Wafer Bin Map (WBM) is a map that visualizes and shows the defective quality/defect status and defective type information of individual semiconductor chips obtained through electrical test results in the semiconductor manufacturing process. As modern semiconductor manufacturing processes become increasingly miniaturized at the nanoscale, accurately detecting defect patterns appearing in the WBM and rapidly identifying their causes are critical challenges for improving semiconductor yield and quality control. While deep learning technology has enabled attempts at automating defect classification, existing supervised learning-based methodologies have limitations in that they only function for predefined defect types. This leads to problems where new types of “unknown defect patterns” arising from product diversification or process miniaturization are either undetected or misclassified under existing definitions. Furthermore, training the model to recognize new patterns incurs significant inefficiency due to the substantial costs of data labeling and model retraining time.

To address this, this study developed an integrated defect detection framework based on active learning that maintains high classification performance for known defect patterns while effectively identifying unknown defect patterns and continuously learning.

The developed system consists primarily of two stages: unknown defect detection and classification/learning. First, an anomaly detector based on One-Class Support Vector Machine (SVM) preliminarily determines whether the input WBM is a previously learned known defect pattern or a new type of unknown pattern. If identified as an existing pattern, the classification model precisely classifies the specific defect type. (See Figure 1)

Conversely, data classified as unknown patterns are clustered into groups with similar characteristics using the DBSCAN algorithm. This clustered data enables efficient labeling with minimal intervention from process engineers. Through active learning techniques, the classifier updates new pattern information in real-time. This process allows the model to adapt to the constantly changing process environment, maintaining and improving its performance autonomously. (See Figure 2)

Experimental results using the WM-811K dataset demonstrated that the developed model maintained high classification accuracy for known defects while effectively filtering out unknown patterns. Furthermore, ‘Eye Defect Patterns,’ which is not present in WM-811K but present in the actual mass production line in Samsung Electronics, was successfully detected and learned these unknown patterns, proving the model's applicability and utility in real industrial settings.

This research is significant in that it presents a methodology demonstrating how artificial intelligence models can practically contribute to building intelligent defect management systems for semiconductor processes. The research findings were published in Expert Systems with Applications, which is famous in industrial engineering discipline.

Figure 1. Multi-step detection process for unknown defect patterns

Figure 2. Process of updating the classifier to an unknown pattern

※ Title: A framework for detecting unknown defect patterns on wafer bin maps using active learning

※ Journal: Expert Systems with Applications
※ DOI: https://www.sciencedirect.com/science/article/pii/S0957417424022450
※ Pure: https://pure.skku.edu/en/persons/donghee-lee/