Diffusion Transformer Detects IC Anomalies Without Labels
A team of researchers has introduced the inaugural unsupervised anomaly detection system for integrated circuits, utilizing a Diffusion Transformer. The initial test data is compressed through an autoencoder and transformed into token sequences that incorporate sinusoidal and wafer-position embeddings. Anomaly scores are derived from noise-prediction errors across mid-range diffusion timesteps, allowing for rapid screening of wafers without the need for labeled defects. This innovative approach demonstrates leading performance on industrial 16nm IC test data, even amidst significant class imbalance, while also providing interpretable failure localization through the analysis of latent-space reconstruction residuals.
Key facts
- First unsupervised anomaly detection framework using a Diffusion Transformer
- Raw test measurements compressed by an autoencoder
- Token sequences enriched with sinusoidal and per-device wafer-position embeddings
- Anomaly scores derived from noise-prediction error over mid-range diffusion timesteps
- No labeled defects or manual feature engineering required
- State-of-the-art performance on industrial 16nm IC test data
- Extreme class imbalance handled effectively
- Interpretable failure localization through latent-space reconstruction residuals
Entities
Institutions
- arXiv