Repository navigation
FabEye v2: wafer-map failure pattern recognition on the real WM-811K dataset, rebuilt from the earlier synthetic pipeline (v1, kept in legacy-v1/).
Results (lot-disjoint test set)
- CNN macro-F1 0.858 (random forest 0.717, GraphSAGE 0.776); a random split changes it by 0.001, so leakage is measured, not assumed
- Lot context does not help: real-neighbour head 0.877 ± 0.010 vs no-neighbour 0.879 ± 0.002, at every label budget
- Class-conditional conformal sets hold on unseen lots (89.3% at a 90% target; worst class 86.2% vs 54.5% with plain calibration)
- Auto-accept rule: 96.2% of wafers accepted at a 1.9% error rate
Service
- CNN exported to ONNX (within 1e-5 of PyTorch), FastAPI with
/predict,/predict/batch(up to 64),/calibration,/health,/metrics(Prometheus) - API-key auth, JSON logs with request IDs, Docker image with a health check; ~412 wafers/s on ONNX Runtime CPU in batches of 32
Used by WaferLens, which calls the service for every sorted wafer of a simulated fab (macro-F1 0.909, Random-class coverage 52%: where the guarantees break under domain shift).