Pre-trained CNN-based TransUNet Model for Mixed-Type Defects in
Wafer Maps
YOUNGJAE KIM1,2, JEE-HYONG LEE1*, JONGPIL JEONG1*
1Department of Computer Science and Engineering,
Sungkyunkwan University,
2066 Seobu-ro Jangan-gu, Suwon, 16419,
REPUBLIC OF KOREA
2Device Solutions Division,
Samsung Electronics,
1, Samsung-ro, Giheung-gu, Yongin-si, Gyeonggi-do 17113,
REPUBLIC OF KOREA
*Corresponding Authors
Abstract: - Classifying the patterns of defects in semiconductors is critical to finding the root cause of
production defects. Especially as the concentration density and design complexity of semiconductor wafers
increase, so do the size and severity of defects. The increased likelihood of mixed defects makes finding them
more complex than traditional wafer defect detection methods. Manually inspecting wafers for defects is costly,
creating a need for automated, artificial intelligence (AI)-based computer vision approaches. Previous research
on defect analysis has several limitations, including low accuracy. To analyze mixed-type defects, existing
research requires a separate model to be trained for each defect type, which is not scalable. In this paper, we
propose a model for segmenting mixed defects by applying a pre-trained CNN-based TransUNet using N-pair
contrastive loss. The proposed method allows you to extract an enhanced feature by repressing extraneous
features and concentrating attention on the defects you want to discover. We evaluated the model on the Mixed-
WM38 dataset with 38,015 images. The results of our experiments indicate that the suggested model performs
better than previous works with an accuracy of 0.995 and an F1-Score of 0.995.
Key-Words: - classification, mixed-type wafer maps, TransUNet, N-pair contrastive loss function, multitask
learning, transformer layer
Received: July 8, 2022. Revised: May 29, 2023. Accepted: June 19, 2023. Published: July 18, 2023.
1 Introduction
The process of making semiconductor wafers is
broadly categorized into front-end and back-end
processes and is made through eight different
processes, [1]. The front-end process is the process
of designing and etching semiconductor chips onto
wafers, while the back-end process is the process of
cutting the chips from the wafers, wrapping them in
insulation, and laying wires to reliably deliver
power, [2], [3]. In specific, the front-end process,
sometimes referred to as the wafer process, is the
process of repeating the formation and cutting of
different types of materials on the face of a wafer to
create electronic circuits to make a single
semiconductor chip, [4]. Previous processes include
photolithography, which prints patterns of
semiconductor circuits on wafers, etching to cut
away parts other than the circuit pattern, deposition
to create insulating thin films to separate and protect
the metal from the circuit for electrical signal
transmission, and metalization to create wiring, [5],
[6]. With such a wide variety of processes, there are
many different patterns of defects that can appear on
a wafer. After wafer manufacturing, we run several
tests to inspect defects and display them as wafer
maps of binary numbers. The result of classifying
the dies on the wafer map in this way shapes a
particular pattern and is visually represented, [7].
The different patterns of defects in the wafer map
are associated with the fabrication process. So,
exactly categorizing the pattern of defects in a wafer
map can help determine the source of defects in the
manufacturing process. These classifications are
important because they give engineers clues for
troubleshooting, [8].
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.27
Youngjae Kim,
Jee-Hyong Lee, Jongpil Jeong
E-ISSN: 2224-3402
238
Volume 20, 2023
A semiconductor wafer is a circular plate made
by growing a column of single crystals, such as
silicon (Si) or gallium arsenide (GaAs), which are
the core materials of semiconductor integrated
circuits. A semiconductor integrated circuit is an
electronic component that integrates many devices
on a single chip to perform and organize various
functions. This means that, since semiconductor
integrated circuits build their circuits on thin
circular plates or wafers, the wafer is the foundation
of the semiconductor.
Wafer mapping is commonly used for data
analysis of semiconductor manufacturing processes.
Wafer mapping generates a color-coded map of
semiconductor device performance on the surface of
a wafer based on the test results of each chip failure.
