Fabrication Layout Retargeting With Differentiable Polygon Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning-based methods for device fabrication retargeting are non-differentiable, limited to individual feature optimization, and fail to account for neighboring pattern influences, leading to inaccurate and inefficient pattern transfer to substrates.
Innovation Solution
Implement a differentiable deep learning network with relative coordinate encoding and position-based masking to enable gradient-based backpropagation, allowing for accurate prediction and adjustment of layout data to achieve target dimensions by considering neighboring features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If existing machine learning-based methods are used for retargeting, then individual feature optimization can be achieved, but the methods are non-differentiable and fail to account for neighboring pattern influences, leading to inaccurate pattern transfer
Solution Approach 1:
The patent replaces traditional non-differentiable machine learning methods with a differentiable deep learning network. This substitution enables gradient-based optimization and allows the model to account for neighboring pattern influences through differentiable operations, thereby improving pattern transfer accuracy while maintaining computational tractability.
Solution Approach 2:
The patent merges individual feature optimization with neighboring pattern consideration into a unified differentiable deep learning framework. By combining these previously separate optimization goals into a single model, the system can simultaneously optimize individual features while accounting for their interactions with neighboring patterns, resolving the contradiction between accuracy and method complexity.
2Manufacturing precision
If traditional retargeting methods are used, then computational efficiency may be maintained, but they fail to provide accurate prediction and adjustment of layout data considering neighboring features
Solution Approach 1:
The patent introduces a dynamic deep learning network that can adaptively learn the complex relationships between layout features and their neighbors. The differentiable architecture allows for flexible optimization while maintaining computational efficiency through gradient-based training, enabling accurate prediction and adjustment of layout data without sacrificing productivity.
Solution Approach 2:
The patent changes the parameter representation and optimization approach by using differentiable operations within the deep learning network. This allows the system to efficiently compute gradients and adjust layout parameters iteratively, achieving high accuracy in layout data prediction while maintaining computational efficiency through optimized training procedures.
3Ease of operation
If individual feature optimization is performed, then each feature can be adjusted independently, but neighboring pattern influences are not accounted for, reducing overall fabrication accuracy
Solution Approach 1:
The patent segments the layout into individual features while maintaining their contextual relationships through the deep learning network's architecture. Each feature can be independently processed and optimized, yet the differentiable model simultaneously considers neighboring pattern influences, resolving the contradiction between operational independence and fabrication accuracy.
Solution Approach 2:
The patent creates a universal deep learning model that serves multiple functions: it can optimize individual features independently while simultaneously accounting for neighboring pattern influences. This multi-functional approach allows the system to maintain ease of operation for individual feature adjustment while achieving high fabrication accuracy through integrated pattern recognition.
Data Source
AI summary
A method includes: obtaining layout data representing a candidate device fabrication pattern; generating an embedding of the layout data; providing the embedding of the layout data as input to a deep learning model; obtaining, as an output of the deep learning model, a predicted fabricated structure formed using the candidate device fabrication pattern; and adjusting the layout data by backpropagation based on a gradient of a loss function representing a difference between the predicted fabricated structure and a target structure.