The field of semiconductor manufacturing is an intricate and intensive process that involves numerous complex chemical and physical operations. The final yield of the process, which signifies the percentage of functional chips produced from a silicon wafer, is a primary measure of a fabrication plant's efficiency.
Design and Development of a Provenance Capture Platform for Data Science
Addressing the Challenge of Wafer Map Classification in Semiconductor Manufacturing.pptx
1. Addressing the Challenge of
Wafer Map Classification in
Semiconductor Manufacturing
https://yieldwerx.com/
2. The field of semiconductor manufacturing is an intricate and intensive process that involves numerous complex chemical and
physical operations. The final yield of the process, which signifies the percentage of functional chips produced from a silicon
wafer, is a primary measure of a fabrication plant's efficiency. However, this yield is often affected by a multitude of factors,
including equipment malfunctions, material impurities, and errors in process control.
Fault detection and classification play a crucial role in optimizing the yield of the process. These faults can be detected and
classified by analyzing wafer maps, which are essentially two-dimensional representations of each die on a silicon wafer. The
wafer map software display defect patterns that correspond to issues in manufacturing, with each defect having a unique
signature that provides clues about the source of the problem.
Machine Learning in Wafer Map Classification
Recently, there have been significant efforts to automate the classification of wafer maps using machine learning (ML),
specifically Convolutional Neural Networks (CNNs). CNNs are a type of deep learning model that excels at recognizing patterns in
images. They are trained on previously annotated wafer maps, and once trained, they can classify new maps into probable defect
categories in real time. This not only reduces costs but also minimizes the manual labor involved in classifying thousands of dies
per wafer manually with the help of die per wafer calculator.
However, a fully automated approach has a significant drawback. Achieving near-perfect accuracy in classification, a requirement
due to the high misclassification cost remains a challenge. Misclassification could potentially result in a high-value chip being
scrapped due to a false-positive detection or a defective chip being passed for packaging due to a false negative, leading to
significant financial losses.
3. A Semi-Automatic Method
To overcome the limitations of a fully automated approach, a semi-automatic method of wafer map pattern classification is proposed. This
approach selectively utilizes the CNN based on its predictive uncertainty on a given wafer map.
Predictive uncertainty refers to the confidence that the model has in its prediction. If the uncertainty is low, indicating that the CNN has a
high degree of confidence in its prediction, the map is classified using the CNN. Conversely, if the uncertainty is high, indicating that the CNN
is unsure about its prediction, the map is classified manually by a process engineer.
This method, designed to improve the accuracy of CNN in the inference phase, ensures near-perfect accuracy while maximizing CNN
coverage and minimizing engineering effort. Essentially, it combines the strengths of automated classification (speed and scalability) with
manual classification (precision and reliability).
Experimental Results
Experimental tests were conducted using the WM-811k dataset, a large-scale wafer map dataset widely used for training and evaluating
machine learning models in wafer map classification. This dataset contains more than 800,000 wafer maps, each labeled with one of 9 classes
representing different defect types.
The results demonstrated that the semi-automatic method achieved an impressive accuracy rate of over 99% with a CNN coverage of 93%,
meaning that the CNN was able to confidently classify 93% of the wafer maps, with the remaining 7% requiring manual classification by an
engineer.
This finding represents a significant improvement over previous efforts, showcasing the potential of this semi-automatic approach to deliver
high accuracy rates while reducing the workload of process engineers.
4. Detailed Overview of Wafer Map Classification Using CNNs
Convolutional Neural Networks (CNNs) are a type of deep learning model that is adept at image recognition tasks, making them particularly
suitable for wafer map pattern classification. A wafer map, being a visual representation of a wafer, can be considered as an image with
different pixel intensities corresponding to the state (functional or defective) of each die.
The CNN model architecture generally consists of an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an
output layer. The convolutional layers are responsible for feature extraction from the input wafer maps. These features are then down
sampled in the pooling layers to reduce computation and provide translation invariance. After several iterations of convolutions and pooling,
the output is flattened and passed through fully connected layers for final classification. The output layer employs a softmax function that
gives a probability distribution over the possible defect classes for a given wafer map.
The CNN is trained using a large dataset of annotated wafer maps, where the model learns to associate certain patterns in the wafer maps
with specific defect classes. The performance of the CNN is typically evaluated using a separate validation set and is fine-tuned to minimize
classification errors.
The Role of Predictive Uncertainty in the Semi-Automatic Approach
The concept of predictive uncertainty is central to the proposed semi-automatic approach. It refers to the confidence the CNN has in its
prediction for a given wafer map. A low predictive uncertainty means that the CNN has a high degree of confidence in its prediction, and
therefore, the map is classified using the CNN. On the other hand, a high predictive uncertainty means that the CNN is unsure about its
prediction, leading to the map being classified manually by a process yield engineer.
Predictive uncertainty can be estimated in multiple ways. One common approach is to use dropout at test time, a method known as Monte
Carlo Dropout. By dropping out random neurons during each forward pass at test time, we obtain a distribution of predictions from which we
can estimate the model's uncertainty.
The threshold that determines whether the predictive uncertainty is high or low is a hyperparameter that can be adjusted based on the
requirements and constraints of the semiconductor manufacturing process.
5. The Practical Implications and Benefits of the Semi-Automatic Approach
In the context of semiconductor manufacturing, the proposed semi-automatic approach offers several practical benefits. First, it combines the
strengths of CNNs (speed and scalability) with those of process engineers (precision and reliability), enabling high-throughput and accurate wafer
map classification.
Second, the semi-automatic method addresses the challenge of misclassification, which can lead to significant financial losses. By utilizing human
expertise when the CNN's predictive uncertainty is high, the method reduces the likelihood of misclassification and thus safeguards the economic
viability of the manufacturing process.
Third, the semi-automatic approach can help manage workforce resources more effectively. By automating the classification of wafer maps with
low predictive uncertainty, process engineers can focus their efforts on more challenging cases that require in-depth analysis, enhancing the
overall productivity of the manufacturing process.
Lastly, by providing an additional layer of scrutiny for wafer map classification, the semi-automatic approach helps improve the overall yield of the
manufacturing process. By detecting and categorizing defects more accurately, process engineers can better understand the root causes of these
defects and take corrective measures, thereby optimizing the yield of the process.
Conclusion
In conclusion, the semi-automatic wafer map pattern classification method offers an efficient solution for achieving near-perfect accuracy in
semiconductor manufacturing. This method successfully manages the high costs of misclassification associated with fully automated systems,
leveraging the strengths of both CNNs and process engineers. By harnessing the power of AI, we can ensure higher wafer yield, reduce wastage,
and streamline the semiconductor manufacturing process.
References
1. Berrar, D., & Grébici, D. (2021). Data Mining in Semiconductor Manufacturing. CRC Press.
2. Ngan, H. Y. T., Lau, H., & Mak, K. L. (2018). Wafer map defect pattern classification and image retrieval using Convolutional Neural
Network. IEEE Transactions on Semiconductor Manufacturing, 31(1), 95-104.
3. Suh, S. C., & Kim, I. S. (2020). A comprehensive review of semiconductor wafer map defect pattern recognition. Journal of Semiconductor
Technology and Science, 20(1), 1-19.