The created wafer maps have one or more patterns
depending on the distribution of the defect chips. In
semiconductor processes, different defect patterns
occur because the defect chip pattern will look
different depending on the source of the irregularity.
Therefore, wafer maps defect analysis of defective
chip patterns provides critical information for
finding anomalies in semiconductor processes and
determining the cause of defects. Fig. 1 shows a
wafer map.
Fig. 1: Example of wafer map.
Modern developments in micronization
techniques and increases in wafer size have
increased the likelihood of creating more than two
defect patterns, [9], [10]. Recent advances in Fig. 1.
Example of wafer map. AI technology has spurred
research focused on deep learning. The study, [11]
suggested using CNN to classify mixed defect
patterns. In [12], the authors suggested a deformable
CNN. In [13], the authors suggested an infinite
warped mixture model for clustering mixed-type
defects. In [14], the authors applied augmentation
techniques for the segmentation of mixed defect
patterns and proposed a masked R-CNN. In [15], the
authors proposed an Improved U-Net with a
Residual Attention Block for mixed-defect wafer
maps.
Taking classification one step further,
segmentation is used as a way to detect defects.
Classification simply categorizes the input image,
but segmentation can provide inferences about the
data on a pixel level. This can greatly help users
make decisions by providing them with additional
information. There is a lot of segmentation research
going on, especially in the medical field, to find
diseases. In [16], the authors applied a model that
applies attention gates on U-Net to perform medical
image segmentation. With attention gates, the model
is trained to maintain target structures of different
patterns and scales in focus automatically. A recent
study suggested a cascaded neutralization dual
attention U-Net for achieving enhanced tumor
segmentation, [17]. In [18], the authors suggested a
new U-Net architecture that uses aggregated
residual blocks and a soft attention mechanism to
segment COVID-19-infected areas. They proposed a
cascading structure to scale low-resolution quality
prediction and a dual-attention module to enhance
the feature representation of tumor segmentation. In
[19], the authors proposed TransUNet for medical
image segmentation
Mixed defects are harder to recognize because as
the defects are mixed, the pattern becomes more
complex, and different defects overlap. So, we use a
pre-trained CNN-based TransUNet to deliver
classification and segmentation results targeted at
defect areas to engineers. We specifically contribute
as follows:
1. N-pair contrastive loss-based pre-training to
generate good Feature maps that focus on
defect details you want to find.
2. Applying multi-task learning to TransUNet
for wafer defects.
3. Reduce unnecessary labor and time by
creating pseudo-label data required to train
segmentation models with automatic defect
masking techniques.
The structure of the paper is as follows. Section 2
describes the background and related work. Section
3 explains in detail the architecture and
characteristics of the proposed model. Section 4
explains the experimental setup and results. Finally,
Section 5 sets out our conclusions and suggestions
for future research.
2 Background & Related Work
2.1 TransUNet
Hybrid CNN-Transformer as Encoder. Instead of
taking a pure transformer for an encoder, TransUNet
uses a CNN transformer. It is a hybrid model that
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.27
Youngjae Kim,
Jee-Hyong Lee, Jongpil Jeong
E-ISSN: 2224-3402
239
Volume 20, 2023
initially uses a Convolutional Neural Network as a
feature extractor to build a feature map for the input.
Patch embedding is performed on a 1×1 patch
extracted from the Convolutional Neural Network
feature map instead of the raw image. Cascaded
Upsampler (CUP). The decoder consists of several
upsampling stages that decode hidden features for
the resulting segmentation mask output. Together
with the hybrid encoder, the CUP forms a U-shaped
architecture, with skip-connections that enable
feature aggregation at different resolution levels.
Fig. 2 shows the architecture of TransUNet
Fig. 2: Overall architecture of TransUNet.
