Deep Generative Model Alignment for Semiconductor Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for semiconductor inspection face challenges in aligning design and specimen images due to the absence of sufficient alignment targets and incomplete design data, leading to difficulties in creating physics-based models for accurate image rendering and defect detection.
Innovation Solution
A system utilizing a deep generative model to transform actual information for alignment targets on a specimen from one type of information to another, enabling the determination of offsets for precise alignment and image-to-design matching without relying on complete design data or sufficient alignment targets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If physics-based models are used for image rendering and alignment, then measurement precision is improved, but device complexity increases due to requiring complete design data and sufficient alignment targets
Solution Approach 1:
A deep generative model is introduced as an intermediary between incomplete design data and the alignment process. This generative model synthesizes missing design data and alignment targets, enabling the physics-based alignment model to function without requiring complete input data. The generative model acts as a bridge that transforms available partial information into the complete data structure needed for precise alignment.
Solution Approach 2:
The system performs preliminary data generation and completion before the actual alignment process. The deep generative model pre-synthesizes missing design data and alignment targets in advance, so that when the physics-based alignment model executes, all necessary data is already prepared. This preliminary action eliminates the need for complete original design data during the critical alignment measurement phase.
2Manufacturing precision
If complete design data is required for accurate image rendering, then manufacturing precision is improved, but loss of information increases when design data is incomplete
Solution Approach 1:
The deep generative model creates synthetic copies of missing design data based on available partial design information. Instead of requiring the original complete design data, the system generates accurate copies of the missing portions through the generative model, which learns the underlying patterns and structures from the available data. These synthesized copies are sufficient for accurate image rendering and alignment purposes.
Solution Approach 2:
The system transforms the problem from requiring complete design data to working with incomplete data by changing the parameters of the data representation. The deep generative model learns to map from partial design data parameters to complete design data parameters, effectively changing the input requirements of the image rendering process. This parameter transformation allows accurate rendering even when original design data is incomplete.
3Measurement precision
If alignment targets are required for PDA training, then measurement precision is improved, but difficulty of detecting and measuring increases when alignment targets are insufficient
Solution Approach 1:
The deep generative model enables the system to be self-sufficient by generating its own alignment targets and complete design data from available partial information. Instead of relying on externally provided alignment targets that may be insufficient, the system creates its own training data through the generative model, which synthesizes alignment targets consistent with the actual specimen features. This self-service capability eliminates the bottleneck of finding sufficient alignment targets.
Data Source
AI summary
Methods and systems for deep learning alignment for semiconductor applications are provided. One method includes transforming first actual information for an alignment target on a specimen from either design data to a specimen image or a specimen image to design data by inputting the first actual information into a deep generative model such as a GAN. The method also includes aligning the transformed first actual information to second actual information for the alignment target, which has the same information type as the transformed first actual information. The method further includes determining an offset between the transformed first actual information and the second actual information based on results of the aligning and storing the determined offset as an align-to-design offset for use in a process performed on the specimen.