Production Systems

Automatic Defect Classification (ADC)

ADC systems automatically classify defects in real-time as part of the production inspection flow:

• Inline ADC: Classification happens on the inspection/review tool itself, immediately after image capture. Low latency but limited compute.
• Offline ADC: Images are transferred to a server for classification by more sophisticated models. Higher accuracy but adds delay.
• Hybrid: Fast pre-classification inline, with uncertain cases sent to a more powerful offline system.

Production ADC requirements:

• Speed: Classify 1,000+ defects per wafer in seconds
• Accuracy: >95% agreement with expert human classification
• Purity: Critical defect categories (e.g., "killer defect") must have very high precision — false negatives are costly
• Adaptability: Models must handle new defect types as processes change

Inline ADC Integration with KLA / AMAT Review Tools

The path from "a CNN that classifies SEM images" to "an inline ADC engine serving 500 wafers/day" runs through three integration layers.

1. Tool ↔ classifier interface

• KLA review tools (eDR-7100, eDR-7600) and AMAT SEMVision expose a SECS/GEM data link that streams each defect image plus metadata (coordinates, tool, recipe, lot, wafer)
• The classifier runs on a tool-side GPU appliance with a hard latency budget — typically ≤100 ms per image to keep up with the SEM acquisition rate
• Classifier returns: class label, confidence, top-2 alternatives, model version hash

2. Confidence-tiered routing

def route_defect(prob_vec, threshold_high=0.92, threshold_low=0.55):
    """Return (label, action) given the per-class probability vector."""
    top_class = int(prob_vec.argmax())
    top_prob = float(prob_vec[top_class])

    if top_prob >= threshold_high:
        return top_class, "auto_classify"        # high-confidence: keep label
    if top_prob < threshold_low:
        return None, "manual_review"             # low confidence: human review
    return top_class, "queue_for_active_learning" # gray zone: label and retrain

3. Active-learning loop

1. Gray-zone images accumulate in a labeling queue
2. Domain expert labels ~50 / week through a lightweight web UI
3. Nightly job appends new labels to the training set and triggers fine-tuning
4. New model is canary-deployed on 5% of traffic; shadow accuracy is compared to the live model for a week before full cutover