SEM Image Shape Estimation Using Simulated 3D Pattern Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for estimating the three-dimensional shape of semiconductor patterns and particles on wafers are limited by the need for extensive data acquisition time and lower measurement accuracy, especially for steep slopes, and fail to provide sufficient height information.
Innovation Solution
An image processing system that stores the detectable range of a charged particle beam device, generates simulated three-dimensional shape patterns, and learns the relationship between simulated and actual images to estimate the shape with high accuracy without requiring extensive data acquisition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If one uses conventional methods (AFM or cross-sectional SEM) to acquire learning data for shape estimation, then measurement accuracy may be improved, but time required for measurement increases significantly
Solution Approach 1:
The patent creates simulated SEM images that replicate the appearance and characteristics of actual SEM images. These simulated images serve as proxies for time-consuming AFM or cross-sectional SEM measurements, allowing the system to learn shape estimation mappings without requiring physical measurements from those instruments. The simulated images are generated with known ground truth shapes, enabling efficient training of the shape estimation model.
Solution Approach 2:
The patent performs preliminary generation of simulated SEM images and shape data before actual measurement is needed. By pre-generating a comprehensive dataset of simulated images covering various shapes, sizes, and imaging conditions, the system prepares training data in advance, eliminating the need for time-consuming reference measurements when actual shape estimation is required.
2Device complexity
If one uses a single top-down SEM image to estimate cross-sectional shape, then device complexity is reduced, but height information becomes insufficient
Solution Approach 1:
The patent introduces simulated images as an intermediary that bridges the gap between simple top-down SEM images and the needed three-dimensional shape information. The simulated images encode height and depth information in their intensity patterns, allowing the system to extract vertical dimension data from what would otherwise be a two-dimensional projection view.
Solution Approach 2:
The patent transforms the problem from extracting three-dimensional shape from a two-dimensional image by introducing a simulated image dimension that contains encoded depth information. The simulation process incorporates imaging physics that preserve height-related contrast variations, effectively adding a dimensional aspect to the image data that reveals vertical structure.
3Measurement precision
If one acquires extensive learning data from reference measurements to improve estimation accuracy, then measurement precision improves, but productivity decreases due to data acquisition time
Solution Approach 1:
The patent creates synthetic copies of measurement data through simulation. Instead of acquiring numerous reference measurements from physical instruments, the system generates simulated images that replicate the imaging process and produce corresponding ground truth shape data, creating a large training dataset without physical measurement overhead.
Solution Approach 2:
The simulation system is self-sufficient in generating training data without requiring external reference measurements. The simulated image generation process inherently produces both the input images and the corresponding ground truth shapes, making the system self-service capable of creating unlimited training examples without external instrument intervention.
Data Source
AI summary
Provided is an image processing system capable of estimating a three-dimensional shape of a semiconductor pattern or a particle by solving problems of measurement reduction in a height direction and taking an enormous amount of time at a time of acquiring learning data. The image processing system according to the disclosure stores a detectable range of a detector provided in a charged particle beam device in a storage device in advance, generates a simulated image of a three-dimensional shape pattern using the detectable range, and learns a relationship between the simulated image and the three-dimensional shape pattern in advance.