Machine Learning SRAF Placement for Lithographic Pattern Fidelity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current lithographic techniques face challenges in reproducing patterns with dimensions smaller than the classical resolution limit, necessitating sophisticated methods like optical proximity correction and customized illumination, which are complex and inefficient.
Innovation Solution
A machine learning model is trained to generate sub-resolution assist features by minimizing a metric between predicted and actual presence values, using training data from reference images to determine optimal placement of assist features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional lithographic techniques are used to reproduce patterns smaller than the resolution limit, then manufacturing precision deteriorates, but device complexity increases due to sophisticated correction methods
Solution Approach 1:
The patent replaces complex mechanical/optical correction systems with a machine learning-based computational system. The ML model trained on synthesis data predicts optimal SRAF placement, substituting iterative optical proximity correction algorithms with a direct predictive model that outputs assist feature locations based on input pattern geometry.
Solution Approach 2:
The patent performs preliminary training of the machine learning model using synthesis data generated from known patterns and their corresponding optimal SRAF placements. This pre-computed knowledge is stored and later applied to new patterns, eliminating the need for real-time complex calculations during actual lithographic processing.
2Manufacturing precision
If optical proximity correction and customized illumination are used to improve pattern reproduction, then manufacturing precision improves, but productivity decreases due to complex processing
Solution Approach 1:
The patent substitutes slow iterative optical proximity correction computations with a pre-trained machine learning model that directly predicts SRAF locations. The model processes input patterns and generates assist feature placement recommendations in a single computational pass, eliminating multiple iterative correction cycles.
Solution Approach 2:
The patent creates a computational model that learns from synthesized examples of optimal SRAF placements. The trained ML model copies the decision-making logic from training data, enabling rapid prediction of assist feature locations for new patterns without re-running complex optical simulations.
3Manufacturing precision
If subresolution assist features are added to enhance pattern reproduction, then manufacturing precision improves, but device complexity increases due to additional features
Solution Approach 1:
The patent applies subresolution assist features selectively at specific locations around target patterns where they are most needed, rather than uniformly across all patterns. The machine learning model analyzes local pattern geometry and determines optimal SRAF placement positions, creating non-uniform assist feature distributions that enhance precision without unnecessary complexity.
Solution Approach 2:
The patent processes patterns by identifying individual target features and determining SRAF placement for each separately. The machine learning model evaluates each pattern element independently and generates localized assist feature recommendations, allowing precise control over where complexity is introduced based on actual need.
Data Source
AI summary
A method for training a machine learning model to generate a characteristic pattern, the method includes obtaining training data associated with a reference feature in a reference image. The training data includes (i) location data of each portion of the reference feature, and (ii) a presence value indicating whether the portion of the reference feature is located within a reference assist feature generated for the reference feature. The method includes training the machine learning model to predict a presence value based on the actual presence value in the training data. The predicted presence value indicates whether a portion of a feature (e.g., a skeleton point on a skeleton of a contour of the feature) is to be covered by an assist feature. The training is performed based on the training data such that a metric between a predicted presence value and the presence value is minimized.