Semiconductor Defect Detection Using Unsupervised Reference Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor defect detection methods rely heavily on human-labeled data, which is time-consuming, prone to errors, and introduces inconsistencies, leading to inaccurate defect classification and increased manufacturing costs due to false positives and negatives.
Innovation Solution
A dual-network approach involving a reconstruction network and a detection network, where the reconstruction network generates clean reference images for defect images, and the detection network is trained using a specially designed loss function combining difference images with predicted defect maps, enabling unsupervised defect detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If human-labeled data is used for defect detection, then defect classification can be performed, but the process becomes time-consuming and error-prone
Solution Approach 1:
The system enables self-supervised learning where the model automatically generates its own training labels through the reconstruction network, eliminating the need for manual human labeling. The detection network and reconstruction network work together to self-train on unlabeled data, making the system self-sufficient and removing the time-consuming manual annotation process while maintaining defect detection accuracy.
2Measurement precision
If human-labeled data is used for defect detection, then defect classification can be performed, but inconsistencies and errors are introduced
Solution Approach 1:
By using self-supervised learning, the system eliminates human involvement in label creation, thereby removing human errors and inconsistencies. The reconstruction network objectively generates labels based on image reconstruction errors, providing consistent and reliable training data that improves defect classification reliability across different batches and operators.
3Productivity
If traditional supervised learning is used, then defect detection can be trained, but human-labeled data is required which increases manufacturing costs
Solution Approach 1:
The system performs unsupervised learning by automatically generating training labels from the data itself through the reconstruction network, eliminating the need to purchase or commission manual labeling services. This self-training capability maintains full defect detection functionality while removing the additional cost associated with human-labeled datasets.
4Measurement precision
If reference images are acquired for each runtime image, then defect detection accuracy improves, but the examination process becomes more complex
Solution Approach 1:
The system performs preliminary training offline where the detection network learns to identify defects using self-generated labels from training images. This pre-training phase enables the network to detect defects in runtime images without requiring reference images during actual examination, simplifying the runtime process while maintaining detection accuracy through the previously learned patterns.
Data Source
AI summary
There is provided a system and method of runtime defect detection in a semiconductor specimen. The method includes obtaining a runtime image of the specimen; and processing, by a detection network, the runtime image to obtain a defect map indicating probabilities of defect distribution thereof. The detection network is previously trained unsupervised in a training phase, comprising, for a training image: obtaining a reference image of the training image; processing, by a detection network to be trained, the training image to generate a predicted defect map thereof; and optimizing the detection network to be trained using a loss function constructed based on the predicted defect map, and a difference image between the training image and the reference image.