Application of Machine Learning Method to Model-Based Library Approach to Critical Dimension Measurement by CD-SEM

Aug 5, 2026

Upload by Ivy

The model-based library (MBL) method enables accurate critical dimension (CD) measurement of semiconductor linewidth from CD-SEM images, and is further advanced here by combining CD-SEM image simulation with neural-network machine learning. Monte Carlo simulation first computes secondary-electron linescan profiles and their dependence on key geometric parameters for Si and Au trapezoidal lines. Machine learning then predicts linescan profiles from randomly selected training sets. Predicted profiles match simulated profiles closely, with 0.1% standard deviation for relative error in Si and 6% for Au, supporting practical library reduction, faster database construction, and richer MBL content.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages30

Reading Time1 h 16 min

AccessFree to download

Abstract

1 Introduction

Semiconductor scaling and CD measurement demand

CD measurement techniques and limitations

Model-based library (MBL) approach

Monte Carlo simulation and database matching

Methods and evaluation (implied)

Application of Machine Learning Method to Model-Based Library Approach to Critical Dimension Measurement by CD-SEM P. Guo 1, H. Miao 1, Y.B. Zou2\#, S.F. Mao3 and Z.J. Ding4 * 1 CAS Key Laboratory ofMechanical Behavior and Design of Materials, Department of Modern Mechanics, University of Science and Technology of China, Hefei, Anhui 230027, China 2School of Physics and Electronic Engineering, Xinjiang Normal University, Urumqi, Xinjiang 830054, China 3Department ofEngineering and Applied Physics, University of Science and Technology of China, Hefei, Anhui 230026. China 4Department ofPhysics and Hefei National Laboratoryfor Physical Sciences at Microscale, University of Science and Technology of China, Hefei, Anhui 230026, China \#e-mail: [zyb0617@mail.ustc.edu.cn](zyb0617@mail.ustc.edu.cn) *[e-mail: zjding@ustc.edu.cn](e-mail: zjding@ustc.edu.cn) Abstract The model-based library (MBL) method has already been established for the accurate measurement of critical dimension (CD) of semiconductor linewidth from a critical dimension scanning electron microscope (CD-SEM) image. In this work the MBL method has been further investigated by combing the CD-SEM image simulation with a neural network algorithm. The secondary electron linescan profiles were calculated at first by a Monte Carlo simulation method, enabling to obtain the dependence of linescan profiles on the selected values of various geometrical parameters (e.g., top CD, sidewall angle and height) for Si and Au trapezoidal line structures. The machine learning methods have then been applied to predicate the linescan profiles from a randomly selected training set of the calculated profiles. The predicted results agree very well with the calculated profiles with the standard deviation of 0.1% and 6% for the relative error distributions of Si and Au line structures, respectively. This result shows that the machine learning methods can be practically applied to the MBL method for the purpose of reducing the library size, accelerating the construction of the MBL database and enriching the content of an available MBL database. Keywords：CD-SEM; MBL; Monte Carlo; machine learning; neural network 1 Introduction The miniaturization of semiconductor devices has been the main direction of the device performance development in the past decades, and the number of transistors that can be accommodated on integrated circuits has been successfully doubled every two years in accordance with the Moore's law [1]. Even though Moore's law is now gradually failing in the post-Moore era at the atomic scale [2], the development of the semiconductor industry continues unabated. To achieve higher performance chips, the linewidth of semiconductors needs to be continuously reduced. Then the control of nanostructure dimensions is becoming more refined, and the critical dimension (CD) measurement technique becomes more important [3] . The CD measurement accuracy needs to be continuously improved, while the level of the CD measurement characterizes the level of development of the semiconductor industry. Transmission electron microscopy [4], atomic force microscopy [5][6] and scanning electron microscopy (SEM) [7][8][9] are useful techniques for the CD measurement. Among them, only critical dimension scanning electron microscope (CD-SEM) is commonly adopted in industry for its fast and convenient measurement procedure. The use of secondary electrons as imaging signals has advantages on the CD measurement for both its sensitive response to the sample morphology and the high image resolution [10] . However, due to the edge effect of secondary electron emission [11][12][13] the linescan profile of the secondary electron signals has a certain spreading at an edge of the sample line structure [14][15][16]; hence, the accurate edge position is hardly to be determined from the secondary electron linescan profile directly. To cope with the edge effect problem in the CD measurement, some al

FAQs

How does the work combine MBL with machine learning for CD-SEM measurement?

It uses Monte Carlo simulation to generate secondary-electron linescan profiles for varied geometric parameters, then trains a neural network to predict linescan profiles from a training set. The predicted profiles support the MBL-based matching approach for CD parameter inference.

Why are secondary-electron linescan edge effects important in CD measurement?

Secondary electrons show strong sensitivity to sample morphology, but edge effects cause spreading in the linescan profile near edges. This makes direct determination of accurate edge positions difficult using the linescan profile alone.

What simulation features are varied to build the training and evaluation data?

The linescan profiles are computed by Monte Carlo simulation as functions of geometric parameters such as top CD, sidewall angle, and height for Si and Au trapezoidal line structures, with the goal of learning the mapping between structure and profile.

Aug 5, 2026

Upload by Ivy