ML Mask Synthesis Using ILT Training Data for Fast Lithography
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The demand for increased feature density in semiconductor manufacturing leads to distortions and artifacts due to the wave nature of light, making inverse lithography technology (ILT) computationally burdensome, while supervised machine learning models require appropriate training data that is challenging to generate effectively.
Innovation Solution
Using Inverse Lithography Technology (ILT) to generate training data for machine learning models, and designing ML models with translational invariance, model grid shift invariance, and symmetry to improve mask synthesis accuracy, converting outputs to suitable representations for lithography tools.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If Inverse Lithography Technology (ILT) is used to synthesize masks for advanced silicon-based compute nodes, then mask quality is improved, but computational burden and tool runtime increase significantly
Solution Approach 1:
The patent pre-generates training data using ILT for common pattern types and pre-trains the neural network model offline. This preliminary action allows the model to learn from ILT-generated high-quality masks without requiring ILT to run during actual mask synthesis, thus achieving fast inference while maintaining high mask quality.
Solution Approach 2:
The patent uses the neural network model to copy the mask synthesis capabilities of ILT without replicating its computational complexity. The model learns the mapping from layout patterns to optimal masks by training on ILT-generated data, then applies this learned knowledge rapidly to new patterns without performing the full ILT optimization process.
2Productivity
If supervised machine learning models are used to accelerate mask synthesis, then computational speed is improved, but appropriate training data becomes difficult to generate
Solution Approach 1:
The patent introduces an intermediary process that automatically generates training data by combining ILT mask synthesis with layout pattern extraction. This intermediary training data generation system bridges the gap between having high-quality ILT masks and creating suitable supervised learning datasets, eliminating the need for manual annotation while ensuring data quality.
Solution Approach 2:
The system performs self-service by automatically generating its own training data using ILT and extracting relevant features from layout patterns. The training data generation process is automated and requires no external intervention, making the complex task of creating appropriate training data self-contained and reproducible.
3Quantity of substance
If feature dimensions are decreased to increase density, then feature density is improved, but wave nature of light causes distortions and artifacts
Solution Approach 1:
The patent changes the parameter space by transforming the mask design problem into a learning problem where the neural network learns optimal mask parameters directly from training data. The model adjusts mask parameters to compensate for optical distortions automatically, enabling high-density patterns to be manufactured with acceptable accuracy despite wave nature effects.
Data Source
Figure 1A
Figure 1B
Figure 2
AI summary
Techniques relating to synthesizing masks for use in manufacturing a semiconductor device are disclosed. A plurality of training masks, for a machine learning (ML) model, are generated by synthesizing one or more polygons, relating to a design pattern for the semiconductor device, using Inverse Lithography Technology (ILT) (106). The ML model is trained using both the plurality of training masks generated using ILT, and the design pattern for the semiconductor device, as inputs (108). The trained ML model is configured to synthesize one or more masks, for use in manufacturing the semiconductor device, based on the design pattern (110).