automatic defect classification (adc)
**Automatic Defect Classification (ADC)** uses **machine learning to categorize defects detected during wafer inspection** — replacing slow manual review with AI-powered classification that identifies defect types (particles, scratches, pattern defects) in seconds, accelerating yield learning and enabling real-time process control.
**What Is ADC?**
- **Definition**: AI-based automatic categorization of wafer defects.
- **Input**: SEM or optical images of detected defects.
- **Output**: Defect type classification with confidence score.
- **Speed**: 10-100× faster than manual review.
**Why ADC Matters**
- **Speed**: Classify thousands of defects in minutes vs days of manual work.
- **Consistency**: Eliminates human subjectivity and variability.
- **Scalability**: Handle increasing defect counts as nodes shrink.
- **Real-Time**: Enable immediate process adjustments.
- **Cost**: Reduce metrology engineer time by 80-90%.
**How ADC Works**
**1. Image Acquisition**: SEM or optical inspection captures defect images.
**2. Preprocessing**: Normalize, enhance contrast, remove noise.
**3. Feature Extraction**: CNN extracts visual features automatically.
**4. Classification**: ML model predicts defect type.
**5. Confidence Scoring**: Probability for each category.
**6. Human Review**: Low-confidence cases flagged for manual check.
**Defect Categories**
**Particles**: Foreign material contamination.
**Scratches**: Mechanical damage, linear features.
**Pattern Defects**: Lithography, etch, or CMP issues.
**Residues**: Incomplete cleaning, polymer buildup.
**Voids**: Missing material in films.
**Bridging**: Unwanted connections between features.
**Pits**: Surface depressions or holes.
**Stains**: Discoloration or chemical residues.
**ML Approaches**
**Convolutional Neural Networks (CNNs)**:
- **Architecture**: ResNet, EfficientNet, Vision Transformer.
- **Training**: Supervised learning on labeled defect images.
- **Accuracy**: 90-98% for common defect types.
**Transfer Learning**:
- **Method**: Pre-train on ImageNet, fine-tune on defect data.
- **Benefit**: High accuracy with limited labeled data (1000-5000 images).
**Few-Shot Learning**:
- **Method**: Learn new defect types from just 10-50 examples.
- **Benefit**: Quickly adapt to new processes or defect modes.
**Quick Implementation**
```python
# ADC with PyTorch
import torch
import torchvision.models as models
from PIL import Image
# Load pre-trained model
model = models.resnet50(pretrained=True)
model.fc = torch.nn.Linear(2048, num_defect_classes)
model.load_state_dict(torch.load('adc_model.pth'))
model.eval()
# Classify defect
def classify_defect(image_path):
image = Image.open(image_path)
image_tensor = transform(image).unsqueeze(0)
with torch.no_grad():
output = model(image_tensor)
probabilities = torch.softmax(output, dim=1)
predicted_class = torch.argmax(probabilities).item()
confidence = probabilities[0][predicted_class].item()
return {
'class': defect_classes[predicted_class],
'confidence': confidence,
'probabilities': probabilities[0].tolist()
}
# Process batch of defects
defects = load_defects_from_inspection()
for defect in defects:
result = classify_defect(defect.image_path)
defect.classification = result['class']
defect.confidence = result['confidence']
# Flag low-confidence for manual review
if result['confidence'] < 0.85:
defect.needs_manual_review = True
```
**Training Data Requirements**
- **Minimum**: 500-1000 images per defect class.
- **Ideal**: 5000-10000 images per class for production.
- **Balance**: Similar number of examples for each class.
- **Quality**: Clean labels, representative of production defects.
**Performance Metrics**
- **Accuracy**: Overall correct classification rate (target: >95%).
- **Precision**: True positives / predicted positives per class.
- **Recall**: True positives / actual positives per class.
- **F1-Score**: Harmonic mean of precision and recall.
- **Confusion Matrix**: Identify which classes are confused.
**Integration**
ADC integrates with:
- **Inspection Tools**: KLA, Applied Materials, Hitachi SEM.
- **Fab MES**: Real-time defect data to manufacturing systems.
- **Yield Management**: Link defect types to electrical failures.
- **Process Control**: Trigger alarms for abnormal defect patterns.
**Best Practices**
- **Start with Common Defects**: Train on high-volume defect types first.
- **Continuous Learning**: Retrain models as new defect modes appear.
- **Human-in-the-Loop**: Manual review of low-confidence predictions.
- **Monitor Drift**: Track classification accuracy over time.
- **Explainable AI**: Use attention maps to understand model decisions.
**Typical Performance**
- **Classification Speed**: 0.1-1 second per defect.
- **Accuracy**: 90-98% depending on defect complexity.
- **Throughput**: 1000-10000 defects per hour.
- **Manual Review Rate**: 5-15% flagged for human verification.
**Advanced Features**
- **Multi-Modal**: Combine SEM + optical + EDX data.
- **Hierarchical**: Coarse category → fine subcategory.
- **Anomaly Detection**: Flag novel defect types not in training.
- **Root Cause Linking**: Connect defect types to process steps.
ADC is **transforming semiconductor metrology** — enabling fabs to process massive defect datasets in real-time, accelerating yield learning cycles from weeks to hours and making data-driven process control a reality at advanced nodes.