Wafer Defect Classification

# Wafer Defect Classification

## Introduction & Motivation

Detecting and classifying wafer defects through ML accelerates semiconductor quality control. ML models identify defect types from images and patterns for rapid wafer screening.

Motivation: Classify wafer defects for quality control.

Applications: Defect detection, classification, spatial analysis, yield prediction.

---

## Core Concepts & Theory

### Defect Types

Scratches, particles, voids.

### Spatial Distribution

Location patterns.

### Defect Size

Dimensions and impact.

### Classification Accuracy

Detection rate.

---

## Mathematical Formulation

Classification:
$$\hat{y} = \arg\max_c P(c|\mathbf{x})$$

Precision:
$$P = \frac{TP}{TP + FP}$$

Recall:
$$R = \frac{TP}{TP + FN}$$

---

## Advanced Theory & Extensions

### Convolutional Networks

Image processing.

### Attention Mechanisms

Region focus.

### Transfer Learning

Pre-trained models.

---

## Computational Considerations

Images: O(W·H·C) input.

CNN: O(D²) network.

Classification: O(D) per wafer.

---

## Practical Implementation Strategies

### Image Preprocessing

Normalization, augmentation.

### Feature Extraction

Pattern recognition.

### Classification Layers

Softmax output.

---

## Benchmark Datasets & Evaluation

WM-811K: Wafer defect data.

Semiconductor Datasets: Industry data.

Literature Defects: Published images.

---

## Key Challenges & Limitations

### Class Imbalance

Rare defects.

### Spatial Variation

Position effects.

Resolution Dependence

Image quality.

---

## Hyperparameter Tuning

Layers: 3-5 conv layers.

Filters: 32-512 per layer.

Learning rate: 1e-4 to 1e-2 schedule.

---

## Real-World Applications & Case Studies

Si Wafers: Logic chips.

GaAs Wafers: RF devices.

Si Carbide: Power electronics.

---

## Integration with Other Methods

Defect ML + imaging; + SEM; + yield models.

---

## Summary & Key Takeaways

ML classifies wafer defects efficiently.

Principles:
1. Imaging: Data acquisition.
2. Preprocessing: Image preparation.
3. Features: Pattern extraction.
4. Classification: Defect typing.
5. Analysis: Spatial patterns.

---

## Appendix: Practical Labs

### Lab 1: Wafer Image Preprocessing

import numpy as np

def preprocess_wafer_image(image, target_size=256):
 """Preprocess wafer defect image"""
 resized = image[:target_size, :target_size]
 normalized = (resized - np.mean(resized)) / (np.std(resized) + 1e-6)
 
 return normalized

image = np.random.randn(512, 512)
processed = preprocess_wafer_image(image)

assert processed.shape == (256, 256), "Preprocessing failed"
print(f"✓ Preprocessed image shape: {processed.shape}")

### Lab 2: Defect Classification

import numpy as np

class DefectClassifier:
 def __init__(self, num_classes=5):
 self.weights = np.random.randn(128, num_classes) * 0.1
 self.bias = np.zeros(num_classes)
 
 def classify(self, features):
 """Classify defect type"""
 logits = features @ self.weights + self.bias
 probs = np.exp(logits) / np.sum(np.exp(logits))
 return np.argmax(probs)

features = np.random.randn(128)
classifier = DefectClassifier()
defect_class = classifier.classify(features)

assert 0 <= defect_class < 5, "Classification failed"
print(f"✓ Predicted defect class: {defect_class}")

### Lab 3: Confusion Matrix

import numpy as np

def compute_confusion_matrix(predictions, labels, num_classes):
 """Compute confusion matrix"""
 cm = np.zeros((num_classes, num_classes))
 
 for pred, true in zip(predictions, labels):
 cm[true, pred] += 1
 
 return cm

preds = np.array([0, 1, 2, 1, 0])
labels = np.array([0, 1, 1, 1, 0])

cm = compute_confusion_matrix(preds, labels, 3)

assert cm.shape == (3, 3), "Confusion matrix failed"
print(f"✓ Confusion matrix shape: {cm.shape}")

### Lab 4: Detection Metrics

import numpy as np

class DefectMetrics:
 def __init__(self):
 pass
 
 def compute_precision_recall(self, tp, fp, fn):
 """Compute precision and recall"""
 precision = tp / (tp + fp + 1e-6)
 recall = tp / (tp + fn + 1e-6)
 f1 = 2 * precision * recall / (precision + recall + 1e-6)
 
 return precision, recall, f1

tp = 90
fp = 10
fn = 20

metrics = DefectMetrics()
p, r, f1 = metrics.compute_precision_recall(tp, fp, fn)

assert 0 <= p <= 1 and 0 <= r <= 1, "Metrics failed"
print(f"✓ Precision: {p:.2%}, Recall: {r:.2%}, F1: {f1:.2%}")

---