Enhancing Wafer Map Inspection Process in Semiconductor Manufacturing Using Deep Learning

1
Enhancing Wafer Map Inspection Process in
Semiconductor Manufacturing Using Deep
Learning https//yieldwerx.com/
2
In the highly competitive semiconductor
manufacturing industry, wafer map inspection is a
crucial step in ensuring product quality and
improving yield. This process, which involves
identifying defects in silicon wafers,
traditionally depends on manual, labor-intensive
techniques. However, with the rise of machine
learning and artificial intelligence, deep
learning methods have emerged as promising
alternatives to these traditional approaches. In
particular, Deep Convolutional Neural Networks
(DCNN) have shown great potential in improving
the efficiency and accuracy of wafer map
inspections. Nevertheless, DCNN's performance
heavily depends on the availability of
high-quality, balanced datasets for training,
which can be challenging to acquire in the
real-world semiconductor manufacturing
environment. To mitigate this issue, this study
proposes a new method that can effectively
utilize imbalanced datasets for training the
DCNN. Dataset Imbalance in Semiconductor
Manufacturing   Dataset imbalance in the context
of wafer map inspections comprises feature and
quantity distribution imbalances. Feature
distribution imbalance refers to the disparity in
the frequency of occurrence of different types of
defects, while quantity distribution imbalance
relates to the difference in the number of
defect-free and defective wafer samples. Feature
Distribution Imbalance and Convolutional Block
Attention Module (CBAM)   To address feature
distribution imbalance, the study incorporates an
enhanced Convolutional Block Attention Module
(CBAM) into the DCNN. The CBAM, which is designed
to improve the representation of wafer map
features in the DCNN, does this by highlighting
relevant features and suppressing irrelevant
ones. By doing so, the DCNN becomes more capable
of identifying defects, even if they appear
infrequently in the dataset. Alongside CBAM, a
novel feature-map-specific direction wafer
mapping software module is also introduced. This
module amplifies the positional information of
defect clusters, which is vital in identifying
the nature of the defects and their likely
causes. By enhancing the representation of defect
positions in the DCNN, the module further
improves the accuracy of defect detection.
3
Addressing Quantity Distribution Imbalance with
Cosine Normalization   On the other hand, the
study proposes the use of a cosine normalization
algorithm to tackle quantity distribution
imbalance. The cosine normalization algorithm is
used to replace the fully connected layer in the
DCNN, which is typically sensitive to the
quantitative distribution of samples in the
dataset. By incorporating the cosine
normalization algorithm, the DCNN becomes less
sensitive to such distributions, thereby
improving its performance even when trained on
imbalanced datasets. Classifier fine-tuning is
achieved through iterative training, which
gradually reduces the DCNN's sensitivity to the
quantitative distribution of samples in the
dataset. By doing so, the DCNN becomes more
robust and less prone to overfitting, thereby
improving its generalizability and accuracy on
unseen data. Experimental Results and Future
Work   The proposed method shows promising
results in improving the robustness of wafer map
inspection processes. Experimental results
indicate that the method performs better than
existing algorithms when applied to imbalanced
datasets. This means that the method is capable
of utilizing these challenging datasets
effectively, which is a significant advantage in
real-world applications where acquiring balanced
datasets can be difficult. Nevertheless, the
recognition of certain defect patterns, such as
Local patterns, remained a challenge with this
methodology. This issue indicates a potential
area for future work, which aims to focus on
improving the representation of defect clusters
in the DCNN to further increase recognition
accuracy.
4

• Conclusion and Impact on Semiconductor
  Manufacturing
• By improving the efficiency and accuracy of wafer
  map inspections, the proposed method has the
  potential to significantly enhance yield in
  semiconductor manufacturing. A more accurate
  inspection process means fewer false positives
  and negatives, leading to more reliable detection
  of defects. This can ultimately lead to a
  reduction in yield loss in manufacturing and a
  corresponding increase in product yield.
• Furthermore, the proposed method can potentially
  streamline the process of wafer map inspections,
  reducing the reliance on labor-intensive manual
  techniques. This can lead to significant cost
  savings and a reduction in the time taken to
  identify and address defects, thereby improving
  the overall efficiency of semiconductor
  manufacturing operations.
• Lastly, the proposed method's ability to utilize
  imbalanced datasets effectively can be a
  game-changer in the industry. This is because the
  availability of balanced, high-quality datasets
  can be a significant challenge in semiconductor
  manufacturing. By being capable of using
  imbalanced datasets, the proposed method can be
  applied more widely and effectively, leading to
  better results in practice.
• References
•  
• Hu, J., Shen, L., Sun, G. (2018).
  Squeeze-and-excitation networks. In Proceedings
  of the IEEE conference on computer vision and
  pattern recognition (pp. 7132-7141).
• Kang, B., Lee, J., Kweon, I. S. (2018,
  September). Self-attention with relative position
  embeddings. In Asian Conference on Computer
  Vision (pp. 323-339). Springer, Cham.
• Kingma, D. P., Ba, J. (2014). Adam A method
  for stochastic optimization. arXiv preprint
  arXiv1412.6980.
• Szegedy, C., Liu, W., Jia, Y., Sermanet, P.,
  Reed, S., Anguelov, D., ... Rabinovich, A.
  (2015). Going deeper with convolutions. In
  Proceedings of the IEEE conference on computer
  vision and pattern recognition (pp. 1-9).