Platform — Defect Detection

23-class defect classification, every die, every wafer

Convolutional neural networks trained on production semiconductor defect datasets. Output includes defect class, confidence score, and die X/Y coordinates for MES integration.

DEFECT MAP — Wafer W-29A-0471
Die X/YClassConfidenceAction
X12/Y08PASS—Continue
X13/Y08Particle ≥1µm0.94Flag
X14/Y08PASS—Continue
X04/Y14Scratch ≥5µm0.97Flag
X22/Y03Edge Chip0.72Review
X19/Y06Particle ≥0.5µm0.91Flag
23 Defect Types

Classification categories

The base model ships with 23 trained defect categories. Custom categories are added through supervised retraining — typically 2–3 weeks with customer-provided labeled samples.

Particle Defects

  • PART-SM Particle <0.5µm
  • PART-MD Particle 0.5–1µm
  • PART-LG Particle ≥1µm
  • PART-CLST Particle Cluster
  • CONT-MTL Metal Contamination
  • CONT-ORG Organic Residue

Surface Defects

  • SCR-FINE Fine Scratch <5µm
  • SCR-WIDE Wide Scratch ≥5µm
  • PIT Surface Pit
  • CRACK Micro-Crack
  • SLIP Crystal Slip Line
  • DIMPLE Surface Dimple

Edge / Handling

  • EDGE-CHIP Edge Chip
  • EDGE-CRK Edge Crack
  • NOTCH-ERR Notch Anomaly
  • HANDLING Handling Artifact

Pattern / Process

  • BRIDGE Metal Bridge
  • OPEN Line Open
  • CMP-ERR CMP Non-uniformity
  • LAYER-SHIFT Layer Misalignment
  • VOID Via Void
  • DIEL-PKG Dielectric Peel
  • CUSTOM Customer-defined
Model Design

Per-recipe adaptive calibration

The base CNN model handles feature extraction — detecting anomalies against the expected die pattern at a given process layer. A per-recipe calibration layer adjusts confidence thresholds based on the specific optical characteristics of your tool, your process node, and your layer.

Calibration runs automatically using the first 50 wafers of a new recipe. As lot data accumulates, the threshold profile self-adjusts within bounds set by your process engineer. This keeps false-positive rates stable across tool aging, resist chemistry changes, and ambient cleanroom variation.

False-Positive Control Methodology
Close-up view of semiconductor wafer surface showing microscopy-level detail of wafer die pattern under inspection lighting

Evaluate defect classification on your wafer set

Provide sample wafers with known defect maps. We run blind classification and compare against your gold standard.

Request Evaluation