纳米压印数据库系统设计与其在神经网络参数优化的应用

NANOIMPRINT LITHOGRAPHY DATABASE SYSTEM DESIGN AND ITS APPLICATION IN NEURAL NETWORKS PARAMETERS OPTIMIZATION

韩非

11849050

硕士

材料加工工程

程鑫

2020-05-27

2020-07-01

哈尔滨工业大学

深圳

纳米压印光刻技术是最有前景的纳米加工技术之一，作为下一代半导体光刻技术，具有低成本、高分辨率和高通量的特征，要实现其在纳米光电子学以及生物医学器件的应用，需要克服纳米压印的高缺陷率的问题。但是，由于纳米压印缺乏标准的工艺参数准则，在实际生产过程中，压印参数复杂且相互制约，难以实现图案的低缺陷率。近几十年来，计算机技术的飞速发展使“数据+人工智能”的模式成为材料研究的主流，采用数据科学的方法处理并分析数据正在加快材料研发的速度。基于以上研究背景，本课题旨在构建“纳米压印+大数据”的新型研究模式，构建纳米压印数据库管理系统，结合扫描电子显微镜与 Pro-SEM 软件对纳米压印后图形进行缺陷检测和分析，并构建以纳米压印材料参数和工艺参数为输入层，缺陷率、图案复制精度为输出层的多层神经网络。通过将纳米压印过程数据化，本研究将首次构建缺陷率与工艺参数、材料特性之间的定量关系，实现一定量级的纳米压印的数据库，从而实现纳米压印复制精度检测和缺陷率预测，为提高纳米压印的图形复制率和大面积图形化打下基础。具体的，本文提供了三种制备紫外压印模板的方法，并对实验所用的制备方法利弊进行了评价。第一种方法，结合光刻、刻蚀、蒸镀等技术制备了不同深宽比的硅模板，该模板涵盖了纳米压印的三种基本图案种类（圆柱、圆孔和光栅）。对于圆柱模板，圆柱平均直径为 1400.7 nm，高度分别为 1220 nm、2287 nm、3490nm；对于圆孔模板，圆孔平均直径为 2104.3 nm，深度分别为 1099 nm、2098 nm、2800 nm；对于光栅模板，光栅平均线宽为 835.9 nm，高度分别为 1219 nm、2087 nm、3011 nm；第二种方法，利用复型技术制备了能透紫外光的 UV-PDMS 软模板；第三种方法，利用紫外固化技术制备了能透紫外光的 OrmoStamp 硬质模板。另外，利用全自动纳米压印设备实现了圆柱、圆孔和光栅的紫外压印实验，并结合扫描电子显微镜和 Pro-SEM 软件开发了一种低成本的纳米压印参数量测和缺陷检测方法。通过对缺陷种类的分类，提供了一种纳米压印缺陷率的计算方法。本文基于纳米压印实验的需求，利用 Dephi 7.0 和 Microsoft Access 数据库建立了一个基于 Client/Server 架构的纳米压印数据库系统。数据库系统包括了模板参数管理、压印胶参数管理、压印工艺参数管理和系统权限管理 4 个模块，实现了模板参数、压印胶参数、压印工艺参数的增、删、改、差四个方面的管理功能，以及扫描电镜图像的查询和录入。该系统可以帮助科研机构和工业人员合理选择纳米压印参数，减少试错法带来的时间和资源浪费。本文最后对人工神经网络进行了概述，介绍了三种主要神经网络，包括 BP 神经网络（Back propagation neural network, BPNN）、广义回归神经网络以及概率神经网络，介绍了神经网络在参数优化方面的应用，按照 SI 标准单位，对模板参数、材料参数、工艺参数进行了标准化单位规定，同时对每个参数进行了计算机语言的转化，并利用文献中已有的压印参数训练了三层的 BP 神经网络，均方误差最小约为 0.00566，训练曲线的相关系数达 0.983，参数拟合效果符合参数优化的需求。

Nanoimprint lithography (NIL) is one of the most promising nanofabrication techniques. As the next generation semiconductor lithography technoloy, NIL has the characteristics of low cost, high throughout, and high resolution. It is necessary to realize its application in nano-electronics, nano-photonics and biomedical devices. However, there is no standard parameters guideline to transfer patterns with low defect. To obtain optimal NIL process parameters, it is often required to develop a reliable tool for predicting the critical dimension and the defect rate of the transferred patterns. In recent decades, the rapid development of computer technologgy has made the “data + artificial intelligence” become the mainstream of materials research. Using data science methods to process and analyze data is accelerating the speed of material research and development. Based on the above research background, aiming to build a new research mode of “nanoimprint + big data”, this research built a nanoimprint database management system, and combined scanning electron microscope and Pro-SEM software to detect printing defects. In the same time, the research built a multi-layer neural network with nanoimprint material parameters and process parameters as the input layer, defect rate and pattern fidelity as the output layer. By digitizing the nanoimprint process, the research built a quantitative relationship between the defect rate, process parameters, and material properties to achieve nanoimprint fidelity and defect rate detection. The study laid the foundation for improving the fidelity and mitigating the dfect rate of the nanoimprint.Specifically, this research provided three methods for fabricating UV imprint mold, and evaluated their advantages and disadvantages. The first method combines photolithography, etching, evaporation, and other techniques to fabricate silicon mold with different aspect ratio, including three patterns (pillar, hole and grating). As for the pillar mold, the diameter is 1400.7 nm, and the height is 1220 nm, 2287 nm, and 3490 nm respectively. As for the hole mold, the diameter is 2104.3 nm, and the depth is 1099 nm, 2098 nm, and 2800 nm respectively. As for the grating mold, the critical dimension is 835.9 nm, and the height is 1219 nm, 2087 nm, and 3011 nm. The second method used a replica technique to fabricate a transparent UV-PDMS soft mold. The third method used a UV curing technique to fabricate a transparent OrmoStamp hard mold. At the same time, the research used automatic nanoimprint machine to carry out UV imprint experiment, combining with scanning electron microscope and Pro-SEM to develop a low cost nanoimprint metrology and defect detection method. Through the classification of graphic types, a defect rate calculation method was provided. Additionally, based on the experimental design of NIL, a NIL database system based based on Client/Server architecture was established using Dephi 7.0 and Microsoft Access database. The database system includes four modules: mold parameter management, imprint resist management, imprint process parameter management and system authority management, which realized the addition, deletion, modification and difference of mold parameters, imprint resist parameters, and imprint process parameters, including the query and entry of scanning electron microscope images. The system can help scientific research institutions and industrial people to choose NIL parameters, reducing the time and resource waste caused by the trial and error method. Finally, this research concluded with an overview of artificial neural networks, introducing three neural networks, including BP neural network (Back propagation neural network, BPNN), generalized regression nerual network, and probabilistic neural network, introducing the application of neural networks in parameter optimization. In accordance with SI standard units, standardized unit specifications for mold parameters, material parameters, and process parameters were established. At the same time, each parameter was converted into computer language, and a BP-ANN mold was trained using the imprint parameters from literature. The minimum mean square error was about 0.00566, and the correlation coefficient reached 0.983, proving the parameters fitting effect satisfies the need of parameter optimization.

纳米压印光刻 数据库 人工神经网络 参数测量 缺陷检测

nanoimprint lithography database artificial neural network metrology defect inspection

中文

联合培养

南方科技大学

韩非. 纳米压印数据库系统设计与其在神经网络参数优化的应用[D]. 深圳. 哈尔滨工业大学,2020.