Semiconductor Process Limits via Generative ML
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for determining process limits in semiconductor lithography, such as exposure dose and focus, are labor-intensive and time-consuming, particularly for small device dimensions, due to the need for manual analysis of CD-SEM images, which are not sufficiently accurate for critical dimensions under 32nm.
Innovation Solution
An unsupervised generative machine learning algorithm is trained on a dataset of CD-SEM images to create a latent vector space, allowing for the synthesis of synthetic images and the attribution of scores to query images based on distances in this space, enabling the extraction of process limits without manual comparison.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual analysis of CD-SEM images is used to determine process limits, then measurement precision is maintained, but productivity deteriorates due to labor-intensive and time-consuming effort
Solution Approach 1:
The patent uses generative machine learning models to create synthetic images that replicate the characteristics of real CD-SEM images. These synthetic images serve as proxies for manual analysis, enabling automated determination of process limits while maintaining measurement accuracy. The model learns from real images and generates synthetic variants that preserve critical features for process window characterization.
Solution Approach 2:
The patent replaces the mechanical/manual process of image analysis with an automated machine learning system. Instead of human operators manually examining CD-SEM images to determine process limits, the system uses trained generative models to automatically analyze images and extract process parameter information, thereby eliminating the labor-intensive nature of the task while preserving measurement precision.
2Ease of operation
If optical inspection is used for analyzing printed dies, then ease of operation is improved, but measurement precision deteriorates for critical dimensions under 32nm due to insufficient resolution
Solution Approach 1:
The patent introduces an intermediary machine learning model that processes images obtained from various sources (including optical inspection and CD-SEM). The generative model learns to recognize critical features and determine process limits without requiring direct high-resolution measurement of sub-32nm features. This intermediary system bridges the gap between ease of optical inspection and the need for precise measurement of small features.
3Measurement precision
If extensive manual comparison of images is performed to determine process limits, then measurement precision is maintained, but loss of time increases significantly
Solution Approach 1:
The patent performs preliminary training of the generative machine learning model on a dataset of real CD-SEM images before actual process limit determination. This preliminary action enables the model to learn the relationships between image characteristics and process parameters in advance. During actual use, the pre-trained model can rapidly analyze new images and determine process limits without requiring time-consuming manual comparison, thus reducing time loss while maintaining precision.
Data Source
Figure 1
Figure 2~4
Figure 5
AI summary
A dataset of images of the same die area is taken as the starting point of the method. The images of the dataset correspond to different settings of one or more process parameters. One or more images from the dataset are used as training images of an unsupervised generative machine learning algorithm. The algorithm creates a latent vector space, comprising latent vectors that allow the algorithm to synthesize a large quantity of synthetic images. The method then presents one or more images from the dataset as query images to the algorithm, and a score is attributed to a set of object images from the dataset, based on the distances between the query image and respective object images, said distances being determined in the latent vector space. According to the invention, the score is representative of one or more characteristics of the images. This allows to extract the process limits directly from the score values, rather than from a manual comparison of the images.