Pattern Selection for ML Training Using Entropy Metrics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for selecting patterns for training models in lithography are inefficient, requiring expensive ground truth data and multiple forward passes through neural networks, and lack a direct approach to identify the most informative patterns from large datasets.
Innovation Solution
A method using an information metric, such as information entropy, to select patterns from a target layout without requiring additional patterning or machine learning model simulations, directly identifying patterns with low entropy as most informative for training purposes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current methods for selecting patterns are used, then training data can be obtained, but computational resources and time are excessively consumed due to multiple forward passes through neural networks and expensive ground truth data generation
Solution Approach 1:
The patent segments the pattern selection process into two distinct stages: (1) an efficient pre-screening stage using information entropy metrics to identify candidate patterns, and (2) a refined selection stage using neural network forward passes only on the pre-selected candidates. This segmentation eliminates the need for computationally expensive operations on the entire pattern dataset, thereby reducing computational time while maintaining selection accuracy.
Solution Approach 2:
The patent extracts and removes the computationally expensive ground truth generation and multiple forward pass operations from the pattern selection process. Instead, it extracts only the essential selection criterion (information entropy) that can be computed efficiently, and uses this extracted metric to guide the subsequent neural network training without requiring the full expensive computation pipeline for every pattern.
2Measurement precision
If current methods for selecting patterns are used, then training data can be obtained, but computational resources are excessively consumed due to multiple forward passes through neural networks
Solution Approach 1:
The patent segments the computational workload into an energy-efficient pre-screening phase using information entropy calculations and a targeted neural network evaluation phase. This segmentation ensures that energy-intensive neural network forward passes are performed only on a small subset of pre-selected candidate patterns rather than the entire dataset, dramatically reducing overall computational resource consumption.
Solution Approach 2:
The patent extracts the essential selection criterion from the expensive neural network computation and represents it through a simple information entropy metric that can be calculated with minimal computational resources. This extracted metric serves as a proxy for pattern informativeness, eliminating the need for energy-intensive multiple forward passes while maintaining selection quality.
3Productivity
If patterns are selected without information metric, then computational resources are saved, but the most informative patterns cannot be identified
Solution Approach 1:
The patent introduces information entropy as an intermediary metric that bridges the gap between computational efficiency and pattern informativeness. This intermediary metric provides a computationally efficient approximation of pattern value, enabling the selection of informative patterns without requiring full neural network forward passes, thus maintaining training efficiency while preventing information loss.
Solution Approach 2:
The patent changes the selection parameter from complex neural network output metrics to a simpler information entropy parameter that can be computed efficiently. This parameter change maintains the ability to identify informative patterns while dramatically improving computational efficiency and training productivity.
Data Source
AI summary
A method and apparatus for selecting patterns from an image such as a design layout. The method includes obtaining an image (e.g., of a target layout) having a plurality of patterns; determining, based on pixel intensities within the image, a metric (e.g., entropy) indicative of an amount of information contained in one or more portions of the image; and selecting, based on the metric, a sub-set of the plurality of patterns from the one or more portions of the image having values of the metric within a specified range. The sub-set of patterns can be provided as training data for training a model associated with a patterning process.