Problem
Wafer inspection produces millions of defect images per fab per week. Manual classification is slow, inconsistent between operators, and misses systematic yield signatures that only emerge across lots.
Solution
An end-to-end defect classification pipeline: a convolutional backbone fine-tuned on labeled inspection imagery, an active-learning loop that routes low-confidence predictions to engineers, and a spatial-signature model that clusters die-level defect maps into known process failure modes.
Architecture
- PyTorch training pipeline with mixed-precision fine-tuning on fab inspection imagery
- ONNX Runtime inference service behind a FastAPI layer for near-line classification
- Active-learning queue that prioritizes uncertain samples for human review
- Wafer-map clustering (DBSCAN on radial features) to detect spatial process signatures
Outcome
- Cut manual review volume dramatically by auto-classifying high-confidence defects
- Surfaced spatial signatures tied to specific process steps, shortening root-cause loops
- Model retraining cycle reduced from weeks to days via the active-learning loop
Lessons
- “Label quality beats model architecture: the active-learning loop mattered more than the backbone”
- “Fab data pipelines are the hard part; the model is the last 10%”
- “Engineers trust a model that shows its uncertainty, not one that hides it”
Stack
- Python
- PyTorch
- ONNX
- FastAPI
- PostgreSQL
- Docker