Charged-Particle Microscopy Defect Localization by Subtractive Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for charged-particle microscopy struggle with accurately localizing defective structures due to the infinite variability of defect types, as finite training datasets fail to capture the nearly infinite structural diversity of damaged specimens.
Innovation Solution
A system utilizing two machine learning models: one trained to localize non-defective structures and another to identify all structures, with the subtraction of non-defective localizations from defect-agnostic localizations to pinpoint defective areas.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a machine learning model is trained to localize defective structures directly, then it can identify defects, but the infinite variability of defect types causes the model to fail due to insufficient training data coverage
Solution Approach 1:
Instead of training a model to directly identify defective structures, the patent inverts the approach by training a model to identify non-defective structures. The defect localization is then achieved by subtracting the non-defective localizations from the total structure localizations. This inversion works because non-defective structures have a constrained state-space that can be adequately captured in training data, whereas defective structures have infinite variability that cannot be fully represented in finite training datasets.
2Reliability
If traditional defect detection methods are used, then they can handle known defect types, but they fail to generalize to the nearly infinite structural diversity of damaged specimens
Solution Approach 1:
The patent applies inversion by shifting the detection target from defective structures (infinite variability) to non-defective structures (constrained state-space). This allows the model to learn reliable patterns from abundant non-defective examples and then use set subtraction to identify defects, achieving both reliability and adaptability to diverse defect types.
Solution Approach 2:
The patent extracts the problem of infinite defect variability by separating the detection task into two components: (1) detecting all structures regardless of defect status, and (2) detecting only non-defective structures. By extracting and analyzing the difference between these two sets, the method handles structural diversity without being overwhelmed by it.
3Ease of manufacture
If finite training datasets are used to train defect detection models, then training is feasible, but the models cannot capture the nearly infinite structural diversity of damaged specimens
Solution Approach 1:
The patent resolves this contradiction by inverting the training target from defective to non-defective structures. Finite training datasets are sufficient for capturing non-defective structures because they have a constrained state-space with limited variations. This inversion maintains training feasibility while dramatically improving adaptability to diverse defect types through the subtraction approach.
Data Source
AI summary
Systems/techniques are provided for facilitating subtractive defect localization for charged-particle microscopy. In various embodiments, a system can access an image captured by a charged-particle microscope, wherein the image depicts a specimen. In various aspects, the system can localize one or more defective instantiations of a structure of interest of the specimen, based on execution of a first machine learning model that is trained to localize non-defective versions of the structure of interest.