Substrate Region Classification for Defect-Aware Analyte Deposition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Graphene substrates used for low-energy electron holography and similar transmission mode analytic procedures are compromised by surface contaminants and structural defects, leading to reduced imaging performance and yields.
Innovation Solution
Implementing low-energy electron point projection (LEEPP) imaging to scan and classify substrate regions, using sparse and secondary scans to identify clean regions, optimize analyte deposition patterns, and employ a trained machine learning tool for consistent image classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a comprehensive scan of the entire substrate is performed to identify all suitable regions, then the coverage of suitable regions is improved, but the scan time increases significantly
Solution Approach 1:
The substrate is divided into multiple smaller regions or tiles that are scanned individually. This segmentation allows the system to perform comprehensive coverage by processing smaller units in parallel or sequentially, reducing the time required to scan the entire substrate while maintaining complete coverage of all suitable regions.
Solution Approach 2:
The system performs preliminary classification of substrate regions using machine learning models to identify and prioritize suitable regions before final analysis. This preliminary action filters out obviously unsuitable regions, allowing the comprehensive scan to focus only on promising areas, thereby reducing overall scan time while maintaining reliability.
2Measurement precision
If traditional imaging methods are used to characterize substrate regions, then the imaging performance is maintained, but surface contaminants and structural defects compromise the yield
Solution Approach 1:
Traditional mechanical scanning and imaging methods are replaced with machine learning-based classification systems that process substrate images automatically. The machine learning models are trained to recognize and filter out regions with surface contaminants and structural defects, maintaining high imaging precision while significantly improving the yield of suitable regions for analyte deposition.
Solution Approach 2:
The system incorporates feedback loops where imaging results are continuously fed back to the machine learning classification models. This feedback mechanism allows the system to learn from previously identified suitable and unsuitable regions, improving its ability to distinguish high-quality substrate areas and thereby increasing the yield of suitable regions while maintaining imaging performance.
3Reliability
If a dense scan pattern is used to maximize coverage, then the number of suitable regions identified is improved, but the scan time and computational resources increase
Solution Approach 1:
The scan pattern is made dynamic rather than static, allowing the system to adjust the scanning density and pattern in real-time based on the substrate characteristics and machine learning predictions. This dynamic approach enables the system to identify a high number of suitable regions by concentrating scan resources on areas with higher probability of containing suitable regions, thereby reducing overall scan complexity while maintaining reliability.
Solution Approach 2:
Different regions of the substrate are scanned with different densities based on their local characteristics and predicted suitability. Areas with higher predicted suitability receive denser scanning, while areas with lower suitability receive sparser scanning. This local quality approach maximizes the number of suitable regions identified while minimizing overall scan complexity and computational resources.
4Measurement precision
If manual classification of substrate regions is performed, then the accuracy of defect detection is maintained, but the time consumption and subjectivity increase
Solution Approach 1:
Manual classification by experts is replaced with automated machine learning classification systems. The machine learning models are trained on extensively labeled datasets to achieve accuracy comparable to or exceeding human experts, while dramatically reducing classification time. The system processes hundreds of substrate regions per second, eliminating the time consumption and subjectivity inherent in manual classification.
Solution Approach 2:
The system creates and uses digital copies of substrate regions in the form of processed images and feature representations that can be rapidly analyzed by machine learning models. These digital copies allow for rapid, repeatable classification without the time and variability constraints of manual inspection, maintaining high accuracy while reducing classification time significantly.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Enhances the yield of analyte characterization data by efficiently identifying suitable substrate regions, reducing scan time, and improving the accuracy of defect detection and analyte deposition, with machine learning achieving over 96% accuracy in image classification.
Implementation Method 1
low-energy electron point projection (LEEPP) imaging
Implementation Method 2
transmission mode analytic procedures
Data Source
AI summary
Methods and apparatus are disclosed for classifying regions of a substrate prior to performing an analytic procedure involving analyte particles supported on the substrate. Regions of the substrate are sparsely scanned using low-energy electron point projection (LEEPP) imaging. Regions are classified as suitable for the analytic procedure (or not) based on defects visible in the respective images. Given one suitable region, neighboring regions are scanned to increase the yield of suitable regions. Based on a distribution of suitable regions, a deposition pattern is planned, and analyte is deposited according to the plan. Following deposition, suitable regions are scanned again to identify or count visible analyte molecules visible. Based on numbers of analyte molecules found, regions are earmarked for analyte characterization, e.g. by low-energy electron holography and reconstruction. A trained machine learning classifier provides consistent, accurate image classification across a range of defect types.