2.2 Related Works on Defect Classification
In [11], the authors suggested using CNN to classify
mixed defect patterns. They train distinct
classification models to classify single defects and
then use all models to find the occurrence of mixed-
type defects. However, this method is not efficient,
as it requires a significant increase in both the
storage and computing power overhead, making it
less scalable. In addition, separate training is
required for every model. They also only study four
single defect patterns: ‘zone’, ‘circle’, ‘scratch’, and
‘ring’. In contrast, we test our model on 8 single-
type and 13 mixed-type defect patterns. For defect
classification, the study, [13], proposed a
deformable convolutional network (DCNet). For
extracting a feature representation of a defect,
DCNet uses transform convolution (DC) to focus on
a sampling region of the defective dies. The output
layer has multi-labels and one-hot encoding. This
converts mixed types of defects into separate, single
defects so that you can effectively identify each
defect.
3 Pre-trained CNN-based TransUNet
3.1 Pre-training TransUNet using N-pair
Contrastive Loss
In this study, we use an N-pair Contrastive Loss
Function to pre-train a CNN on the encoder of the
TransUNet and then trained the entire TransUNet
using cross-entropy loss and dice loss. The cross-
entropy loss is the most commonly used loss
function for supervised learning. But, we use N-pair
contrastive loss in pre-train to ensure that features of
the same class are closer than features of different
classes. In our experiments, using this loss
outperform supervised learning using cross-entropy
loss. Encoder training based on N-pair contrast
losses can provide a better representation of the
latent dimension of wafer maps. It increases the
accuracy of the whole network.
By pre-training the encoder with N-pair contrast
loss, the distance between similar embeddings is
reduced. This allows for better feature learning in
the encoder stage and helps with the segmentation
of the decoder. To produce a feature representation
of the input image, we pre-trained the encoder on
150 epochs using N-pair contrast loss. Fig. 3 shows
the proposed pre-training structure.
Fig. 3: Overall architecture of proposed pre-training.
N-pair loss is a generalized version of triplet
loss, consisting of one anchor, one positive sample,
and negative samples. If , this is
equivalent to a triplet loss. This is optimized for
identifying a positive sample from multiple negative
samples. Consider the training data
, where is a positive sample
of and are negative. The -
tuple loss is defined as follows, where f is the
embedding kernel defined by the deep neural
network.
(1)
The multi-class N-pair loss (N-pair-mc) is
defined as follows:
(2)
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.27
Youngjae Kim,
Jee-Hyong Lee, Jongpil Jeong
E-ISSN: 2224-3402
240
Volume 20, 2023
3.2 Pseudo-Defect Masking
To train the segmentation model, we need images
and masks that serve as a label. Generally, labeling
is performed by humans using an image masking
program. These traditional approaches are time-
consuming and need a lot of labor. To overcome
that problem, we use a technique that automatically
masks defects. A defect represented by a wafer map
consists of a set of defects on a die. So, we perform
the masking in a way that separated the connected
pixels. We used the measure. label feature of Scikit-
image, and Fig. 4 shows the connectivity option. We
use 2-connectivity based on what we did with the
connectivity option, as shown in Fig. 5.
Fig. 4: Connectivity option description.
Fig. 5: Execution result by option.
4 Experiment and Results
4.1 Experiment Environments
To check our proposed model’s performance, we
conduct classification and segmentation with wafer
maps including both single and mixed defects.
Every experiment was run on four V100 GPUs with
16 CPU cores, 528 GB of MEM, and 32 GB of
RAM, using the torch open-source library. Table 1
summarizes the system specification.
Table 1. System specification
Hardware
Environment
Software Environment
CPU: 16Cores
MEM: 528GB
GPU: V100 x 4
Linux
Torch 1.4.0
Python3.7
4.2 Datasets
Mixed-Type Defects. We are aiming to identify
mixed defects. Therefore, we chose to use a mixed
wafer defect dataset offered by the Intelligent
Manufacturing Institute and Donghua University. It
contains both single defects and mixed defects.
Single defect classes are organized as follows
Center, Donut, Edge-Loc, Edge-Ring, Loc, Scratch,
Random, and Near-full. The center is defects
clustered in the center. Donut is a ring formed by
defects in the center. Edge-Loc is a cluster localized
at the edge. Edge-Ring is Ringed clusters around
boundaries. Loc is localized clusters that occur
regularly. Scratch is a distribution of defects in long,
narrow areas. Near-full is abnormal failure patterns.
