Machine Learning Model Error Clustering for Lithography
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current lithographic technologies face challenges in accurately predicting and correcting errors in pattern representation, particularly in regions with significant deviations from the target pattern, due to the limitations of metrics like root mean square error (RMSE) which do not effectively localize or prioritize prediction errors.
Innovation Solution
The implementation of a mechanism to generate cluster error data and error cluster maps, which identify and prioritize regions with errors in the predicted pattern representation, allowing for the adjustment and training of machine learning models to improve prediction accuracy by focusing on error-prone areas.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models are trained using global metrics like RMSE, then overall prediction accuracy is improved, but the ability to localize and prioritize specific error regions deteriorates
Solution Approach 1:
The patent segments the error analysis by dividing the predicted pattern into multiple regions or zones and calculating local error metrics for each region rather than using a single global RMSE metric. This segmentation enables identification of specific error-prone regions while maintaining overall accuracy assessment, directly resolving the contradiction between global accuracy measurement and local error localization.
Solution Approach 2:
The patent implements local quality assessment by computing error metrics individually for different regions of the pattern, allowing each region to be evaluated based on its specific characteristics and error profile. This enables the system to prioritize corrections in regions with highest local error magnitude while preserving the ability to assess overall model performance.
2Manufacturing precision
If error correction focuses on regions with significant deviations, then manufacturing precision is improved, but the complexity of error analysis and model training increases
Solution Approach 1:
The patent extracts and isolates error clusters from the overall pattern by identifying regions where local error metrics exceed predefined thresholds. These extracted error regions can then be targeted for specific correction operations, separating the correction task from the entire pattern analysis and reducing the complexity of error handling by focusing only on problematic areas.
Solution Approach 2:
The patent implements feedback mechanisms where local error analysis results are fed back into the model training process, allowing the model to learn from and adjust its predictions in error-prone regions. This feedback loop enables continuous improvement of manufacturing precision while managing complexity through automated, data-driven correction rather than manual analysis.
3Productivity
If traditional metrics like RMSE are used to evaluate predictions, then computational efficiency is maintained, but the ability to prioritize error correction regions is lost
Solution Approach 1:
The patent applies partial action by calculating error metrics only for specific regions or pixels where deviations exceed thresholds, rather than uniformly processing the entire pattern. This selective approach maintains computational efficiency by avoiding unnecessary calculations in low-error regions while providing detailed error prioritization where needed, balancing productivity with operational ease.
Data Source
AI summary
Identification of error clusters in an image predicted by a simulation model (e.g., a machine learning model), and training or adjusting the simulation model by feeding the error cluster information back to the simulation model to improve the prediction in regions of the image having the error clusters. Further, embodiments are disclosed for scoring the predicted images, or the simulation models generating those predicted images, based on a severity of errors in the error clusters. The score may be used in evaluating the simulation models to select a specific simulation model for generating a predicted image that may be used in manufacturing a mask to print a desired pattern on a substrate.