Circuit Layout Defect Detection Using AI-Generated Training Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques face difficulties in detecting defects caused by circuit layout in semiconductor devices due to the scarcity of available samples, making it challenging to prepare sufficient training data for classification AI.
Innovation Solution
Utilizing a generative AI to generate virtual images based on defective circuit images, which are used as training data for a classification AI, thereby increasing the number of samples available for learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional defect analysis methods are used with limited samples, then the analysis process remains simple, but the detection capability for rare defects is insufficient
Solution Approach 1:
The patent uses generative AI to create virtual copies of defective circuit images. The system acquires real defective circuit images, uses a generative AI model to synthesize multiple virtual defective circuit images from these real samples, and combines them to form an augmented training dataset. This copying approach enables the classification AI to learn from sufficient training data without requiring numerous real defective samples, thereby resolving the contradiction between limited sample availability and detection capability.
2Measurement precision
If more training samples are collected to improve classification accuracy, then detection performance improves, but the cost and time of data collection increase
Solution Approach 1:
The patent performs preliminary data augmentation by generating virtual defective circuit images before the actual classification task. The system pre-processes real defective images through generative AI to create a comprehensive training dataset in advance, which is then used to train the classification AI. This preliminary action of data synthesis eliminates the need for time-consuming real-world defect collection and annotation during the actual detection phase, thereby improving classification accuracy without proportionally increasing data preparation time.
3Quantity of substance
If real defective samples are used for training, then the training data is authentic, but the number of available samples is insufficient
Solution Approach 1:
The patent introduces generative AI as an intermediary between real defective circuit images and the training dataset. The generative AI model acts as a mediator that transforms a small number of authentic real defective images into a large number of virtual defective images. This intermediary process maintains the authenticity characteristics of real defects while dramatically increasing the quantity of training samples, making data acquisition easier without sacrificing sample quality.
Data Source
AI summary
A learning support device includes: an image acquisition unit acquiring a plurality of defective circuit images representing a circuit layout causing a defect extracted from an existing circuit layout and a plurality of normal circuit images representing a normal circuit layout; a virtual image acquisition unit inputting each of the plurality of defective circuit images into a generative AI (artificial intelligence), and acquiring a plurality of virtual images generated based on the plurality of defective circuit images by using the generative AI; a learning processing unit inputting the plurality of defective circuit images, the plurality of virtual images, and the plurality of normal circuit images as training data into a classification AI; and an output unit outputting a classification AI 150 having already learned.