Unsupervised Reference Selection for Semiconductor Defect Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional defect inspection systems in semiconductor wafers face challenges in accurately detecting defects due to noise from radial dependency in wafer processing, as they typically compare images from adjacent dies, which may not be the most similar, leading to artifacts in difference images and reduced sensitivity.
Innovation Solution
An unsupervised learning-based reference image selection algorithm that identifies and uses the most similar reference images throughout the wafer for defect detection, rather than relying solely on adjacent dies, to generate a reference image that minimizes noise and enhances defect detection sensitivity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If images from adjacent dies are used for reference comparison, then the inspection process is simple and fast, but noise from radial dependency causes reduced defect detection sensitivity
Solution Approach 1:
The system performs self-characterization by automatically analyzing its own inspection data to identify and cluster dies with similar noise profiles. The clustering algorithm autonomously groups dies based on dissimilarity metrics calculated from image data, enabling the system to select appropriate reference dies without external intervention or manual configuration.
Solution Approach 2:
The system changes the selection criterion from spatial adjacency to noise profile similarity by calculating dissimilarity values based on image characteristics. This parameter transformation allows the system to select reference dies that minimize radial dependency noise, improving defect detection sensitivity while adapting to the specific noise characteristics of each wafer.
2Measurement precision
If unsupervised learning-based cluster analysis is performed to select reference images, then defect detection sensitivity is enhanced, but computational complexity and processing time increase
Solution Approach 1:
The system performs cluster analysis and reference die selection in advance during a characterization step before actual defect inspection. By pre-identifying and grouping dies with similar noise profiles, the system prepares optimal reference selections that can be quickly applied during subsequent inspections, reducing real-time processing requirements.
Solution Approach 2:
The wafer is divided into multiple clusters based on noise profile similarity, with each cluster containing dies that share comparable characteristics. This segmentation allows the system to select references within the same cluster, ensuring compatibility while distributing computational work across multiple manageable groups rather than analyzing all dies uniformly.
3Measurement precision
If reference images are generated from multiple combinations within clusters, then noise from radial dependency is reduced, but the complexity of reference generation increases
Solution Approach 1:
The system merges multiple image combinations within each cluster to generate composite reference images. By combining images from multiple dies that share similar noise profiles, the system averages out random noise while preserving systematic radial dependency patterns, thereby reducing noise impact on defect detection.
Solution Approach 2:
The system creates reference copies from selected dies within clusters and uses these synthesized references for comparison. Instead of directly comparing test images to original die images, the system generates representative reference copies that capture the typical noise characteristics of each cluster, simplifying the comparison process while maintaining noise reduction benefits.
Data Source
AI summary
An optical characterization system and a method of using the same are disclosed. The system comprises a controller configured to be communicatively coupled with one or more detectors configured to receive illumination from a sample and generate image data. One or more processors may be configured to receive images of dies on the sample, calculate dissimilarity values for all combinations of the images, perform a cluster analysis to partition the combinations of the images into two or more clusters, generate a reference image for a cluster of the two or more clusters using two or more of the combinations of the images in the cluster; and detect one or more defects on the sample by comparing a test image in the cluster to the reference image for the cluster.