CFNO Mask Optimization for Large-Scale Lithography Turnaround
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional mask optimization techniques, including model-based and machine learning approaches, face challenges in achieving fast turnaround times and scalability for large-scale mask design optimization, often relying on sub-optimal training sets and failing to account for global information pertinent to mask optimization tasks.
Innovation Solution
The implementation of a Convolutional Fourier Neural Operator (CFNO) structure that incorporates a token-shared FNO unit and token-wise convolution layer, enabling efficient local-global mixing and addressing limitations of conventional FNO structures by utilizing high-resolution and large-tile inputs to capture global circuit features, thereby improving computing efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional model-based or heuristic optimization methods are used for mask optimization, then manufacturing precision can be achieved, but productivity is insufficient due to slow turnaround times
Solution Approach 1:
The patent replaces conventional iterative numerical optimization methods with a neural network-based system. The neural network is trained on training data containing target patterns and corresponding optimized masks, then uses this trained model to directly generate optimized masks for new target patterns without iterative querying of lithography models, significantly reducing computation time while maintaining optimization accuracy
Solution Approach 2:
The patent performs preliminary training of the neural network using training data that includes target patterns and their corresponding optimized masks generated by conventional methods. This pre-computed knowledge is stored in the neural network's weights and biases, allowing the system to rapidly retrieve and apply optimization results for new patterns without repeating the iterative optimization process
2Productivity
If conventional machine learning models focus on fixed-sized small tile mask optimization, then training speed may be acceptable, but the models do not scale well to large-scale mask design optimization problems
Solution Approach 1:
The patent segments the large mask optimization problem into smaller manageable tiles that can be processed independently by the neural network. Each tile is optimized separately using the trained model, and the results are combined to form the complete optimized mask. This approach allows the model to maintain fast training speeds on small tiles while effectively scaling to large-scale mask design problems through systematic composition of tile results
Solution Approach 2:
The patent creates a universal neural network model that can handle both small tile optimization and large-scale mask optimization. The model is trained on small tile data but designed with architectural features that allow it to generalize to larger problem sizes, making it multi-functional for different scale requirements without requiring separate models for different problem sizes
3Ease of manufacture
If conventional CNN-based machine learning models are used, then the models can process mask data, but they have limited capacity to take into account global information pertinent to mask optimization tasks
Solution Approach 1:
The patent enhances the conventional CNN architecture by incorporating additional dimensional processing capabilities. The neural network processes both local tile information and global mask context simultaneously, adding a global information dimension to the traditional local convolutional processing. This allows the model to capture long-range dependencies and global patterns in addition to local features, preventing information loss while maintaining implementation feasibility
Data Source
AI summary
A circuit mask optimizer utilizes a Convolutional Fourier Neural Operator (CFNO) to efficiently learn layout tile dependencies, enabling stitch-less largescale mask optimization with limited intervention of legacy tools. Litho-guided self training via a trained machine learning model provides non-convex optimization, enabling iterative model and dataset refinements at a substantial performance improvement over conventional solutions.