Conditional Masked Autoencoder for Multi-Modal Defect Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for defect detection in semiconductor fabrication face challenges in accurately predicting the location, type, and structure of defects due to high computational costs and the inability to handle images of different modalities, such as design and inspection images.
Innovation Solution
A system and method utilizing a conditional masked autoencoder (MAE) with two encoders, one for optical images and one for design images, that applies targeted masking to generate a reference image by leveraging defect-free regions to learn design-to-SEM mapping, effectively reconstructing images to identify defects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If physics-based approaches or deep convolutional neural networks are used for optical image modeling, then defect detection capability is improved, but computational cost increases significantly
Solution Approach 1:
The patent uses a generator network to create a synthesized reference image (copy) that mimics the characteristics of the actual reference image. This synthesized reference image is then used for defect detection instead of requiring complex physics-based modeling or extensive computational resources, thereby reducing computational cost while maintaining defect detection capability
Solution Approach 2:
The patent introduces a conditional masked autoencoder as an intermediary component that bridges the gap between the inspection image and the reference image. This autoencoder processes the inspection image to generate a masked version, which is then combined with design image data to create the synthesized reference, reducing the need for direct complex comparisons
2Adaptability or versatility
If traditional optical image modeling methods are used, then defect detection is performed, but the system cannot handle images of different modalities (design and inspection images)
Solution Approach 1:
The patent employs a universal neural network architecture that can process multiple types of input images (inspection images and design images) through the same encoder networks. The conditional masked autoencoder is designed to handle different image modalities uniformly, enabling the system to integrate and compare diverse image types without requiring separate processing pipelines, thus improving adaptability while maintaining detection accuracy
Data Source
AI summary
A method for defect detection includes receiving an optical image of a sample and dividing the received optical image include a plurality of image patches. The method includes generating a masked optical image by selectively masking the image patches based on characterization data. The method includes providing the masked optical image to a fist encoder. The method includes receiving a design image and providing it to a second encoder. The method includes generating a first set of feature vectors using the first encoder and generating a second set of feature vectors using the second encoder. The method includes concatenating the feature vectors from the first encoder and the feature vectors from the second encoder and generating a reference image of the sample based on the concatenated feature vectors using a decoder, where the generated reference image is a reconstructed image of the optical image and the design image.