DECOR: Deep Embedding Clustering with Orientation Robustness

In semiconductor manufacturing, early detection of wafer defects is critical for product yield optimization. However, raw wafer data from wafer quality tests are often complex, unlabeled, imbalanced and can contain multiple defects on a single wafer, making it crucial to design clustering methods that remain reliable under such imperfect data conditions. We introduce DECOR, a deep clustering with orientation robustness framework that groups complex defect patterns from wafer maps into consistent clusters. We evaluate our method on the open source MixedWM38 dataset, demonstrating its ability to discover clusters without manual tuning. DECOR explicitly accounts for orientation variations in wafer maps, ensuring that spatially similar defects are consistently clustered regardless of its rotation or alignment. Experiments indicate that our method outperforms existing clustering baseline methods, thus providing a reliable and scalable solution in automated visual inspection systems.

💡 Research Summary

The paper introduces DECOR (Deep Embedding Clustering with Orientation Robustness), a three‑stage unsupervised framework designed to cluster wafer‑map images in semiconductor manufacturing where data are unlabeled, highly imbalanced, and often contain multiple defect types per wafer. The authors identify three practical challenges: (1) traditional clustering methods require fixed hyper‑parameters such as the number of clusters (k) or distance thresholds, which are unrealistic in a production environment where defect distributions evolve; (2) wafer maps frequently appear in different rotations or flips due to handling, causing standard algorithms to treat identical patterns as distinct; (3) multi‑defect wafers make evaluation and outlier detection difficult.

Stage 1 – Rotation‑and‑Flip‑Invariant Embedding (RCAE).
The authors build a convolutional autoencoder whose encoder consists of three equivariant blocks based on R2Conv, a group convolution that is equivariant to the dihedral group D₄ (four rotations and two mirror flips). After each block, a GroupPooling operation collapses the orientation channels, yielding a rotation‑ and flip‑invariant 128‑dimensional latent vector. The decoder mirrors the encoder and is trained with a mean‑squared‑error loss. Compared with a vanilla CAE and a contrastive MoCo model, RCAE produces more compact, well‑separated embeddings while explicitly handling symmetry.

Stage 2 – Non‑Parametric Clustering (DeepDPM).
The latent vectors are fed into DeepDPM, a deep Dirichlet Process Mixture Model. DeepDPM implements a two‑layer MLP (128 → 50 → K) that outputs soft cluster membership probabilities. Because the Dirichlet Process concentration parameter is learned, the model automatically infers the appropriate number of clusters, eliminating the need for a pre‑specified k. Soft assignments are converted to hard labels by taking the argmax. This non‑parametric approach is robust to the unknown and potentially overlapping defect categories present in the MixedWM38 dataset.

Stage 3 – Ensemble Outlier Detection.
Within each inferred cluster, the authors run two complementary detectors: Isolation Forest (global partition‑based isolation) and Local Outlier Factor (local density deviation). Raw scores are transformed into binary decisions using a robust threshold τ = median(s) + k·MAD(s), where MAD is the median absolute deviation. The per‑cluster neighborhood size for LOF is set adaptively as k = clip(√N, k_min, k_max). An instance is flagged as an outlier only if both detectors agree, which empirically improves precision on rare defect patterns while preserving sensitivity.

Experimental Setup.
The framework is evaluated on the open‑source MixedWM38 dataset (≈38 k wafer maps, 38 distinct defect combinations, including single‑defect and mixed‑defect patterns). Images are resized to 128 × 128, normalized, edge‑masked, and lightly blurred (Gaussian kernel = 5, σ = 1.0). A multilabel‑stratified train‑test split preserves defect‑type proportions. Baselines include CAE, MoCo, and each of these combined with K‑Means (parametric) as well as the same embeddings paired with DeepDPM.

Results.
Table 1 shows that RCAE + DeepDPM achieves the highest clustering quality: NMI = 0.543 ± 0.03 and ARI = 0.296 ± 0.00, outperforming the best parametric baseline (CAE + K‑Means, NMI = 0.529, ARI = 0.205). Visualizations (t‑SNE/UMAP) confirm that clusters are compact and invariant to rotations; images of the same defect pattern appear together regardless of orientation. The ensemble outlier detector reduces false positives compared with using either Isolation Forest or LOF alone.

Computational Efficiency.
Training RCAE for 1,000 epochs takes ~6 hours on a single NVIDIA H100 GPU; DeepDPM training (max 200 epochs) adds ~2 hours. Total GPU consumption across all experiments is ≈30 GPU‑hours, and model checkpoints occupy <1 GB, indicating suitability for integration into real‑time inspection pipelines.

Limitations and Future Work.
The non‑parametric clustering still requires an initial guess for k_init and a suitable training schedule, which the authors treat as hyper‑parameters tuned empirically. Moreover, evaluating multi‑label clustering with NMI/ARI necessitates collapsing multi‑defect ground truth to a dominant label, potentially obscuring nuanced performance. Future directions include developing multi‑label‑aware clustering metrics, incorporating temporal modeling of soft cluster assignments to track defect evolution across production cycles, and benchmarking against additional parametric models such as Gaussian Mixture Models and deep clustering frameworks (e.g., Deep Embedded Clustering).

Conclusion.
DECOR successfully combines a rotation‑and‑flip‑invariant autoencoder, a Dirichlet‑process‑based non‑parametric clustering engine, and a robust ensemble outlier detector to address the three major pain points of wafer‑map analysis: lack of labels, multi‑defect complexity, and orientation variability. The framework delivers superior clustering performance with modest computational overhead, offering a scalable solution for automated visual inspection in semiconductor fabs and, by extension, for other domains where data exhibit similar imperfections.