Synthetic Defect Generation for Semiconductor DNN Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor examination processes face challenges in generating accurate training data for Deep Neural Networks (DNNs) due to the time-consuming and error-prone nature of manual annotation, which affects the quality and efficiency of defect detection and classification in high-resolution images.
Innovation Solution
A computerized system that generates training data for DNNs by extracting features from semiconductor images, including contextual and statistical features, and iteratively refines the training process based on user feedback to improve segmentation accuracy, enabling automated defect detection, classification, and metrology operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation is used to generate training data for DNNs, then training data can be created, but the process is time-consuming and error-prone
Solution Approach 1:
The system enables self-service by allowing the DNN to automatically generate its own training data through synthetic defect generation and automated segmentation, eliminating the need for time-consuming manual annotation while maintaining high accuracy through iterative refinement based on performance metrics
Solution Approach 2:
The system performs preliminary action by pre-generating synthetic defect training data before actual defect detection is needed, using simulated defect patterns and automated segmentation to create a comprehensive training dataset in advance, reducing real-time processing requirements
2Reliability
If manual annotation is used to generate training data, then training data can be created, but the process is error-prone
Solution Approach 1:
The system uses copying by creating synthetic defect patterns that replicate real defect characteristics through simulation, generating multiple copies of training samples with varied parameters to build a robust training dataset that improves reliability without introducing manual annotation errors
Solution Approach 2:
The system implements feedback by continuously evaluating DNN performance on validation data and using the results to iteratively refine synthetic defect generation parameters and segmentation algorithms, systematically reducing errors and improving detection reliability over time
3Productivity
If automated examination processes are implemented, then defect detection efficiency is improved, but the complexity of the system increases
Solution Approach 1:
The system applies universality by designing a multi-functional automated examination platform that integrates synthetic defect generation, automated image segmentation, DNN training, and defect detection classification into a single unified system, improving productivity while managing complexity through functional integration
Solution Approach 2:
The system uses segmentation by dividing the automated examination process into distinct modular components including synthetic defect generation module, automated segmentation module, DNN training module, and defect classification module, allowing each to be optimized independently while maintaining overall system efficiency
Data Source
AI summary
Provided is a system and method of generating training data for training a Deep Neural Network usable for examination of a semiconductor specimen. The method includes: obtaining a first training image and first labels respectively associated with a group of pixels selected in each segment, extract a set of features characterizing the first training image, train a machine learning (ML) model using the first labels, values of the group of pixels, and the feature values of each of the set of features corresponding to the group of pixels, process the first training image using the trained ML model to obtain a first segmentation map, and determine to include the first training image and the first segmentation map into the DNN training data upon a criterion being met, and to repeat the extracting of the second features, the training and the processing upon the criterion not being met.