Random is random defects with no pattern. Fig. 6
shows a Single wafer map defect. Fig. 7 shows a
mixed-type wafer map defect. For mixed-type
Defects, we used 10,400 of the two-type mixed
defects images as training data and 2,600 as
evaluation data.
Fig. 6: Single wafer map defect: (a) Center (C); (b)
Donut (D); (c) Edge-Loc (EL); (d) Edge-Ring (ER);
(e) Loc (L); (f) Near-full; (g) Scratch (S); (h)
Random.
Fig. 7: Mixed-type wafer map defect: (a) Center +
Edge-Loc; (b) Center + Scratch; (c) Donut + Edge-
Ring; (d) Donut + Scratch; (e) Center + Edge-Loc +
Loc; (f) Center + Edge-Ring + Scratch; (g) Donut +
Edge-Loc + Loc; (h) Edge-Loc + Loc + Scratch.
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.27
Youngjae Kim,
Jee-Hyong Lee, Jongpil Jeong
E-ISSN: 2224-3402
241
Volume 20, 2023
4.3 Results
Mixed-Type Defect Result. We compared our
performance to previous works in Table 2. Our
model was better than previous studies. After pre-
training the CNN among TransUNet’s encoders
using the N-pair contrastive loss function, the self-
attention-based transformer layer concentrated on
the defect and treated the neighboring defective dies
as noise, resulting in improved results. We can also
see better performance compared to previous
studies. For accuracy or F1-Score, our model shows
improvement. Fig. 8 illustrates the confusion matrix
of mixed-type defects. In problems such as
statistical classification in machine learning, a
confusion matrix is a table that allows us to
visualize the performance of a classification
algorithm trained with supervised learning. Each
column of the matrix represents an instance of the
predicted class and each row represents an instance
of the true class (or vice versa). Precision, indicated
by the yellow diagonal line, was able to achieve a
result of 0.995. Table 3 details the model’s
performance in detecting two types of mixed
defects. The F1-Socre for two-type mixed defects
reached 0.995.
Table 2. Comparison with other models
Model
Accuracy
F1-Score
Our Model
[12]
[13]
[14]
[15]
0.995
0.826
0.962
0.977
0.980
0.995
0.824
0.962
0.977
0.974
Fig. 8: Testing results of mixed-type defects:
normalized confusion matrix (C + EL, C + ER, C +
L, C + S, D + EL, D + ER, D + L, D + S, EL + L,
EL + S, EL + L, ER + S, L + S).
Table 3. Mixed-type testing result
Defect
Type
F1-Score
baseline
w/multitask
w/pretrain+multitask
C + EL
C + ER
C + L
C + S
D + EL
D + ER
D + L
D + S
EL + L
EL + S
ER + L
ER + S
L + S
0.903
0.955
0.954
0.981
0.930
0.968
0.924
0.975
0.875
0.890
0.852
0.917
0.965
0.895
0.959
0.953
0.984
0.941
0.972
0.942
0.979
0.887
0.921
0.875
0.918
0.969
0.990
0.993
0.997
1.000
1.000
1.000
0.995
0.995
0.998
0.984
0.999
0.986
0.999
Fig. 9 shows some samples of inference results
for two-type mixed defects. It is composed of the
original image, pseudo-label, and inference results.
When benchmarked against the pseudo-label mask,
we can check that it correctly predicts a single
defect. Since the pseudo label data serves as the
label, segmentation results are produced as close to
the pseudo label. Since the masking was done
programmatically, there were mismatches in the
defects, but there was no problem detecting the
defects. It could even correctly predict the majority
of defects that were a mix of both types, but it was
difficult to recognize when two defects overlapped.
By more than a certain percentage, such as when
Local and Edge-Loc defects overlapped, or when
Scratches overlapped with other defects.
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.27
Youngjae Kim,
Jee-Hyong Lee, Jongpil Jeong
E-ISSN: 2224-3402
242
Volume 20, 2023
Fig. 9: Inference results in two-types mixed defects:
(a) Center + Edge-Ring defect; (b) Center + Loc
defect; (c) Center + Scratch defect; (d) Donut +
Edge-Loc defect; (e) Donut + Loc defect; (f) Donut
+ Scratch defect; (g) Edge-Loc + Loc defect; (h)
Edge-Loc + Scratch defect; (i) Edge-Ring + Scratch
defect; (j) Loc + Scratch defect.
