Diffuse to Detect: Generative Diffusion Models for Unsupervised IC Anomaly Detection

Researchers propose an unsupervised anomaly detection framework using Diffusion Transformers to identify defects in semiconductor manufacturing at the 16nm node. The method combines autoencoders with diffusion models to screen for rare defects without labeled training data, achieving state-of-the-art results on industrial test data.

This research addresses a critical pain point in semiconductor manufacturing: detecting extremely rare defects in high-dimensional test data without labeled examples. Traditional supervised approaches fail when anomalies represent less than 0.1% of production, making labeled datasets impractical to generate. The proposed framework leverages recent advances in diffusion models, which excel at learning complex distributions from unlabeled data, paired with Diffusion Transformers for structured sequence processing.

The technical approach first compresses raw measurements through an autoencoder to reduce dimensionality, then constructs token sequences enriched with positional embeddings specific to wafer location and device characteristics. Anomaly detection operates by measuring prediction error in the diffusion model's noise estimation at mid-range timesteps, where signal-to-noise ratios favor subtle defect detection. This eliminates manual feature engineering—a historically labor-intensive step in semiconductor quality control.

For the semiconductor industry, this represents meaningful operational efficiency gains. Current defect screening relies on expensive electrical testing or human inspection, with false negatives resulting in field failures and costly recalls. Achieving detection without labeled data accelerates deployment and reduces dependency on historical defect catalogs. The interpretability through latent-space reconstruction residuals provides actionable feedback for process engineers to understand failure mechanisms.

Longer-term implications extend to scaling advanced node manufacturing (5nm, 3nm) where process variations create novel, previously-unseen defect modes. This framework's unsupervised nature makes it adaptable to emerging process nodes without retraining on new labeled datasets. Success at 16nm validates applicability to tighter tolerances, though computational requirements at production scale warrant continued optimization.

• →Diffusion Transformers enable effective anomaly detection in semiconductors without labeled defect data, solving a critical manufacturing challenge.
• →The approach eliminates manual feature engineering by learning directly from raw test measurements, reducing deployment time.
• →State-of-the-art performance on industrial 16nm data demonstrates practical applicability beyond academic benchmarks.
• →Interpretable failure localization through latent-space residuals provides actionable insights for process optimization.
• →Unsupervised framework scales to new process nodes without requiring expensive retraining on novel defect types.