Neural Network Semiconductor Pattern Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional semiconductor manufacturing processes face challenges in accurately correcting patterns due to errors in develop and etch processes, requiring human intervention and predefined correction rules, which are time-consuming and inefficient.
Innovation Solution
The use of a processor-implemented method involving backward and forward neural networks for pattern correction, where the backward correction neural network adjusts parameters to minimize errors between desired and simulated pattern images, utilizing gradient descent for optimization without requiring human-defined correction rules.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If conventional OPC and PPC methods are used for pattern correction, then correction can be performed based on predefined rules, but the process requires human intervention and is time-consuming
Solution Approach 1:
The system uses neural networks that automatically learn and perform pattern correction without human intervention. The backward correction network and forward simulation network work together to self-correct patterns by minimizing errors between desired and simulated results, eliminating the need for human operators to apply predefined correction rules manually.
Solution Approach 2:
The patent replaces traditional mechanical/optical correction methods (OPC, PPC) with an intelligent software-based neural network system. Instead of using predefined correction rules and manual processes, the system uses machine learning models that automatically learn optimal correction strategies and apply them dynamically.
2Productivity
If neural networks are used for automated pattern correction, then efficiency is improved, but computational complexity increases
Solution Approach 1:
The correction system is divided into two specialized neural networks: a backward correction network that generates corrected patterns and a forward simulation network that validates corrections. This segmentation allows each network to focus on specific tasks, improving overall efficiency while managing computational complexity through functional decomposition.
Solution Approach 2:
The system implements a feedback loop where the forward simulation network evaluates the output of the backward correction network by comparing simulated results with desired patterns. Error information is fed back to adjust and optimize the correction process, enabling continuous improvement while maintaining computational efficiency through iterative refinement.
3Manufacturing precision
If iterative updating of corrected result images is performed, then manufacturing precision is improved, but processing time increases
Solution Approach 1:
The backward correction neural network is pre-trained to generate highly accurate corrected patterns in a single pass, reducing the need for multiple iterative updates. The network learns optimal correction strategies during training, enabling it to produce precise results quickly without extensive iterative refinement during actual operation.
Solution Approach 2:
The system uses computationally efficient approximation methods where the neural networks provide sufficiently accurate corrections without requiring exhaustive iterative optimization. The forward simulation provides quick validation rather than exhaustive verification, balancing precision requirements with processing speed by accepting near-optimal solutions.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A processor-implemented method including generating a first corrected result image of a first desired pattern image using a backward correction neural network provided an input based on the first desired pattern image, the backward correction neural network performing a backward correction of a first process, generating a first simulated result image using a forward simulation neural network based on the first corrected result image, the forward simulation neural network performing a forward simulation of a performance of the first process, and updating the first corrected result image so that an error between the first desired pattern image and the first simulated result image is reduced.