Semiconductor Defect Detection With Synthetic Reference Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current defect examination methods in semiconductor fabrication face challenges such as increased image acquisition time due to the need for multiple image comparisons and insufficient training data for machine learning models, leading to low detection sensitivity and high false alarm rates.
Innovation Solution
A system utilizing a pre-trained first ML model for feature extraction and comparison with a bank of reference features generated from synthetic images produced by a second ML model, enabling efficient detection of defective patches in semiconductor specimens.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple actual reference images are acquired and compared for defect detection, then detection sensitivity is improved, but image acquisition time increases
Solution Approach 1:
The patent uses a generative model to create synthetic reference images that copy the essential characteristics of actual reference images. These synthetic images serve as substitutes for multiple actual images, providing sufficient variation for accurate defect detection without requiring acquisition of multiple physical images. The synthetic images are generated by transforming a small set of actual reference images through the generative model, thereby reducing acquisition time while maintaining detection sensitivity.
2Measurement precision
If multiple actual reference images are acquired for training machine learning models, then model accuracy is improved, but data acquisition time and resources increase
Solution Approach 1:
The generative model creates synthetic reference images that serve as training data for machine learning models. These synthetic images replicate the statistical properties and variations of actual reference images, enabling effective model training with fewer actual images. The synthetic data generation process transforms a limited set of actual images into a diverse training set, reducing data acquisition time while maintaining model accuracy.
3Reliability
If traditional defect examination methods are used with multiple image comparisons, then defect detection capability is maintained, but false alarm rate increases
Solution Approach 1:
The synthetic reference images generated by the generative model provide a more consistent and controlled reference basis for defect detection. By using synthetic images that capture the essential variations of normal structures without actual defects, the system reduces false alarms while maintaining defect detection capability. The synthetic references eliminate some of the variability present in actual reference images that can lead to false positives.
Data Source
AI summary
There is provided a system and method of examination a semiconductor specimen. The method includes obtaining a runtime image of the specimen; processing the runtime image using a first machine learning (ML) model to extract a set of runtime features representative of a set of patches in the runtime image; and comparing the set of runtime features with a bank of reference features, giving rise to an anomaly map indicative of one or more defective patches in the runtime image. The bank of reference features is previously generated by obtaining a plurality of synthetic reference images generated by a second ML model based on a plurality of actual images; and processing the plurality of synthetic reference images by the first ML model to extract, for each synthetic reference image, a set of reference features representative thereof, giving rise to the bank of reference features.