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

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

Problem Statement
The paper addresses the critical challenge of screening latent defects in semiconductor manufacturing. Latent defects are physical imperfections (e.g., resistive shorts or opens) that allow devices to pass all functional test criteria but cause field failures under operating stress. Detecting these is difficult due to three compounding factors:

1. Extreme Class Imbalance: Failure rates are often in the parts-per-million range (e.g., <0.12% in the studied datasets), rendering supervised learning impractical due to the lack of labeled anomaly data and the tendency of classifiers to collapse to the majority class.
2. High Dimensionality: Modern test programs generate thousands of parametric measurement features per device, creating a high-dimensional feature space where traditional distance- or density-based methods struggle without careful feature engineering.
3. Strict Tolerance: Screening systems must balance maximizing the recall of rare anomalies against minimizing false positives to avoid yield loss, a balance that naive thresholding often fails to achieve.

Existing diffusion-based anomaly detection methods have primarily focused on image data or low-dimensional tabular data (e.g., TabDDPM), which treats feature vectors as flat, unordered inputs. These approaches lack the structural inductive bias required for high-dimensional, semantically heterogeneous IC test data.

Methodology
The authors propose Diffuse to Detect, a fully unsupervised framework that incorporates a Diffusion Transformer (DiT) to model the distribution of healthy silicon. The methodology proceeds through the following stages:

1. Dimensionality Reduction and Tokenization:

   • Raw high-dimensional measurement vectors ($x \in \mathbb{R}^F$, where $F \sim 10^3$) are first compressed by a two-layer MLP encoder into a compact latent representation ($h \in \mathbb{R}^{D_r}$, where $D_r = 128$).
   • This latent vector is reshaped into a structured 2D token sequence ($Z_0 \in \mathbb{R}^{C \times L}$), analogous to patch embedding in Vision Transformers. This allows the model to treat the measurement data as a sequence rather than a flat vector.

2. Two-Level Positional Encoding:

   • Feature-Level: Fixed sinusoidal embeddings are added to encode the position of tokens within the latent sequence, capturing the semantic structure of the test flow.
   • Die-Level: A learned, gated MLP generates a per-device embedding based on the physical wafer coordinates ($d_x, d_y$). This accounts for systematic spatial variations (e.g., edge vs. center effects) inherent in wafer-level test data.

3. Diffusion Process:

   • The framework employs a 1D Diffusion Transformer (DiT1D) as the denoising backbone. Unlike standard tabular diffusion models that use ResNet-based MLPs, DiT1D utilizes self-attention mechanisms to capture compound inter-feature correlations across the token sequence.
   • The model is trained exclusively on healthy devices using the standard DDPM objective to predict the noise added during the forward diffusion process.

4. Anomaly Scoring:

   • During inference, an anomaly score is derived from the noise-prediction error over a fixed set of mid-range diffusion timesteps.
   • Devices lying outside the learned distribution of healthy silicon induce elevated reconstruction errors.
   • The approach avoids iterative sampling for scoring, requiring only a fixed number of forward passes, making it suitable for high-throughput production screening.

Key Contributions

• First Diffusion Framework for IC Test Data: The paper introduces the first diffusion model-based framework for unsupervised anomaly detection on IC electrical test data, eliminating the need for manual feature engineering.
• Structured Tokenization and Positional Encoding: The authors propose a novel two-level positional encoding scheme that combines fixed sinusoidal token positions with learned, per-device die embeddings. This encodes both the test flow structure and the spatial geometry of the wafer.
• Latent-Space Diffusion with Transformers: The work demonstrates that performing diffusion in a compact latent space using a Transformer denoiser outperforms methods that diffuse in raw feature space or use MLP-based denoisers.

Experimental Results
The framework was evaluated on two industrial 16nm automotive IC test datasets characterized by extreme class imbalance (anomaly rates of 0.22% and 0.12%).

• Performance: The proposed DiT1D model achieved state-of-the-art performance, outperforming classical methods (e.g., Isolation Forest, PCA), deep learning baselines (e.g., GANomaly, DAGMM), and existing diffusion baselines (TabDDPM, DTE variants).
  • On Dataset 1, DiT1D achieved an AUROC of 0.771 and an AUCPR of 0.0250, significantly surpassing the second-best method (LOF and DTE-IG, respectively). It successfully recalled 3 out of 7 confirmed anomalies at a 95% yield threshold, whereas most baselines recalled zero or one.
  • On Dataset 2, DiT1D was the only method to recall all 7 anomalies present in the test split at the 95% yield operating point.
• Ablation Studies:
  • Removing the die-level positional embedding caused the largest drop in performance (AUROC decreased by 0.234), highlighting the importance of modeling spatial wafer variations.
  • Removing the autoencoder (diffusing in raw space) severely degraded performance, confirming the necessity of latent compression for high-dimensional data.
  • Replacing the Transformer with an MLP backbone (TabDDPM style) resulted in lower AUROC and recall, validating the benefit of self-attention for capturing inter-token correlations.

Significance and Claims
The paper claims that Diffuse to Detect offers a robust solution for the "Zero Defect" quality goals required in automotive and high-performance computing sectors. By leveraging generative diffusion models trained solely on healthy data, the approach:

• Eliminates the need for labeled anomalies, addressing the scarcity of defect data in mature processes.
• Avoids manual feature engineering, which has historically been a bottleneck in semiconductor quality pipelines.
• Provides interpretability: Through latent-space reconstruction residuals, the method enables failure localization. The authors note that these residuals can be mapped back to specific test programs, offering test engineers actionable insights into root causes (e.g., identifying specific leakage measurement deviations) without requiring knowledge of the latent representation.

The work positions itself as a significant step forward in applying advanced generative modeling to the specific constraints of high-dimensional, spatially correlated, and extremely imbalanced semiconductor test data.