How we keep false positives below 0.1%
Per-recipe adaptive calibration is the core mechanism. The calibration layer adjusts confidence thresholds based on accumulated lot data, keeping false-positive rates stable even as tools age and process conditions drift.
Four-stage false-positive control
The first 50 production wafers run with conservative confidence thresholds. All flagged items are reviewed by the process engineer to identify false positives. The calibration layer builds a baseline threshold profile for this recipe × tool combination.
False positive feedback from engineering review adjusts calibration thresholds in bounded increments. Convergence typically completes by lot 100. Once false-positive rate drops below 0.1%, the recipe transitions to production mode.
In production mode, the calibration layer monitors false-positive rate per lot and adjusts thresholds within process engineer-defined bounds when drift is detected. Tool aging, resist chemistry changes, and cleanroom temperature variations are all handled automatically.
Each tool gets its own calibration profile even when running the same recipe. Tool T47-B and T47-C will have independent threshold profiles derived from their individual optical characteristics. A tool swap doesn't reset calibration history — profiles transfer with re-validation.
How confidence thresholds work in practice
The calibration layer controls false-positive rate by adjusting the per-class confidence threshold — the minimum softmax probability required for a detection to generate a defect record in the output event stream.
What the classifier outputs
For each inspection region that passes the initial anomaly scan, the CNN outputs a softmax probability vector across all defect classes — particle small (PART-SM), particle medium (PART-MD), particle large (PART-LG), scratch fine (SCR-FINE), scratch wide (SCR-WIDE), CMP error (CMP-ERR), layer shift (LAYER-SHIFT), and eight additional classes.
The highest-probability class becomes the defect type label, and the probability value becomes the confidence score. A threshold of 0.75 means: only generate a defect event if the classifier assigns ≥75% probability to its top-class prediction.
Raising the threshold reduces false positives but risks letting borderline real defects through at reduced confidence. The calibration layer tracks this tradeoff per class by monitoring confirmed-real vs. confirmed-false events from engineering review dispositions.
How thresholds are updated
After each lot, the calibration layer computes the per-class FPR from any engineering review dispositions recorded in the review queue. If a class exceeds its FPR budget, the threshold is incremented by a bounded step (default: +0.02 per lot, configurable by process engineer). If a class has no false positives and a confirmed-real rate above baseline, the threshold may decrease by a smaller step (+0.01 per lot) to recover sensitivity.
The bounds are hard: threshold cannot be set above 0.95 (would miss too many borderline real defects) or below 0.50 (classifier is below discriminative threshold). The process engineer can override any auto-tuned value and lock it for a defined number of lots.
What drives false positives in wafer inspection
Most false positives trace to one of four sources. Per-recipe calibration addresses all four.
Optical aberrations
Every inspection tool has optical quirks — edge diffraction, vignetting, focus gradient. These create detection events that aren't real defects. Tool-specific calibration suppresses optical artifact signatures.
Resist chemistry variation
Resist reflectance changes with lot-to-lot chemistry variation. A model calibrated on one resist batch may see surface contrast differences as particles when running a new batch.
Ambient cleanroom variation
Temperature and humidity fluctuations affect optical capture. The calibration layer treats these as drift events and adjusts thresholds accordingly.
Tool aging
Optical components degrade over tool lifetime. A static threshold profile that worked at tool qualification will drift toward higher false-positive rates without continuous recalibration.
Download the full methodology white paper
46-page report covering adaptive calibration algorithm, benchmark comparison against static threshold systems, and validation protocol for process engineers.