Neural Network Mask Generation for Reticle Enhancement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current reticle enhancement technologies face challenges in efficiently producing masks with high dose margin and resilience to manufacturing variation, particularly with smaller feature sizes and complex geometries, leading to issues with critical dimension uniformity and line-edge roughness.
Innovation Solution
The method involves generating a Continuous Tone Mask (CTM) and Quantized Tone Mask (QTM) using a trained neural network, allowing for optimized mask creation across large sections of the mask layer, reducing stitching artifacts, and employing Function Sample Arrays to represent smooth functions, enabling efficient computation and improved manufacturability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional reticle enhancement technologies are used to produce masks with high dose margin and resilience to manufacturing variation, then manufacturing precision is improved, but device complexity increases due to smaller feature sizes and complex geometries
Solution Approach 1:
The mask design is divided into multiple tiles that can be processed independently. Each tile represents a manageable section of the overall mask pattern, allowing complex geometries to be handled in smaller, more tractable units while maintaining overall pattern integrity and manufacturing precision
Solution Approach 2:
The patent transforms the mask design from binary (present/absent) to continuous tone values, enabling gradual transitions in material density or thickness. This parameter transformation allows for smoother geometries that reduce manufacturing variation while maintaining design flexibility for complex features
2Manufacturing precision
If conventional reticle enhancement technologies are used to produce masks with high dose margin and resilience to manufacturing variation, then manufacturing precision is improved, but line-edge roughness increases
Solution Approach 1:
By transitioning from binary mask patterns to continuous tone patterns, the invention enables gradual transitions in material properties. This parameter transformation smooths edge transitions and reduces abrupt changes that cause line-edge roughness, while still achieving the desired critical dimension uniformity through controlled tone variations
Solution Approach 2:
The continuous tone approach allows different regions of the mask to have locally optimized properties. Areas requiring high precision can have smoother, more gradual tone transitions, while other areas can be optimized for different manufacturing constraints, thereby reducing overall line-edge roughness while maintaining critical dimension uniformity
3Productivity
If neural networks are used to generate optimized masks across large sections, then productivity is improved, but memory requirements increase
Solution Approach 1:
The large mask design is divided into smaller tiles that can be processed independently by the neural network. This segmentation reduces the memory footprint required for each neural network operation while maintaining the ability to generate optimized masks across the entire large section through systematic processing of individual tiles
Solution Approach 2:
The neural network is trained to generate optimized patterns for local tile regions rather than requiring the entire large mask section to be processed simultaneously. This local processing approach reduces memory requirements by focusing computational resources on smaller, manageable regions while still achieving overall productivity improvement through automated generation
Data Source
AI summary
Methods and systems for reticle enhancement technology (RET) include inputting a target wafer pattern, where the target wafer pattern spans an entire design area. The entire design area is divided into a plurality of tiles, each tile having a halo region surrounding the tile. An optimized mask is calculated, wherein the optimized mask is generated by a first trained neural network using the target wafer patter. The calculating is performed for each tile in the plurality of tiles including its halo region.