5 Conclusion
In this study, we used an N-pair Contrastive Loss
Function to pre-train a CNN on the encoders of the
TransUNet and then learn the entire TransUNet,
resulting in good feature maps that emphasize the
details of defects we were looking for. Human
defect labeling is a very inefficient method because
it is subjective and labor-intensive. In this study, we
used automatic masking to solve the inefficiency
problem. Our suggested approach delivered better
results than the previous approach. Accuracy was
0.995, and F1-Score was 0.995. These results
provide engineers with the exact location of the
error, which helps them determine the cause of the
problem. The study allowed us to detect mixed
defects, and the automatic masking technique saved
us unnecessary labor and time. This saves labor for
existing workers and provides accurate defect
detection.
We tested the model in public large wafer map
datasets, but further validation on real-world
datasets can be considered in future work. We can
also use methods like transfer learning. In the future,
we will concentrate our research on decreasing the
size of the model and increasing performance.
Acknowledgment:
This research was supported by the SungKyunKwan
University and the BK21 FOUR(Graduate School
Innovation) funded by the Ministry of
Education(MOE, Korea) and the National Research
Foundation of Korea(NRF). And this work was
supported by the National Research Foundation of
Korea (NRF) grant funded by the Korean
government (MSIT) (No. 2021R1F1A1060054).
References:
[1] Shu-Kai S. Fan, Chia-Yu Hsu, Du-Ming Tsai,
Fei He, Chun-Chung Cheng, Data-driven
approach for fault detection and diagnostic in
semiconductor manufacturing, IEEE
Transactions on Automation Science and
Engineering, Vol.17, Issue.4, 2020, pp. 1925–
1936.
[2] Erik H.M. Heijne, Future semiconductor
detectors using advanced microelectronics
with post-processing, hybridization and
packaging technology. Nuclear Instruments
and Methods in Physics Research Section A:
Accelerators, Spectrometers, Detectors and
Associated Equipment, Vol.541, Issues.1-2,
2020, pp. 274–285.
[3] Chih-Hung Chang, Brian K. Paul, Vincent T.
Remcho, Sundar Atre, James E. Hutchison,
Synthesis and post-processing of
nanomaterials using microreaction
technology, Journal of Nanoparticle
Research, Vol.10, 2008, pp. 965–980.
[4] P. Grybos, Front-End Electronics for
Multichannel Semiconductor Detector
Systems, Institute of Electronic Systems,
Warsaw University of Technology: Warsaw,
Poland, 2020, pp. 132–135.
[5] Ofer Sneh, Robert B. Clark-Phelps, Ana R.
Londergan, Jereld Winkler, Thomas E. Seidel,
Thin film atomic layer deposition equipment
for semiconductor processing, Thin Solid
Film, Vol.403, Issues.1-2, 2020, pp. 248–261.
[6] S. Joe Qin, Gregory Cherry, Richard Good,
Jin Wang, Christopher A. Harrison,
Semiconductor manufacturing process control
and monitoring: A fab-wide framework,
Journal of Process Control, Vol.16, Issue.3,
2006, pp. 179–191.
[7] Seokho Kang, Sungzoon Cho, Daewoong An,
Jaeyoung Rim, Using wafer map features to
better predict die-level failures in final test,
IEEE Transactions on Semiconductor
Manufacturing, Vol.28, Issue.3, 2015, pp.
431–437.
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.27
Youngjae Kim,
Jee-Hyong Lee, Jongpil Jeong
E-ISSN: 2224-3402
243
Volume 20, 2023
[8] Minghao Piao, Cheng Hao Jin, Jong Yun Lee,
Jeong-Yong Byun, Decision tree ensemble-
based wafer map failure pattern recognition
based on radon transform-based features.
IEEE Transactions on Semiconductor
Manufacturing, Vol.31, Issue.2, 2018, pp.
