Convolutional Neural Network for Photomask Pattern Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing pattern generation processes for photomasks often result in binary patterns that differ from the desired patterns due to corner rounding and other physical phenomena, leading to inaccuracies such as linewidth variations and proximity effects.
Innovation Solution
A method involving the generation of training data for a convolutional neural network using a physical model and reinforcement learning to derive corrected digital pattern descriptions, which iteratively updates candidate descriptions based on similarity with desired patterns, ensuring closer alignment between predicted and desired binary patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a digital pattern description is directly used to produce a photomask, then the production process is simple and fast, but the resulting binary pattern differs from the desired pattern due to corner rounding and physical phenomena
Solution Approach 1:
The system performs preliminary correction of the digital pattern description before photomask production by using a trained convolutional neural network to predict and compensate for corner rounding and other physical effects that will occur during manufacturing. This advance correction ensures that the final photomask pattern matches the desired pattern despite the inherent physical distortions in the production process
Solution Approach 2:
The patent replaces traditional mechanical or manual pattern correction methods with a data-driven convolutional neural network approach. The system uses machine learning models trained on physical models and empirical data to automatically predict and correct pattern deviations, substituting complex iterative manual adjustment processes with an automated intelligent system
2Manufacturing precision
If corner rounding and physical phenomena are compensated for through iterative correction, then pattern accuracy improves, but the computational time and processing steps increase
Solution Approach 1:
The system performs preliminary correction of the digital pattern description before photomask production by using a trained convolutional neural network to predict and compensate for corner rounding and other physical effects that will occur during manufacturing. This advance correction ensures that the final photomask pattern matches the desired pattern despite the inherent physical distortions in the production process
Solution Approach 2:
The system creates a virtual model or copy of the photomask production process using physical models and simulation data. By working with these digital copies and predictions rather than actual iterative physical production cycles, the system can quickly evaluate and correct patterns without the time cost of repeated physical manufacturing and measurement cycles
Data Source
AI summary
Methods, a non-transitory computer-readable storage medium, devices, and a system in relation to training a convolutional neural network for deriving corrected digital pattern descriptions from digital pattern descriptions for use in a process for producing photomasks are disclosed. A reinforcement learning agent is trained to derive corrected digital pattern descriptions from respective digital pattern descriptions. The training is based on a first plurality of generated digital pattern descriptions and an obtained physical model using which predicted binary patterns of photomasks can be derived that would result from inputting digital pattern descriptions to the process for producing photomasks. A second plurality of digital pattern descriptions is then generated, and corresponding corrected digital pattern descriptions are generated using the trained reinforcement learning agent, thereby generating training data. The training data can be used to train a convolutional neural network to derive corrected digital pattern descriptions from digital pattern descriptions, the trained neural network can be used to derive a corrected digital pattern description, and the corrected digital pattern description can be used to produce a photomask according to the corrected digital pattern description.