Deep Learning Nuisance Filtering for Semiconductor Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor inspection methods are inefficient in filtering nuisances from defect candidates, leading to prolonged setup times and sub-optimal performance due to high nuisance rates and limited availability of defects of interest, especially in complex layers like RDL and BEOL where grain noise and other noise sources overwhelm defect detection.
Innovation Solution
A system utilizing two deep learning networks to filter nuisances from defect candidates, with the first network processing high-volume patch images to reduce nuisance rates and the second network applying high-resolution image analysis to further suppress noise, enabling effective detection of defects in noisy backgrounds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If conventional nuisance filtering methods are used, then setup time is reduced, but nuisance filtering performance deteriorates due to high nuisance rates overwhelming defect detection
Solution Approach 1:
The patent divides the nuisance filtering process into multiple stages using different deep learning networks. A first DL network performs initial filtering on high-volume patch images, followed by a second DL network that applies high-resolution image analysis for further noise suppression. This segmented approach maintains reliability by progressively filtering nuisances while managing computational load efficiently.
Solution Approach 2:
The patent transitions from analyzing low-resolution patch images to high-resolution images across different dimensional scales. The first DL network processes coarse patch images for rapid initial filtering, then the second DL network examines high-resolution images of remaining defect candidates for refined nuisance suppression, leveraging multi-dimensional analysis to improve filtering performance.
2Reliability
If deep learning networks with high-resolution image analysis are used, then nuisance filtering performance is improved, but computational complexity and processing time increase
Solution Approach 1:
The patent segments the computational workload across two specialized deep learning networks. The first DL network handles the bulk of initial nuisance filtering on down-sampled patch images, while the second DL network focuses computational resources on analyzing only the remaining defect candidates at high resolution. This segmentation reduces overall computational complexity compared to applying high-resolution analysis to all defect candidates.
Solution Approach 2:
The patent applies high-resolution image analysis selectively rather than universally. The second DL network performs computationally intensive high-resolution analysis only on defect candidates that survived the first filtering stage, rather than processing all detected defects. This partial application of excessive computational action maintains high filtering performance where needed while reducing overall system complexity.
3Productivity
If traditional defect detection methods are used, then processing speed is maintained, but detection accuracy deteriorates in noisy backgrounds with high nuisance rates
Solution Approach 1:
The patent segments the defect detection pipeline into rapid initial screening followed by refined analysis. The first DL network quickly processes high-volume patch images at reduced resolution to identify potential defects, maintaining processing speed. The second DL network then applies high-resolution analysis to refine detection accuracy for the subset of candidates, achieving both speed and precision in noisy environments.
Solution Approach 2:
The patent employs multi-resolution analysis across different dimensional scales. Low-resolution patch images enable rapid initial defect candidate identification, while high-resolution images provide the detailed view needed for accurate defect verification. This dimensional transition allows the system to maintain processing speed through efficient low-resolution screening while achieving high detection accuracy through selective high-resolution analysis.
Data Source
AI summary
Methods and systems for detecting defects on a specimen are provided. One system includes a first deep learning (DL) network configured for filtering nuisances from defect candidates detected on a specimen. Output of the first DL network includes a first subset of the defect candidates not filtered as the nuisances. The system also includes a second DL network configured for filtering nuisances from the first subset of the defect candidates. Computer subsystem(s) input high resolution images acquired for the first subset of the defect candidates into the second DL network. Output of the second DL network includes a final subset of the defect candidates not filtered as the nuisances. The computer subsystem(s) designate the defect candidates in the final subset as defects on the specimen and generate results for the defects.