Symmetric CycleGAN Registration for SEM and Design Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image alignment techniques for semiconductor inspection, such as linear filters and pattern-to-design alignment, fail to accurately align scanning electron microscope (SEM) images with design images due to differing modalities and noise, making defect detection challenging, especially at smaller design rules.
Innovation Solution
Employing a symmetric CycleGAN neural network architecture for bidirectional mappings between design and SEM image domains, using symmetric convolutional kernels and focal normalized cross-correlation regularization to align SEM and design images without paired samples, thereby generating high-fidelity synthetic SEM images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional alignment techniques (linear filters or pattern-to-design alignment) are used to align SEM images with design images, then the alignment process is simple and fast, but alignment accuracy deteriorates due to differing modalities and SEM-specific noise, resulting in registration errors up to 4-5 pixels
Solution Approach 1:
The patent introduces a symmetric CycleGAN as an intermediary system that translates design images into synthetic SEM images. This intermediary translation layer bridges the modality gap between design and SEM images, enabling accurate alignment by comparing images in the same modality (SEM-to-SEM) rather than directly aligning design images with noisy SEM images.
Solution Approach 2:
The patent creates a copy of the design image in the SEM modality through the symmetric CycleGAN translation process. By generating a synthetic SEM image that replicates the appearance and noise characteristics of real SEM images, the system enables accurate defect detection without directly comparing incompatible image modalities.
2Measurement precision
If paired and aligned samples of design and SEM images are used to train a design-to-SEM translator, then translation accuracy improves, but obtaining precise correspondences becomes difficult and time-consuming
Solution Approach 1:
The patent inverts the traditional training approach by using unpaired images instead of requiring paired aligned samples. The symmetric CycleGAN learns from independent design and SEM image datasets without needing explicit correspondences, reversing the conventional supervised learning paradigm to eliminate the time-consuming pairing process.
Solution Approach 2:
The symmetric CycleGAN performs self-alignment through cycle consistency loss, where the model automatically learns to translate images back and forth between domains without external supervision or paired training data. The system serves itself by using the translation consistency as its own training signal rather than requiring externally provided paired samples.
3Ease of manufacture
If asymmetric CycleGAN is used for image translation, then training without paired samples is enabled, but geometric offsets are introduced in the translated images
Solution Approach 1:
The patent employs asymmetric convolutional kernels in the generator architecture to capture directional features and patterns in semiconductor images. This controlled asymmetry in the kernel design allows the model to learn geometric transformations while maintaining accuracy, preventing the geometric offsets that would result from purely symmetric processing.
Solution Approach 2:
The patent modifies the CycleGAN architecture by incorporating symmetric loss functions and specific kernel constraints that change the optimization parameters during training. These parameter changes ensure that while the network learns from unpaired data, it maintains geometric fidelity by enforcing consistency constraints that prevent offset introduction.
Data Source
AI summary
A synthetic image of a semiconductor device structure is generated from a design image using a symmetric CycleGAN. The symmetric CycleGAN includes a neural network architecture that learns bidirectional mappings between a design image domain and a workpiece image domain. The synthetic image is aligned to a workpiece image (e.g., a SEM image) of the device structure on the workpiece corresponding to the synthetic image thereby generating an aligned image. The aligning is provided by domain translation performed by the symmetric CycleGAN.