Machine learning approach for wafer frontend test yield loss prediction

• Main Author: Zhong, Qinhong
• Other Authors: Wang Hong
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Electrical and electronic engineering::Microelectronics
• Online Access: https://hdl.handle.net/10356/166864
• Institution: Nanyang Technological University

Nowadays, as automation and digitalization are deeply integrated into the semiconductor industry, a tremendous amount of data generated from IC Design to Final Test plays a vital role in boosting innovation and productivity. The interval time between fabrication and testing could be a few weeks or even months due to wafer transportation. If any misbehavior occurred during fabrication, faulty instruction would be applied to wafer manufacturing continuously until yield loss events are tested and confirmed by fabless companies, causing great loss. Process Control Monitoring (PCM) data provided by foundries is an important accessible data source for fabless companies to detect and intervene wafer yield loss events before testing. However, currently used statistical monitoring methods on PCM data are no longer effective and reliable. After continuous optimization of fabrication process, monitored data of abnormal wafers becomes indistinguishable statistically. The primary objective of this project is to figure out if it is possible to use Machine Learning to predict wafer test yield loss categorically based on PCM data. To eliminate test-related impact, wafer frontend test is selected, because it would not be influenced by defects from slicing, packaging and assembly compared with wafer backend test. In this project, explorative data analysis (Eda) is conducted first. The analysis results of Eda show the limitation of sensitive PCM feature pairs and high potential difficulty in yield loss numerical prediction. After feature re-extraction and ML model training, the introduction of feature selection and train set imbalanced class processing is proven to improve specific ML models’ performance, but a trade-off should be considered. Model performances are interpreted in detail from different perspectives, giving in-depth explanations of prediction principle. By grouping and combining different classifiers, stacking ensemble also contributes to further performance improvement. The machine learning approach for wafer frontend test yield loss event classification gives very promising prediction results on test set compared with the baseline, showing huge potential for further study. Hence, analytic reasoning about possible obstacles to performance improvement and suggestions for future work are also given in the end.