AI-based explainable wafer scoring for semiconductor quality management
- Title
- AI-based explainable wafer scoring for semiconductor quality management
- Authors
- 김희진
- Date Issued
- 2022
- Publisher
- 포항공과대학교
- Abstract
- Semiconductor quality links to corporate reliability and competitiveness
directly. Therefore, detecting defects in advance and taking precautions before
shipping is important for the company’s reliability and competitiveness. The
semiconductor process consists of four steps: FAB process, probe test, assembly
process, and package test. The probe test is the first stage to generate data for each
die during the semiconductor process, so it plays an important role in predicting the
final quality. Various studies have been conducted to predict and improve
semiconductor yield by modeling based on the probe test data. However, the causes
of quality and yield are not analyzed mainly through the prediction model. The
traditional cause analysis of yield includes electrical tests and physical or structural
analysis methods. In the case of conventional analysis methods, it takes a lot of
money and time due to human resources and equipment limitations. Accordingly, if it
is possible to explain the relationship between variable and prediction result from the
prediction model, efficient cause analysis of yield will be possible.
This study analyzes the probe and the package test data. Then, it develops a
probe test defect prediction model using tree-based XGBoost. XGBoost is the latest
machine learning model. While the previous studies focused solely on prediction
performance, this study considers an explainable model. Therefore, it can explain
prediction performance and results using the partial dependence function. After model
development, wafer quality is defined based on the prediction probability of dies.
That proposes a wafer scoring method for improving semiconductor yield. Also, the
variables of the prediction model that affect wafer quality would be derived and
analyzed. It is expected that it will contribute to interpreting wafer quality by
deriving significant probe test variables.
Package yield is an important indicator for the final quality of
semiconductor finished products. Package yield improves after disposing of low-quality
wafers based on the proposed wafer scoring method. The analysis method based on
the explainable prediction model can significantly reduce the time required to analyze
the cause of package yield using wafer scores. If this methodology is applied to the
field, it will reduce the costs incurred in the semiconductor process, such as deriving
an optimal probe test process that satisfies the target package quality.
- URI
- http://postech.dcollection.net/common/orgView/200000601076
https://oasis.postech.ac.kr/handle/2014.oak/112182
- Article Type
- Thesis
- Files in This Item:
- There are no files associated with this item.
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.