Image-Based Pattern Selection for Diverse Lithography Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current pattern selection processes for training data in semiconductor manufacturing are prone to overfitting and require significant time and resources, often missing key patterns and resulting in inadequate model coverage and accuracy.
Innovation Solution
An image-based pattern selection process that identifies both special (outlier) and representative patterns from a data pool, using clustering algorithms and feature vectors to enhance machine learning model coverage and accuracy with reduced training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional pattern selection processes are used to train machine learning models, then model training can be performed, but the models suffer from overfitting and inadequate coverage due to insufficiently diverse training data
Solution Approach 1:
The patent segments the pattern selection process into multiple stages: initial pattern selection, model training, performance evaluation, and iterative refinement. By dividing the training data selection into distinct phases with specific selection criteria for each phase, the system ensures diverse pattern coverage while preventing overfitting through structured data subset creation
Solution Approach 2:
The patent changes selection parameters dynamically during the training process. Different selection criteria are applied at different stages: initial diversity-focused selection, then performance-focused selection, with parameters adjusted based on model performance metrics. This parameter adaptation ensures both broad coverage and high accuracy
2Reliability
If comprehensive pattern data is collected for model training, then model coverage is improved, but the time and resources required for data processing and model training increase significantly
Solution Approach 1:
The patent applies partial action by selecting only the necessary subset of patterns for training rather than processing all available data. Through iterative selection and performance evaluation, the system identifies the minimum sufficient training set that achieves desired model coverage, avoiding unnecessary processing of excess data
Solution Approach 2:
The patent performs preliminary pattern selection and data preprocessing before actual model training. By pre-processing and selecting patterns in advance based on diversity and representativeness criteria, the system reduces the computational burden during the main training phase, saving significant time and resources
3Ease of manufacture
If traditional pattern selection methods are used, then the process is simple to implement, but key patterns are missed resulting in inadequate model performance
Solution Approach 1:
The patent implements feedback loops where model performance on validation data continuously informs pattern selection decisions. Selection criteria are refined based on performance metrics, and poorly performing patterns are identified and replaced. This feedback mechanism ensures high selection accuracy while maintaining automated simplicity
Solution Approach 2:
The system performs self-service through automated pattern selection and evaluation. The model itself identifies which patterns are most valuable for training by evaluating its own performance on different pattern subsets, eliminating the need for manual pattern identification while maintaining high selection accuracy
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Described herein is a method for selecting patterns for training a model to predict patterns to be printed on a substrate. The method includes (a) obtaining images of multiple patterns, wherein the multiple patterns correspond to target patterns to be printed on a substrate; (b) grouping the images into a group of special patterns and multiple groups of main patterns; and (c) outputting a set of patterns based on the images as training data for training the model, wherein the set of patterns includes the group of special patterns and a representative main pattern from each group of main patterns.