ASML通过机器学习开发用于半导体制造的虚拟计量技术

解决方案

Use MATLAB to create and train a neural network that predicts overlay metrology from alignment metrology

结果

建立了行业领导
潜在的制造改进已确定
Maintenance overhead minimized

“作为一名工艺工程师我没有经验不ural networks or machine learning. I worked through the MATLAB examples to find the best machine learning functions for generating virtual metrology. I couldn’t have done this in C or Python—it would’ve taken too long to find, validate, and integrate the right packages.”
Emil Schmitt-Weaver，ASML

Cutaway of a TWINSCAN and Track as wafers receive alignment and overlay metrology.

在纳米制作中，光刻是控制微芯片大小的基本图案步骤。在光刻期间，低波长的电源通过光学元件通过图像进行调节，然后将其尺寸降低，并以更多的光学功能降低到覆盖底物（通常为硅）的光敏化学薄膜中。重复此步骤，直到底物上的所有可用表面积都以相同的图像暴露为止。结果称为层。需要多个裸露的层来创建构成芯片的复杂显微镜结构。为了防止由于层之间的连接故障而引起的收益问题，层之间的所有模式都必须按预期排列。

为了确保不影响吞吐量的层对齐，ASML的Twinscan光刻系统必须限制其在曝光步骤之前测量的对齐标记的数量。一般规则是，测量对齐标记所需的时间不能超过序列中前一个晶圆所需的时间。由于适当的覆盖模型校正所需的大量覆盖标记，因此测量来自Twinscan系统的每个晶圆是不可行的。

ASML used MATLAB^®and Statistics and Machine Learning Toolbox™ to develop virtual overlay metrology software. This software applies machine learning techniques to come up with a predicted estimate of overlay metrology for every wafer, using alignment metrology data.

“The work we’ve done with MATLAB and machine learning demonstrates industry leadership in the best use of existing metrology,” says Emil Schmitt-Weaver, applications development engineer at ASML. “The papers we’ve published on this work have attracted the interest of customers looking to improve their manufacturing processes with ASML products.”

挑战

尽管错过的覆盖错误可能会降低产量，但大多数制造商仅占晶圆人群的24％的覆盖层。通过与Twinscan系统收集的每个晶圆的对齐计量学，ASML试图将机器学习技术应用于晶状体的覆盖层，并将其与现有的fardingstar计量学进行了比较。

由于Schmitt-Weaver以前没有开发机器学习算法的经验，因此他决定不开发Python，C或其他低级语言的算法。他想快速开发一个原型，依靠已在ASML大型，多样化的用户群中部署并由专业专业人员维护的功能。

解决方案

Schmitt-Weaver used MATLAB, Statistics and Machine Learning Toolbox, and Deep Learning Toolbox™ to develop a method for generating virtual metrology.

To start, Schmitt-Weaver used the Neural Network Time Series Prediction and Modeling app to learn how to prepare data for use with Deep Learning Toolbox. Using the app, he generated and exported the example code, which gave him a more detailed understanding of how the functions could be used together. As his competence increased, he was able to build on the generated code using examples from the vast multidisciplinary user community on MATLAB Central.

Schmitt-Weaver使用FARDSTAR系统收集了来自Twinscan系统和覆盖计量数据的对齐计量数据。然后，他将数据集分为两组，一组用于培训网络，另一个用于验证它。

Using Deep Learning Toolbox and Statistics and Machine Learning Toolbox, he designed a nonlinear autoregressive network with exogenous inputs (NARX) and trained it with data from the training group.

To avoid overfitting the neural network to the training group, he used Deep Learning Toolbox to implement automated regularization with a Bayesian framework.

After the network was trained, he supplied it with input from the test data and verified its results against the measured results from the YieldStar system.

ASML used the data collected to develop a prototype real-time overlay controller in MATLAB. The network provided the foundation for potentially improving yield, as well as the ability to identify wafers that might not have received overlay metrology.

结果

建立了行业领导。“By using MATLAB to improve overlay metrology, we showed our customers that we are leaders in developing innovative ways to reach their overlay performance goals,” says Schmitt-Weaver.
潜在的制造改进已确定。“The network we designed and trained in MATLAB identified systematic and random overlay errors that might otherwise have gone undetected,” notes Schmitt-Weaver. “This degree of improvement to overlay performance is necessary for microchip fabrication up to and below the 5nm node.”
Maintenance overhead minimized。Schmitt-Weaver说：“自从公司的开端以来，ASML系统就使用了编译的MATLAB算法。”“通过在经过同行评审的旧脚本的大型数据库上构建，我能够将注意力集中在新的机器学习功能上。”

“通过集成计量和机器学习支持的故障检测的虚拟叠加层计量学”，”金宝appProc。Spie9424，微观志xxix的计量，检查和过程控制，94241T（2015年3月19日），，doi：10.1117/12.2085475。

下载188bet金宝搏使用的产品

Learn more about ASML

Showcase Your Success

加入客户参考计划

Get started