ML-Based Semiconductor Specimen Examination With Low SNR Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor examination technologies face challenges in accurately detecting defects on sensitive layers with low signal-to-noise ratio (SNR) images, as high SNR images required for annotation are difficult to obtain without damaging the specimen, leading to inaccurate training of machine learning models for defect detection and classification.
Innovation Solution
A system and method that utilize a machine learning model trained with specific composition, using a first training image with low SNR and a second high SNR image, both with similar design patterns, to enable accurate examination of semiconductor specimens by aligning and annotating label data from the high SNR image for use with low SNR images, ensuring robust and accurate defect detection and classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high SNR images are used for annotation, then annotation accuracy is improved, but specimen damage increases
Solution Approach 1:
The patent segments the annotation process into two distinct phases: (1) acquiring a high SNR training image to enable accurate annotation, and (2) using the annotated data to train an ML model that can then annotate low SNR inspection images. This segmentation allows the harmful high-dose imaging to be confined only to the training phase rather than applied to all inspection images.
Solution Approach 2:
The patent performs the annotation action preliminarily on a high SNR training image before actual inspection. By pre-annotating the training image with ground truth data, the system prepares the ML model in advance so that during runtime inspection, only low SNR images need to be processed without requiring repeated high-dose annotation attempts.
2Object-affected harmful factors
If low SNR images are used for inspection, then specimen damage is reduced, but defect detection accuracy deteriorates
Solution Approach 1:
The patent introduces an ML model as an intermediary that bridges the gap between low SNR inspection images and accurate defect detection. The model is trained on paired low SNR images and their corresponding high SNR annotated versions, learning to infer ground truth from low quality images, thereby enabling accurate defect detection without directly imaging at high dose.
Solution Approach 2:
The patent creates a computational copy of the annotation process through the ML model. Instead of physically re-acquiring high SNR images for each inspection case, the trained model generates predicted annotations from low SNR images, effectively copying the annotation capability without requiring the original high-dose imaging condition.
3Object-affected harmful factors
If traditional annotation methods are used on low SNR images, then specimen damage is reduced, but annotation quality deteriorates
Solution Approach 1:
The patent segments the imaging process into a dedicated training phase using high SNR images for annotation, and an inspection phase using low SNR images for model inference. This segmentation ensures that low SNR images are only used where they are acceptable (for model input), while high SNR imaging is reserved solely for the annotation task where quality is critical.
Data Source
AI summary
There is provided a system and method of runtime examination of a semiconductor specimen. The method includes obtaining a runtime image representative of an inspection area of the specimen, the runtime image having a relatively low signal-to-noise ratio (SNR); and processing the runtime image using a machine learning (ML) model to obtain examination data specific for a given examination application, wherein the ML model is previously trained for the given examination application using one or more training samples, each training sample representative of a respective reference area sharing the same design pattern as the inspection area and comprising: a first training image of the respective reference area having a relatively low SNR; and label data indicative of ground truth in the respective reference area pertaining to the given examination application, the label data obtained by annotating a second training image of the respective reference area having a relatively high SNR.