Weak Labeling for Semiconductor Defect Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor defect detection methods require high-resolution imaging and manual annotation, which are time-consuming and costly, especially for classifying defects in ultra-large scale integration devices, necessitating a more efficient automated process.
Innovation Solution
A system utilizing a machine learning model trained with high-resolution images and weak labels derived from low-resolution inspections to classify defectiveness in semiconductor specimens, employing neural networks to generate per-pixel-block scores indicative of defect likelihood, enabling automated defect classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution imaging and manual annotation are used for defect classification, then measurement precision is improved, but loss of time and productivity deteriorate
Solution Approach 1:
The patent applies preliminary action by using low-resolution inspection to pre-identify and locate potential defects before high-resolution imaging. This preliminary step creates a defect map that guides subsequent high-resolution analysis, avoiding the need for manual annotation of entire wafer surfaces and significantly reducing annotation time while maintaining classification accuracy
Solution Approach 2:
The patent segments the defect detection process into two distinct phases: (1) low-resolution screening to identify suspect regions and generate defect maps, and (2) high-resolution analysis focused only on identified defect locations. This segmentation allows automated classification to handle the time-consuming low-resolution phase while preserving manual or high-precision automated analysis only where needed, thus reducing overall annotation time without sacrificing precision
2Measurement precision
If high-resolution imaging is used for entire wafer surface inspection, then measurement precision is improved, but productivity deteriorates
Solution Approach 1:
The patent divides the inspection process into two resolution levels: low-resolution rapid scanning of the entire wafer surface to generate defect maps, followed by high-resolution imaging only of identified defect regions. This segmentation enables fast initial screening at low resolution while concentrating computational and imaging resources only where defects are suspected, thereby maintaining high detection accuracy without sacrificing overall inspection productivity
Solution Approach 2:
The patent applies partial action by performing high-resolution imaging only on portions of the wafer surface where defects have been identified by low-resolution inspection, rather than applying high-resolution imaging to the entire surface. This selective approach reduces the total number of high-resolution images required, significantly improving inspection speed while maintaining defect detection accuracy through focused analysis of suspect regions
3Productivity
If automated examination processes are implemented, then productivity is improved, but device complexity increases
Solution Approach 1:
The patent segments the automated examination system into distinct functional modules: low-resolution imaging subsystem, defect map generation algorithm, high-resolution imaging subsystem, and classification engine. This modular segmentation allows each component to be optimized independently and facilitates easier implementation and maintenance of the automated system, reducing the practical complexity despite high productivity gains
Data Source
AI summary
A system for classifying a pattern of interest (POI) on a semiconductor specimen is disclosed. The system comprises a processor and memory circuitry. The memory circuitry is configured to obtain a high-resolution image of the POI, and to generate data usable for classifying the POI in accordance with a defectiveness-related classification. To generate the data, a machine learning model is utilized that has been trained in accordance with training samples. The training samples include a high-resolution training image captured by scanning a respective training pattern on a specimen, the respective training pattern being similar to the POI. The training samples also include a label associated with the image, the label being derivative of low-resolution inspection of the respective training pattern.