Early in new-process ramp, defect types appear sporadically, often fewer than ten samples per class, leaving negative-sample-driven supervised models starved.
During new-process introduction at this advanced packaging plant, defect morphologies kept evolving, and some rare defects appeared only a few times across an entire lot. Conventional supervised classification needs hundreds to thousands of negative samples per class to converge, which is simply unattainable for rare defects in the short term, leaving them in a persistent model blind spot.
To work around the sample shortage, the team could only loosen criteria or add manual backstops, leading to either soaring overkill or rare-defect escapes. Ramp-up most fears exactly these unknown new defects; once one escapes to the customer, the traceback cost is steep.
DaoAI Solution
DaoAI adopted the APDT positive-sample learning route: the model mainly learns the feature distribution of good samples, treating any deviation as suspect, so no large negative-sample collection is needed per rare defect. Paired with the DaoAI World model to generalize defect morphologies, even classes with single-digit samples are stably detected and classified at microscope/micron-level precision.
- APDT positive-sample learning models good samples, no negatives to gather for rare defects
- DaoAI World model generalizes unseen defect morphologies
- New defects can be onboarded fast without waiting for batch accumulation
- Rare and common classes share one classification output into the existing grading flow
Modeling on good samples ends the rare-defect blind spot; new defects can be onboarded same-day.
After deployment, rare-defect classification accuracy exceeded 96%, the rare-defect escape rate was around 0.2%, and the cycle from a new defect's appearance to model onboarding shrank from weeks to days. Unknown-defect risk during ramp-up was effectively contained, freeing the process team to focus on process improvement rather than sample collection.