CNN reads the fail-die pattern → names the signature → points at a root-cause direction.
Demo on the public WM-811K dataset (real production wafer maps); true labels shown for verification.
auto-decide when ≥ 90% sure less sure → goes to human review
How it works
Input — a wafer map. The pass/fail result for every die on one wafer, straight from test. No images, no sensor data — just the map.
A neural network names the pattern. Trained on 173,000 engineer-labeled maps, it answers one question: which of 9 known failure signatures does this shape match (edge ring, scratch, center cluster, …) — and how sure is it.
The pattern points at the cause. Each signature physically constrains what can have produced it — a ring at the edge implicates chamber uniformity, not handling; a thin line implicates handling, not plasma. That signature→cause mapping is built-in engineering knowledge, shown alongside every call. The model finds the pattern; the physics does the diagnosing.
Sure → automatic, unsure → human. High-confidence calls are auto-cleared or auto-flagged for SPC trending; ambiguous maps are escalated to engineer review. The cutoff is the slider above. The Cluster map tab shows wafers grouped purely by learned similarity — a failure mode that fits none of the 9 categories would appear there as a new cluster.
sampling…
Each dot is a wafer, positioned by the CNN's learned embedding (PCA to 2-D).
Similar failure patterns land together — a novel signature would appear as its own cluster.
Hover to inspect, click to open that wafer in Single view.