automated defect classification
**Automated Defect Classification (ADC)** is the **application of machine learning and computer vision algorithms to automatically categorize SEM defect images into predefined defect types** — enabling fabs to process the thousands of defect images generated daily at production throughput without human review of each image, accelerating yield learning feedback cycles from days to hours.
**The Throughput Problem ADC Solves**
A modern 300 mm fab running patterned wafer inspection generates 10,000–100,000 flagged defect coordinates per day across all inspection layers. Review SEMs can image ~50 defects/hour. Even with multiple DR-SEM tools, manual classification of every defect image is impossible — creating a bottleneck where yield engineers are overwhelmed and feedback to process teams is delayed by days.
**ADC System Architecture**
**Image Collection**: DR-SEM captures high-resolution images of each defect at standardized conditions (magnification, beam energy, detector mode). Images are stored in a database linked to wafer, lot, layer, and coordinate metadata.
**Feature Extraction**: Classic ADC systems extract numerical features from each image — area, aspect ratio, perimeter, texture metrics (GLCM), Hu moments, Fourier descriptors. These ~50–200 features form a feature vector representing the defect's visual characteristics.
**Classification Engine**:
- **Rule-Based (Legacy)**: Hand-crafted decision trees — "if aspect ratio > 5 AND area > 1000 pixels, call Scratch." Fast but brittle; breaks when process conditions change.
- **SVM/Random Forest**: Statistical classifiers trained on labeled defect libraries. More robust than rules, but requires manual feature engineering.
- **Deep Learning (CNN)**: Convolutional neural networks trained end-to-end on thousands of labeled SEM images. Automatically learn relevant features from raw pixels. Current state-of-art achieves >95% classification accuracy on well-defined defect categories.
**Training Data Requirements**
CNN-based ADC requires large labeled training datasets — typically 500–2,000 images per defect class, reviewed and labeled by expert engineers. Dataset construction is the bottleneck: labels must be accurate, classes must be balanced, and training data must represent the full range of appearance variation for each defect type across different process conditions.
**Nuisance Filtering**
A critical ADC function is distinguishing real defects from "nuisances" — optical artifacts, pattern roughness events, and false triggers from the inspection tool. A poorly tuned nuisance filter floods engineers with false alarms; an over-aggressive filter misses real yield-critical defects. ADC systems maintain separate nuisance classifiers that precede the defect type classifier.
**Closed-Loop Yield Learning**: ADC output feeds directly into Yield Management Systems — defect type counts by layer are SPC-monitored, Pareto-ranked, and correlated to electrical test failures, creating the closed-loop feedback that drives process improvement.
**Automated Defect Classification** is **AI-powered quality inspection at factory scale** — replacing the impossible task of human review of millions of SEM images with machine learning classifiers that tirelessly categorize defects in real time, compressing yield learning cycles from weeks to days.