Synthetic Defect Image Generation for ML Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for training machine learning models for defect detection in semiconductor manufacturing rely on scarce and often absent examples of defects of interest (DOIs), leading to inadequate model performance and high nuisance rates due to the difficulty in collecting sufficient and accurate training data.
Innovation Solution
A system and method that allows users to create synthetic defects using image editing tools within a graphical user interface, enabling the training of machine learning models even without real examples of DOIs, by generating modified images that are used to enhance model performance and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual defect detection and annotation is used to create training sets, then model training can be performed with real defect examples, but the process is extremely time-consuming and impractical due to the sparsity of defects requiring extensive visual scanning
Solution Approach 1:
The patent creates synthetic defect images by copying and modifying existing defect-free wafer images, adding simulated defects through image processing techniques. This allows generation of unlimited training examples without manual defect scanning, resolving the contradiction between training quality and data collection time
Solution Approach 2:
The system enables automatic generation of annotated training data through automated defect simulation and annotation processes, eliminating the need for manual defect detection and labeling. The system serves itself by generating its own training data without human intervention in the data collection process
2Reliability
If training is performed with a limited set of collected defect examples, then model training can proceed with available data, but the model fails to generalize to unrepresented defect variants leading to poor classification performance
Solution Approach 1:
The system varies multiple parameters in synthetic defect generation including defect size, shape, intensity, position, and type. By changing these parameters across many generated images, the model learns to generalize across diverse defect variants rather than memorizing specific examples, resolving the contradiction between reliability and adaptability
3Quantity of substance
If extensive manual scanning is performed to compile a comprehensive training set, then more defect examples can be collected, but the process becomes increasingly impractical and error-prone due to the sparse distribution of defects
Solution Approach 1:
Instead of manually scanning to find defects, the system copies defect-free images and programmatically adds synthetic defects. This generates unlimited training examples automatically, resolving the contradiction between quantity of examples and operational ease
Solution Approach 2:
The patent replaces the manual mechanical process of visual scanning and annotation with automated image processing algorithms. This substitution eliminates human error and dramatically reduces the complexity of data compilation while increasing the quantity of training examples
Data Source
AI summary
Methods and systems for training a machine learning model using synthetic defect images are provided. One system includes one or more components executed by one or more computer subsystems. The one or more components include a graphical user interface (GUI) configured for displaying one or more images for a specimen and image editing tools to a user and for receiving input from the user that includes one or more alterations to at least one of the images using one or more of the image editing tools. The component(s) also include an image processing module configured for applying the alteration(s) to the at least one image thereby generating at least one modified image and storing the at least one modified image in a training set. The computer subsystem(s) are configured for training a machine learning model with the training set in which the at least one modified image is stored.