Defect classification systems convert inspection candidates into labels, confidence scores, Pareto distributions, and process dispositions. Detection, review, classification, and yield attribution are different operations. An optical inspector finds differences; a review SEM relocates and images them; a classifier assigns a bin; and downstream evidence determines yield consequence. Preserve raw signal, coordinate, image, prediction, correction, process context, and disposition instead of reducing a lot to one count.

Detection sensitivity and classification accuracy solve different problems. Brightfield inspection is often sensitive to pattern and reflectance changes, while darkfield emphasizes scattered light from particles, scratches, and topography. Electron-beam inspection or review can resolve smaller morphology at lower throughput. Every detector operates at a threshold: lowering it may capture more subtle defects but also creates more nuisance candidates. Recipe care areas, reference choice, pixel size, illumination, focus, die registration, and background subtraction shape the candidate population before classification begins. A classifier cannot recover a killer defect that inspection never detected, and high classifier accuracy cannot justify an inspection recipe with poor capture efficiency.

The taxonomy must represent morphology, mechanism, and uncertainty without confusing them. Particle, scratch, stain, residue, bridge, missing feature, void, crack, and crystal-originated pit are useful visual bins, but a visual appearance is not automatically a root cause. A particle can originate in deposition, clean, handling, or chamber wear; a bridge may be resist scum, etch residue, or deposited material; a stain may be an optical signature without electrical consequence. Keep an unknown or other class rather than forcing every image into a familiar bin. Record layer, process step, product, die region, inspection channel, review modality, and model version beside the label. Hierarchical labels can separate morphology from suspected mechanism so later XPS, SIMS, AFM, or cross-section evidence can refine causality without rewriting the original observation.

Sampling must preserve rare critical classes. If 25 wafers produce 20 candidates each, the lot has 500 detections; reviewing only the highest-signal 100 biases toward large defects. Stratify by wafer, zone, die region, signal, suspected class, and channel, then add random controls. Auditing 50 images across 8 classes gives 400 images before unknowns. Targeted oversampling can protect rare killers, but population estimates must restore sampling weights. Deduplicate clusters while retaining their size as process evidence.

Confidence thresholds require class-specific precision and recall, not one accuracy score. For a killer class, recall = TP/(TP+FN) measures the fraction of true killers found, while precision = TP/(TP+FP) measures how many killer predictions are correct. In an illustrative validation set with 100 true killer images, 90 true positives, 10 false negatives, and 30 false positives, recall is 90% and precision is 90/(90+30) = 75%. Overall accuracy can still appear high when nuisance examples dominate. A 99% accurate model evaluated on a population containing only 0.1% killers can miss most killers and remain numerically impressive. Report a confusion matrix by class, precision, recall, false-negative rate, support count, and confidence calibration. Choose thresholds from the asymmetric cost of missed killers and unnecessary review, not from a round 90% confidence value.

Nuisance filtering must suppress nonactionable signal without hiding excursions. A nuisance defect is not simply a small defect or one that does not fail the current electrical screen. Repeating design edges, harmless color variation, prior-level visibility, and systematic optical signatures may dominate raw count while contributing little incremental yield loss. Filters can use design context, location recurrence, multi-channel signal, size, shape, or validated class confidence. Their escape rate must be measured with challenge sets containing known killers. Store filtered candidates or a statistically meaningful sample so a recipe or model change can be reconstructed. If 400 of 500 detections are suppressed as nuisance, the dashboard should retain the 80% suppression rate, sampling policy, and audited false-negative result rather than showing only 100 surviving defects.

Killer attribution needs evidence beyond the image. Overlay defects with wafer sort, parametric test, memory bitmap, scan diagnosis, or final-test failure at a declared tolerance. Keysight or Keithley electrical results can connect a bridge, open, leakage site, or resistance shift to a location. Correlation alone is not causality; split lots, process-window experiments, cross-sections, and chamber or reticle recurrence strengthen inference. Estimate impact by class and location rather than assigning every severe-looking image 100% kill probability.

