Rotation-Invariant Wafer Map Pattern Classification With Convolutional Neural Networks

Seokho KangIEEE Access 2020 · OpenReview

Abstract

The enhancement of production yield is a continuous challenge in semiconductor manufacturing. Analyzing the spatial defect patterns of previously processed wafers is a key step in identifying the root causes of yield degradation. Predictive modeling approaches have been successful in automated wafer map pattern classification. The classification performance depends significantly on the quantity and diversity of data that can be acquired, which are often limited in practice. In this study, we demonstrate that rotation-based data augmentation can effectively improve wafer map pattern classification when training data are scarce. As rotation is a label-preserving transformation for wafer maps, we construct a convolutional neural network with rotation-augmented training data to render the classification invariant with respect to rotation. This enables us to provide consistent predictions for rotational variations of new wafer maps, thereby achieving higher classification performance. The effectiveness of our method is verified based on real-word data from a semiconductor manufacturer.

Introduction

In semiconductor fabrication, wafer maps serve as critical diagnostic tools, visualizing the pass/fail status of dies across a circular silicon wafer. Specific spatial arrangements of defects—such as rings, scratches, or clusters—often act as "fingerprints" for equipment malfunctions. Identifying these patterns promptly is essential for maintaining high production yields. While manual inspection by domain experts is accurate, it is unscalable for modern high-volume manufacturing. Automated solutions using Convolutional Neural Networks (CNNs) have become the standard, yet they face two significant challenges: data scarcity for rare defect types and the inherent rotational variance of standard CNN architectures.

Methodology: Achieving Rotational Invariance

To address these challenges, this research proposes a rotation-based data augmentation strategy. Because wafers are circular, rotating a wafer map is a "label-preserving transformation"—a "Scratch" remains a "Scratch" regardless of its orientation.

Augmentation Strategy

The researchers implemented a systematic augmentation process where each original wafer map was rotated at $k-1$ equally spaced counter-clockwise angles. Additionally, horizontal flips were applied to these rotations, resulting in a dataset expansion factor of $\lambda = 2k$. This forces the model to learn features that are independent of orientation, effectively baking rotational invariance into the network.

Model Architecture

The study utilized a modified LeNet-5 CNN architecture. Given the relatively low resolution of the input wafer maps ($32 \times 32$), this smaller, more efficient architecture was preferred over deeper models like ResNet. The network consists of three convolutional layers followed by three fully-connected layers, optimized for high-speed inference.

Datasets

The methodology was validated using the WM-811K dataset, a comprehensive real-world benchmark containing 811,457 wafer maps.

Example of a 'Near-Full' defect pattern from the WM-811K dataset.

Of the total dataset, approximately 20% (172,950 maps) are labeled across nine categories: Center, Donut, Edge-Loc, Edge-Ring, Loc, Random, Scratch, Near-Full, and Non-Pattern. The dataset is characterized by extreme class imbalance; for instance, "Non-Pattern" wafers account for over 85% of the data, while rare defects like "Near-Full" (shown above) represent less than 0.1%.

Results and Discussion

The impact of rotation-based augmentation was most pronounced in scenarios with limited training data.

  • Performance Gains: On a small training set of 100 samples, the Macro F-score (which treats all classes with equal importance) improved from 34.06 to 56.28 when using an augmentation factor of $\lambda=64$.
  • Consistency Metric: The researchers introduced a metric $\kappa$ to measure "average agreement" across rotations. Models trained with the proposed augmentation achieved $\kappa$ values exceeding 0.9, indicating that the model provides the same classification for a pattern regardless of its orientation over 90% of the time.
  • Optimization: Performance gains were found to saturate at an augmentation factor between 16 and 32, suggesting a practical balance between accuracy and training time.

Conclusion

This study demonstrates that rotation-based data augmentation is a powerful tool for improving the robustness and accuracy of wafer map classification, particularly when labeled data is scarce. By ensuring the model is invariant to rotation, manufacturers can achieve more consistent and reliable defect identification, leading to faster root-cause analysis and improved production yields. Future work may explore selective augmentation strategies to further optimize training efficiency for rare defect classes.

Summary updated

Abstract

The enhancement of production yield is a continuous challenge in semiconductor manufacturing. Analyzing the spatial defect patterns of previously processed wafers is a key step in identifying the root causes of yield degradation. Predictive modeling approaches have been successful in automated wafer map pattern classification. The classification performance depends significantly on the quantity and diversity of data that can be acquired, which are often limited in practice. In this study, we demonstrate that rotation-based data augmentation can effectively improve wafer map pattern classification when training data are scarce. As rotation is a label-preserving transformation for wafer maps, we construct a convolutional neural network with rotation-augmented training data to render the classification invariant with respect to rotation. This enables us to provide consistent predictions for rotational variations of new wafer maps, thereby achieving higher classification performance. The effectiveness of our method is verified based on real-word data from a semiconductor manufacturer.
BibTeX
@inproceedings{kang2020rotation,
  author = {Seokho Kang},
  title = {{Rotation-Invariant Wafer Map Pattern Classification With Convolutional Neural Networks}},
  booktitle = {IEEE Access},
  year = {2020},
  url = {https://openreview.net/forum?id=QeQm-Tc824P},
}
Figure