Image-Based Pattern Selection for Lithography Model Training Coverage
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current pattern selection methods for training data in semiconductor manufacturing, particularly in lithography, face challenges such as overfitting and incomplete coverage due to manual selection processes or limited data processing, leading to inaccuracies in predicting patterns on substrates.
Innovation Solution
An image-based pattern selection method that groups patterns into special and main patterns using clustering algorithms, generating feature vectors, and selecting a set of patterns for training models to improve coverage and accuracy, including both outlier and representative patterns, thereby enhancing the model's predictive capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual pattern selection methods are used for training data, then the selection process is simple to implement, but the coverage and accuracy of the trained model are insufficient
Solution Approach 1:
The patent replaces manual pattern selection (mechanical/human process) with an automated image-based processing system that uses feature extraction and clustering algorithms. This substitution improves model prediction accuracy while managing complexity through automation, as the system objectively processes patterns without human intervention.
Solution Approach 2:
The patent transforms patterns into feature vectors by changing their representation parameters. This transformation enables quantitative comparison and automated clustering of patterns based on extracted features, improving selection accuracy through objective parameter-based differentiation rather than subjective manual evaluation.
2Reliability
If limited data processing is used in pattern selection, then the computing resources required are reduced, but overfitting and incomplete coverage occur
Solution Approach 1:
The patent extracts essential features from patterns to create compact feature vectors, taking out only the most relevant characteristics needed for model training. This extraction approach improves model generalization by focusing on key discriminative features while reducing computing resource consumption by working with condensed representations rather than complete raw data.
Solution Approach 2:
The patent applies clustering algorithms to identify and select representative patterns from the dataset, performing partial processing that captures the essential diversity of patterns without processing every single pattern. This partial action approach achieves good model generalization while avoiding the excessive computing resources that would be required to process the entire dataset in detail.
3Adaptability or versatility
If a diverse set of patterns is selected for training, then the model coverage is improved, but the data processing time and complexity increase
Solution Approach 1:
The patent performs preliminary feature extraction and clustering analysis on the pattern dataset before model training. This preliminary action identifies and groups similar patterns, allowing the selection of diverse representative patterns from each cluster. This approach improves model coverage by ensuring diverse pattern representation while reducing processing time by pre-organizing data into meaningful groups that can be efficiently sampled.
Data Source
AI summary
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.