250–257.
[9] Cheng Hao Jin, Hyuk Jun Na, Minghao Piao,
Gouchol Pok, Keun Ho Ryu, A novel
DBSCAN-based defect pattern detection and
classification framework for wafer bin map,
IEEE Transactions on Semiconductor
Manufacturing, Vol.32, Issue.3 2019, pp.
286–292.
[10] Hyuck Lee, Heeyoung Kim, Semi-supervised
multi-label learning for classification of wafer
bin maps with mixed-type defect patterns,
IEEE Transactions on Semiconductor
Manufacturing, Vol.33, Issue.4, 2020, pp.
653–662.
[11] Kiryong Kyeong, Heeyoung Kim,
Classification of mixed-type defect patterns in
wafer bin maps using convolutional neural
networks, IEEE Transactions on
Semiconductor Manufacturing, Vol.31,
Issue.3, 2018, pp. 395–402.
[12] Junliang Wang, Chuqiao Xu, Zhengliang
Yang, Jie Zhang, Xiaoou Li, Deformable
convolutional networks for efficient mixed-
type wafer defect pattern recognition, IEEE
Transactions on Semiconductor
Manufacturing, Vol.33, Issue.4, 2020, pp.
587–596.
[13] Jinho Kim, Youngmin Lee, Heeyoung Kim,
Detection and clustering of mixed-type defect
patterns in wafer bin maps, IISE Transactions,
Vol.50, Issue.2, 2018, pp. 99–111.
[14] Ming-Chuan Chiu, Tao-Ming Chen, Applying
Data Augmentation and Mask R-CNN-Based
Instance Segmentation Method for Mixed-
Type Wafer Maps Defect Patterns
Classification, IEEE Transactions on
Semiconductor Manufacturing, Vol.34,
Issue.4, 2021, pp. 455–463.
[15] Jaegyeong Cha, Jongpil Jeong, Improved U-
Net with Residual Attention Block for Mixed-
Defect Wafer Maps. MDPI Applied Sciences
Journal, Vol.12, Issue.4, 2022.
[16] Ozan Oktay, Jo Schlemper, Loic Le Folgoc,
Matthew Lee, Mattias Heinrich, Kazunari
Misawa, Kensaku Mori, Steven McDonagh,
Nils Y Hammerla, Bernhard Kainz, Ben
Glocker, Daniel Rueckert, Attention U-Net:
Learning Where to Look for the Pancreas,
arXiv 2018, arXiv:1804.03999.
[17] Yu-Cheng Liu, Mohammad Shahid,
Wannaporn Sarapugdi, Yong-Xiang Lin, Jyh-
Cheng Chen, Kai-Lung Hua, Cascaded atrous
dual attention U-Net for tumor segmentation,
Multimedia Tools and Applications, Vol.80,
2021, pp.30007-30031.
[18] Xiaocong Chen, Lina Yao, Yu Zhang,
Residual Attention U-Net for Automated
Multi-Class Segmentation of COVID-19
Chest CT Images, arXiv.2004,
arXiv:2004.05645.
[19] Jieneng Chen, Yongyi Lu, Qihang Yu,
Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu,
Alan L. Yuille, Yuyin Zhou, TransUNet:
Transformers Make Strong Encoders for
Medical Image Segmentation, arXiv 2021,
arXiv:2102.04306.
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
-Youngjae Kim set the research topic and goals,
developed the software, conducted the experiments,
validated, and wrote the paper.
-Jee-Hyong Lee and Jongpil Jeong conducted the
review and the editing.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
This research was supported by the SungKyunKwan
University and the BK21 FOUR(Graduate School
Innovation) funded by the Ministry of
Education(MOE, Korea) and the National Research
Foundation of Korea(NRF). And this work was
supported by the National Research Foundation of
Korea (NRF) grant funded by the Korean
government (MSIT) (No. 2021R1F1A1060054).
Conflict of Interest
The authors have no conflicts of interest to declare.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.27
Youngjae Kim,
Jee-Hyong Lee, Jongpil Jeong
E-ISSN: 2224-3402
244
Volume 20, 2023
