ML Model Reduces Charging Artifacts in SEM Inspection Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for reducing charging artifacts in SEM images are either slow, lack physics-based explainability, or are unable to effectively eliminate charging artifacts, which can lead to errors in defect detection and critical dimension measurement.
Innovation Solution
A physics-based machine learning model, such as an autoencoder, is trained using simulated inspection images with and without charging artifacts to learn decoupled features. This model is then applied to actual inspection images to generate output images with reduced charging artifacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If experimental techniques such as quadscan are used to avoid charging artifacts, then charging artifacts are reduced, but throughput decreases due to slow processing
Solution Approach 1:
The patent applies preliminary action by pre-training a machine learning model using simulated inspection images with known charging artifacts before actual inspection. The model learns to identify and correct charging artifacts in advance, enabling rapid processing during actual inspection without requiring slow experimental techniques like quadscan during production
Solution Approach 2:
The patent replaces the mechanical/experimental quadscan technique with a computational machine learning approach. Instead of physically scanning the sample multiple times in different directions, the system uses a trained neural network to predict and remove charging artifacts from single-scan images, dramatically improving throughput while maintaining artifact reduction
2Productivity
If numerical image processing techniques are used to reduce charging artifacts, then processing speed improves, but physics-based explainability is lost
Solution Approach 1:
The patent changes the parameters of the machine learning model during training by exposing it to simulated images with varying charging conditions, beam energies, and scan parameters. This enables the model to learn physics-based relationships between imaging parameters and charging artifacts, providing explainability while maintaining fast processing speed during actual inspection
3Reliability
If the beam dose per frame is reduced to minimize charging artifacts, then charging artifacts are reduced, but image quality and signal-to-noise ratio deteriorate
Solution Approach 1:
The patent introduces a machine learning model as an intermediary between the raw inspection image and the final output. The model takes low-dose images with charging artifacts as input and produces corrected images with reduced artifacts, enabling the use of lower beam doses while maintaining image quality through computational correction rather than requiring high-dose imaging
4Reliability
If multiple SEM frames are used to reduce charging artifacts, then charging artifacts are reduced, but acquisition time increases
Solution Approach 1:
The patent uses simulated inspection images as copies during the training phase to create a virtual dataset that mimics real charging artifact conditions without requiring multiple actual scans. This allows the model to learn from numerous examples rapidly, and during production, only single frames need to be processed, dramatically reducing acquisition time compared to requiring multiple actual SEM frames
Data Source
AI summary
Systems and methods for reducing charging artifacts in an inspection image include obtaining a set of inspection images, in which each of the set of inspection images includes a charging artifact; and training a machine learning model using the set of inspection images as input, in which the machine learning model outputs a set of decoupled features of the set of inspection images.