Using Automatic Defect Classification To Reduce The Escape Rate Of Defects

Automated optical inspection (AOI) is a cornerstone in semiconductor manufacturing, assembly and testing facilities, and as such, it plays a crucial role in yield management and process control. Traditionally, AOI generates millions of defect images, all of which are manually reviewed by operators. This process is not only time-consuming but error prone due to human involvement and fatigue, which can negatively impact the quality and reliability of the review.

In the Industry 4.0 era, the integration of a deep learning-based automatic defect classification (ADC) software solution marks a significant advancement in manufacturing automation. For one, ADC solutions reduce manual workload – meaning less chance of human error and higher accuracy – and two, they are poised to lower the costs associated with high-volume manufacturing (HVM).

Deep learning, a branch of machine learning based on artificial neural networks, is at the core of these ADC solutions. It mimics the human brain’s ability to learn and make decisions; this enables the system to recognize complex patterns in data without explicit programming. Compared to traditional methods, this approach offers a significant leap in processing efficiency and accuracy.

In the last two decades, the surge in processing power and data availability has made deep learning and machine learning algorithms increasingly practical and effective. Deep learning also has shown exceptional capabilities in finding, through self-learning, hidden patterns within large data sets, making it a suitable fit for ADC applications.

An ADC solution can be integrated with AOI tools to develop a robust ADC library, a library which can use a combination of machine learning algorithms, such as K-nearest neighbors (KNN) and convolutional neural network (CNN) models, to improve accuracy but reduce the escape rate of defects. ADC supports both inline, or tool centric ADC, and offline applications, making it adaptable for various AOI machines across the HVM spectrum.

The primary function of ADC is to automatically classify defect codes. This is crucial for quality control in semiconductor manufacturing. By automating the classification process, ADC systems effectively streamline the inspection process, addressing common defects like scratches and pad defects with higher precision and reliability.

For our study, we employed an ADC approach using two sophisticated machine learning algorithms to ensure precise and efficient defect classification:

K-Nearest Neighbors (KNN): This supervised learning algorithm excels in pattern recognition tasks such as handwriting and image recognition. KNN operates by identifying the closest data points (a.k.a., neighbors) in a feature space to classify unknown inputs with high precision and accuracy. The KNN algorithm evaluates the similarity based on the proximity of data points, refining the classification process (table 1).

Table 1: The performance matrix result of KNN model.

Convolutional Neural Networks (CNN): A cornerstone of deep learning techniques, CNN is instrumental in image processing tasks, including classification, segmentation and object detection. It utilizes multiple convolutional layers to automatically extract and learn features from images. This process involves the analysis of each image to discern intricate patterns; this is crucial for accurate defect identification in semiconductor manufacturing (table 2).

Table 2: The performance matrix result of CNN only.

By combining the robust feature analysis of KNN with the advanced pattern recognition capabilities of CNN, the ADC method employed in our study was able to achieve significant accuracy in classifying a wide range of defect types. This dual approach not only enhances the overall classification process but also adapts to the intricate and varied nature of defects encountered in semiconductor assembly.

Table 3: The performance matrix result of KNN+CNN.

Our study determined that a multi-engine ADC process demonstrates a breakthrough in defect classification performance, leveraging advanced AI engines like CNN and KNN. This approach not only meets but often exceeds industry standards, with significant achievements including:

• Attaining an accept/pass/classify rate of approximately 99%.
• Reducing the defect escape rate to about 0.2%.
• Lowering the overkill rate to around 0.2%.

The integration of different models, particularly KNN and CNN, with our study’s inline ADC system provides a robust solution for AOI tools. This integration ensures high accuracy in classifying defects and reduces the risk of misclassification. Meanwhile, an offline ADC variant is able to focus solely on deep learning for defect image analysis.

Our findings show that while CNN-based models alone offer substantial accuracy, the incorporation of KNN significantly minimizes the likelihood of defect misclassification. This dual-model strategy, particularly effective in inline ADC applications, ensures reliable and precise defect classification across various manufacturing scenarios.