Two-Pass ML Defect Detection for Semiconductor Specimens
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor defect examination technologies face challenges in accurately detecting defects due to limited training data, particularly scarce true defects which are hard to detect and annotate, leading to insufficient variance in training samples and high false alarm rates.
Innovation Solution
A two-pass machine learning approach is employed, where a first ML model is trained using synthetic defective images and nominal images, and a second ML model is trained using a subset of images labeled with detected defects from the first model, enhancing defect detection performance by improving precision and recall.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If synthetic defective images are used to train the first ML model, then the model can be trained with available data, but the detection precision and recall remain insufficient due to lack of true defect variance
Solution Approach 1:
The first ML model is trained in advance using synthetic defective images and nominal images to establish an initial defect detection capability. This preliminary training enables the system to identify potential defects in real images, which are then used to create a more diverse second training set for the second model, ultimately improving detection precision and recall.
2Reliability
If more true defects are collected for training, then the detection performance would improve, but the annotation cost and time increase significantly
Solution Approach 1:
Instead of collecting and annotating numerous true defect images directly, the system uses the first ML model to generate candidate defect locations in real images, which serve as proxy training data for the second model. This copying approach creates a second training set with sufficient variance without requiring manual annotation of every defect, significantly reducing annotation time while maintaining detection reliability.
3Quantity of substance
If the first ML model detects all potential defects, then the training data would be comprehensive, but the false alarm rate increases
Solution Approach 1:
The system applies different quality thresholds to different regions or types of detected defects. By analyzing the characteristics of defects identified by the first model, the system selectively uses high-confidence defect detections as training samples for the second model, while filtering out low-confidence detections that would contribute to false alarms. This local quality differentiation ensures comprehensive training data while maintaining high precision.
Data Source
AI summary
There is provided a system and method of semiconductor specimen examination. The method includes obtaining a plurality of images of a semiconductor specimen acquired by an examination tool; processing the plurality of images using a first machine learning (ML) model for defect detection, thereby obtaining, from the plurality of images, a set of images labeled with detected defects, wherein the first ML model is previously trained using a first training set comprising a subset of synthetic defective images each containing one or more synthetic defects, and a subset of nominal images; and training a second ML model using a second training set comprising at least part of the set of images labeled with detected defects, wherein the second ML model, upon being trained, is usable for defect detection with improved detection performance with respect to the first ML model.