Semiconductor Image Denoising for Low-Dose High-SNR Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor examination processes face challenges in obtaining high-quality images with high Signal-to-Noise Ratio (SNR) due to low electron beam doses, leading to physical damage and image artifacts, which affect metrology operations and annotation accuracy.
Innovation Solution
A machine learning (ML) model is trained using a training set of low SNR images associated with high SNR images to enhance image denoising, processing low SNR images to produce high SNR images, thereby reducing electron beam dose and physical effects while improving measurement precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If low electron beam dose is used to reduce physical damage, then specimen damage is reduced, but image quality and SNR deteriorate
Solution Approach 1:
The ML model is trained in advance using paired low-SNR and high-SNR images to learn the mapping relationship. During examination, the pre-trained model processes low-dose images to generate high-quality output images, eliminating the need for high-dose imaging while maintaining image quality.
Solution Approach 2:
The ML model acts as an intermediary between the low-dose input images and the desired high-quality output. It processes the noisy low-SNR images and transforms them into high-SNR images, effectively mediating the trade-off between dose and quality.
2Measurement precision
If high electron beam dose is used to improve image quality, then SNR is improved, but physical damage and image artifacts increase
Solution Approach 1:
The ML model is trained on low-SNR images that inherently contain noise and artifacts. During training, the model learns to recognize and correct these artifacts by comparing with corresponding high-SNR ground truth images. The harmful noise patterns in low-dose images become training data that teaches the model how to remove similar artifacts.
3Measurement precision
If multiple runs of scanning are performed to improve image quality, then SNR is improved, but examination time increases
Solution Approach 1:
The ML model learns from training data consisting of multiple scanned runs and their corresponding high-SNR references. Once trained, it can generate high-SNR images from single or few runs by applying the learned transformation, effectively copying the quality improvement without requiring multiple actual scans during examination.
Data Source
AI summary
There is provided an image generation system and method. The method comprises obtaining a runtime image of a semiconductor specimen with a low Signal-to-noise ratio (SNR), and processing the runtime image using a machine learning (ML) model to obtain an output image with a high SNR. The ML model is previously trained using a training set comprising a plurality of low SNR images associated with a high SNR image. The plurality of low SNR images correspond to a plurality of sequences of frames acquired in a plurality of runs of scanning a first site of the specimen. The high SNR image is generated based on the plurality of low SNR images. The training comprises, for each low SNR image: processing the low SNR image by the ML model to obtain predicted image data, and optimizing the ML model based on the predicted image data and the high SNR image.