Traditional AOI only flags suspect defects; verification and classification still fall on inspectors, where runaway overkill drags capacity and distorts yield data.
This fab's front-end AOI generated a large volume of suspect defect points per wafer, but the equipment could not separate real defects from false alarms such as scratch glare or particle residue. Every point had to be manually verified and classified by inspectors on microscope images; even three shifts struggled to cover peak output, and the review backlog pushed straight to shipment deadlines.
Over time the fab's AOI overkill rate held above 20%, meaning roughly one in five alarms was a good die wrongly rejected. Large numbers of qualified wafers were sent back for re-inspection, burning labor hours and distorting defect distribution statistics, leaving process engineers unable to locate the true anomaly source.
DaoAI Solution
DaoAI integrated the AI-ADC auto defect classification module into the AOI review station: suspect point images from AOI feed directly into AI-ADC, which separates true from false at microscope/micron-level precision and sorts them into predefined defect classes. For classes with imbalanced historical samples, APDT positive-sample learning combined with the DaoAI World model generalization reduces reliance on heavy manual labeling.
- AI-ADC re-verifies AOI alarm points and filters out false alarms
- Auto-classifies by particle, scratch, pattern defect, etc. and writes back to MES
- APDT positive-sample learning covers low-frequency defects, shortening ramp-up
- Classification results feed back to process engineering to locate true anomalies
Overkill fell from above 20% to under 2%, shifting inspectors from per-wafer review to anomaly arbitration.
After deployment, the fab's AOI overkill rate dropped from above 20% to under 2%, the man-to-machine ratio in review fell about 90%, and inspectors only handle a small set of borderline samples. Review is no longer a bottleneck, whole-line throughput rose about 30%, and defect classification data became usable again.