Cascaded machine learning model for reconstruction of surface topography from light scattering
Cascaded machine learning model for reconstruction of surface topography from light scattering
复制标题
DOI：
10.1117/12.2555035
复制
发表时间：
2020-04
期刊：
影响因子：
--
通讯作者：
Mingyu Liu;N. Senin;Rong Su;R. Leach
中科院分区：
文献类型：
--
作者：
Mingyu Liu;N. Senin;Rong Su;R. Leach
光散射法是一种很有前途的在线表面测量方法。许多研究人员已经研究了用于评估表面纹理的光散射方法。散射测量的研究人员已经开发出通过求解逆散射问题来获得表面纹理参数的方法。然而，大多数研究仅集中在纹理测量或特征尺寸小于光源波长的临界尺寸的确定上。在本文中，我们提出了一种新的光散射方法来重建光栅图案的表面形貌，使用级联机器学习模型。实验散射信号可以作为输入被馈送到机器学习模型中，并且表面形貌可以被确定为输出。训练数据集，即不同表面的散射信号，通过基于边界元法（BEM）的验证严格的表面散射模型生成。通过这种方式，可以使用包括数万个数据集的大数据方法来训练机器学习模型，这些数据集代表了真实的情况下的大多数场景。级联机器学习模型被设计为使用神经网络实现的组合自顶向下的两层模型。第一层由分类模型组成，该分类模型被设计为在一组预定义的设计变量中确定正在测量的结构化表面的类型。第二层包含回归模型，该回归模型被设计成确定限定已被识别的结构化表面的特定类型的设计参数的值，例如其周期性特征的标称节距和高度。我们已经开发了一个原型系统，并进行了实验，以验证所提出的方法。考虑了包含光栅图案的结构化表面，并分析了不同类型的光栅。通过与原子力显微镜进行的测量进行比较，结果进行了验证。
Light scattering methods are promising for in-process surface measurement. Many researchers have investigated light scattering methods for evaluating surface texture. Researchers working on scatterometry have developed methods to derive surface texture parameters by solving the inverse scattering problem. However, most of the research has been focused only on texture measurement or determination of critical dimensions where feature sizes are less than the wavelength of the light source. In this paper, we propose a new light scattering method to reconstruct the surface topography of grating patterns, using a cascaded machine learning model. The experimental scattering signal can be fed into the machine learning model as the input and the surface topography can be determined as the output. The training dataset, i.e. scattering signals of different surfaces, are generated through a validated rigorous surface scattering model based on a boundary element method (BEM). In this way, the machine learning model can be trained using a big data approach including tens of thousands of datasets, which represent most of the scenarios in real cases. The cascaded machine learning model is designed as a combined top-down, two-layer model implemented using neural networks. The first layer consists of a classification model designed to determine which type of structured surface is being measured, amongst a set of predefined design variants. The second layer contains a regression model, designed to determine the values of the design parameters defining the specific type of structured surface which has been identified, for example the nominal pitch and height of its periodic features. We have developed a prototype system and conducted experiments to verify the proposed method. Structured surfaces containing grating patterns were considered, and different types of gratings were analysed. The results were validated by comparison with measurements performed with atomic force microscopy.