A Pareto becomes actionable only after counts are normalized and weighted by consequence. Raw counts answer how often bins appear in the sampled data. Density normalizes by inspected area such as defects per cm²; opportunity normalization may use die, feature, contact, or inspected care area. A yield-weighted score can multiply estimated population count by a validated kill probability, but both inputs carry uncertainty. In the SVG, nuisance has the largest raw bar while a lower-count bridge carries the largest killer contribution. That reversal is the purpose of a yield lens: engineering resources should follow preventable loss, not the tallest unqualified bar. Show unknowns and SEM non-visuals separately because removing them from the denominator can make classification performance and Pareto stability look better than they are.

Model governance is part of process control, not an offline data-science detail. Training data need immutable images, definitions, weights, product coverage, and adjudicated labels. Split datasets by wafer or lot so duplicates cannot leak across sets. Version preprocessing, crop, normalization, model, thresholds, and taxonomy. A model trained on 200 nm optical crops may not transfer to 20 nm SEM morphology. Confidence is not probability without calibration; a 0.95 score can drift with contrast, layer, product, or class mix. Monitor prevalence, confidence, unknowns, not-found sites, overrides, precision, and recall; retrain only through controlled evidence.

Metrology escalation should resolve a defined ambiguity. AFM measures profile; XPS identifies surface chemistry; SIMS supplies depth-resolved species; ellipsometry constrains modelable films; four-point probe and Hall effect connect broad electrical shifts; DLTS examines traps on suitable structures. Semilab platforms can add maps, while NIST practices inform calibration and uncertainty. Use the least destructive discriminating method and preserve its coordinate transform and uncertainty.

{ "rows": [
  { "type": "nodes", "items": [
    { "title": "Inspect wafer", "sub": "brightfield, darkfield, e-beam; save coordinates", "tone": "neutral" },
    { "title": "Select review sample", "sub": "stratify signal, wafer, zone, class, and random sites", "tone": "neutral" }
  ] },
  { "type": "arrow" },
  { "type": "group", "title": "Review and classification control loop", "note": "retain raw evidence and every override", "cycle": true, "loop": "audit confusion, drift, unknowns, and not-found rate", "items": [
    { "title": "ADR and SEM image", "sub": "re-detect, center, image, preserve source", "tone": "green" },
    { "title": "ADC prediction", "sub": "class, confidence, model and taxonomy version", "tone": "green" },
    { "title": "Threshold gate", "sub": "auto-bin qualified classes; review uncertain or risky", "tone": "orange" },
    { "title": "Human adjudication", "sub": "correct label, reason code, unknown retained", "tone": "orange" }
  ] },
  { "type": "arrow" },
  { "type": "nodes", "items": [
    { "title": "Build weighted Pareto", "sub": "population estimate, density, kill correlation, uncertainty", "tone": "green" },
    { "title": "Disposition and feedback", "sub": "contain excursion, find source, verify fix, monitor recurrence", "tone": "neutral" }
  ] }
] }

Release decisions must remain reproducible across tools, products, and time. Lock inspection, care areas, sampling, transforms, imaging, dictionary, thresholds, model, human-review policy, normalization, correlation window, and escalation. Challenge the chain with adjudicated defects at multiple signals. At 200 ms per site, 500 sites require 100 s imaging before motion, autofocus, re-detection, recovery, and transfer. Speed matters only while not-found rate, image quality, and killer recall remain controlled.

Read defect classification systems through a killer-defect-yield lens rather than a raw-count lens. Inspection establishes which candidates were detectable, ADR establishes which sites were reviewable, ADC proposes reproducible labels, human adjudication controls ambiguity, and electrical or process correlation establishes consequence. In the illustrative validation set, 90 of 100 true killers are found for 90% recall, yet 30 false alarms limit precision to 75%; both facts matter. Likewise, suppressing 400 of 500 candidates as nuisance may accelerate disposition, but only a retained audit proves that the 80% filter did not hide the defects that control yield. The defensible output is therefore not the tallest Pareto bar—it is a versioned chain from detection opportunity through classification uncertainty to verified, preventable yield loss.