10.1145/3728468.
[24]	古天龙, 张清智, 李晶晶. 基于时-频注意力机制网络的水声目标线谱增强[J]. 电子与信息学报, 2024, 46(1): 92–100. doi: 10.11999/JEIT230217. GU Tianlong, ZHANG Qingzhi, and LI Jingjing. Line spectrum enhancement of underwater acoustic targets based on a time-frequency attention network[J]. Journal of Electronics & Information Technology, 2024, 46(1): 92–100. doi: 10.11999/JEIT230217.
[25]	王开正, 曾瑶, 张占喜, 等. FCSNet: 频域感知的跨特征融合烟雾分割网络[J]. 电子与信息学报, 2025, 47(7): 2320–2333. doi: 10.11999/JEIT241021. WANG Kaizheng, ZENG Yao, ZHANG Zhanxi, et al. FCSNet: A frequency-domain aware cross-feature fusion network for smoke segmentation[J]. Journal of Electronics & Information Technology, 2025, 47(7): 2320–2333. doi: 10.11999/JEIT241021.