基于深度学习的端到端掩模优化任务

End-to-end Mask Optimization Task Based on Deep Learning

目的 针对光刻系统与特征尺寸不匹配导致的光刻图案与掩模图案严重偏差的问题,提出了一个基于深度学习的端到端掩模优化框架TransU-ILT。方法 该框架使用CNN-Transformer的混合模型作为特征提取模块提取目标布局的深度特征,在特征重构模块中加入像素重组层来重构掩模;此外,在训练过程中,加入深度监督机制提高对布局图案特征的提取精度,从而进一步提高掩模的可印刷性。结果 实验定量结果表明:与最先进的方法相比,所提出的框架可以实现4倍的周转时间加速,在平方L_(2)误差和工艺变化带指标方面分别降低了13.4%和4.3%,且框架生成的掩模晶圆图案边缘更加平滑,更接近目标布局。结论 TransU-ILT在时间性能和掩模可印刷性方面总体上优于对比的先进方法,可以为掩模优化方法提供一种有效的解决方案。

Objective Aiming at the problem of serious deviation between lithography pattern and mask pattern caused by the mismatch between lithography system and feature size,an end-to-end mask optimization framework TransU-ILT based on deep learning was proposed.Methods The framework used the CNN-Transformer hybrid model as a feature extraction module to extract the depth features of the target layout,and added a pixel reorganization layer to the feature reconstruction module to reconstruct the mask.In addition,in the training process,the depth supervision mechanism was added to improve the extraction accuracy of layout pattern features,so as to further improve the printability of the mask.Results Quantitative experimental results showed that compared with the most advanced methods,the proposed framework achieved a 4X turnaround time acceleration,reduced the square L_(2) error and the process change band index by 13.4% and 4.3%,respectively,and the wafer pattern edge of the mask generated by the framework was smoother and closer to the target layout.Conclusion TransU-ILT is superior to other advanced methods in terms of time performance and mask printability,which can provide an effective solution for mask optimization methods.