Efficient Convolutional Neural Networks for Semiconductor Wafer Bin Map Classification

EunMi Shin, Chang D. Yoo

TL;DR

This paper tackles the challenge of automatically classifying defect patterns on semiconductor wafers, a crucial step for quality control. They propose using lightweight convolutional neural networks (CNNs) like MobileNetV3 to achieve high accuracy with significantly reduced computational resources, making advanced defect detection feasible even on less powerful hardware.

What Problem Does This Paper Address?

The semiconductor manufacturing process is complex and lengthy, with the wafer test process being critical for identifying defects and determining yield. Wafer bin maps visually represent these defects, but manual inspection is time-consuming and prone to human error. Automating this classification is desirable, but existing deep learning models often require substantial computational resources and high-performance hardware, making them impractical for many manufacturing environments. Furthermore, wafer map datasets often suffer from class imbalance, where "normal" wafers far outnumber specific defect types, posing a challenge for model training.

Methods

The researchers explored several state-of-the-art lightweight CNN architectures: EfficientNetV2, ShuffleNetV2, MobileNetV2, and MobileNetV3. The core idea is to leverage these models' efficiency to achieve good performance without the need for expensive hardware.

Key methodological aspects include:

  • Lightweight Architectures: The focus was on models known for their efficiency in terms of parameters, memory usage, and computational operations (MAdds, FLOPs).
  • Data Augmentation: To combat the class imbalance issue, various image transformation techniques (RandomRotation, RandomErasing, Resize, etc.) were applied to the training data to artificially increase the number of samples for underrepresented defect classes.
  • Normalization: Input images were normalized to ensure consistency across different models.
  • K-fold Cross-Validation: This technique was used to provide a more robust evaluation of the models' performance by systematically splitting the data into training, validation, and testing sets.

Datasets

The primary dataset used in this study was WM-811K. This dataset contains approximately 811,457 wafer images, with 172,950 labeled data points. The labeled data covers nine defect categories: Center, Donut, Edge-loc, Edge-ring, Loc, Near-full, Random, Scratch, and None (normal). The dataset exhibits a significant class imbalance, with the "None" class comprising over 85% of the labeled data.

Key Insights and Ideas

  • Lightweight is Right: The study strongly suggests that efficient, lightweight CNNs can achieve performance comparable to larger, more resource-intensive models for wafer map classification.
  • MobileNetV3 Shines: Among the tested lightweight models, MobileNetV3 demonstrated the best balance of classification accuracy and resource efficiency.
  • Data Augmentation is Crucial: The effectiveness of data augmentation in mitigating class imbalance and improving model robustness for defect detection is highlighted.
  • Beyond Accuracy: The paper emphasizes that accuracy alone is insufficient for evaluating models on imbalanced datasets. Metrics like the Macro F1 score, which gives equal weight to all classes, are more informative.

Key Results

  • Performance: MobileNetV3 achieved an accuracy of 98% and an F1 score of 89.5% on the testing dataset.
  • Resource Efficiency: Compared to ResNet, MobileNetV3 used 7.5 times fewer parameters.
  • Speed: MobileNetV3 offered a training speed that was 7.2 times faster and an inference speed that was 4.9 times faster than ResNet.
  • Comparative Performance: While all tested lightweight models performed slightly lower than ResNet18 on training and validation, MobileNetV3 matched ResNet18's F1 score on the testing dataset and outperformed CNN-WDI in training and validation F1 scores.
  • Image Size Trade-off: Reducing input image size significantly improved computation and throughput but led to a notable decrease in F1 score, indicating a trade-off between efficiency and classification performance.

Limitations and Future Work

The authors acknowledge two primary limitations:

  • New Defect Types: The current model struggles to recognize novel defect patterns not present in the training data.
  • Mixed Defect Patterns: The model has difficulty classifying images where multiple defect patterns are present simultaneously, often predicting only one.

Future work will focus on:

  • Unknown Defect Recognition: Developing a method to identify and classify new, unseen defect types by treating them as an "unknown" class during training.
  • Handling Mixed Defects: Expanding the dataset to include mixed defect patterns and training the model to recognize combinations of defects.

Summary updated

BibTeX
@inproceedings{shin2023efficient,
  author = {EunMi Shin and Chang D. Yoo},
  title = {{Efficient Convolutional Neural Networks for Semiconductor Wafer Bin Map Classification}},
  booktitle = {Sensors},
  year = {2023},
  url = {https://openreview.net/forum?id=OefwWbV0JL},
}
Figure