Challenge
It is required to automatically identify the cause of defects using scanning electron microscope (SEM) images showing defects detected by semiconductor wafer defect inspection equipment.
Although machine learning is an effective method, it has the following problems.
- Imbalanced data
- Lack of training data
- Differences of images in a class may outweigh differences of images between classes
Solution
Our solution is a combination of many methods: Anomaly Detection, Object Classification, Region Analysis
- The Anomaly Detection algorithm only needs good samples to train, so the lack of training data does not affect the result.
- Next, we based on the characteristics of the error area to distinguish two main types of defects, scratches or foreign bodies.
- With scratches, the severity of the defect is assessed by size. With foreign bodies, use the Object Classification algorithm to distinguish the type of foreign body.
Outcome
Automatically classifies known defect causes and detects unknown defects that do not belong to any class. Achieved 96% classification accuracy.
