IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 2, NO.

1, JANUARY 2017 179

A Scheme of High-Dimensional Key-Variable Search

Algorithms for Yield Improvement
Fan-Tien Cheng, Fellow, IEEE, Yao-Sheng Hsieh, Jing-Wen Zheng, Szu-Min Chen, Ren-Xiang Xiao, and Chin-Yi Lin

Abstract—Product yield directly affects production cost. Thus

manufacturers seek to quickly enhance product yield during the
development and mass-production processes. In other words, when
a yield loss occurs, the root causes should be found rapidly in both
the development and mass-production phases. When a yield loss
is encountered, the traditional yield enhancement approach is to
collect all production-related data to perform big data analysis in
order to find out the root causes of affecting yield and remedy
them. However, production-related data are extremely large and
complicated, which makes it hard to search for the root causes of a
yield loss. This leads to the problem of high-dimensional variable Fig. 1. Yield and cost changes during product life cycle.
selection. To solve this problem, a scheme of Key-variable Search
Algorithms (KSAs) is proposed in this paper. The inputs of this
KSA scheme include production routes, process data, inline data, quality in the mass-production phase as well as finding out and
defects, and final inspection results; while the outputs are search resolving the yield-loss problem encountered.
results and their corresponding reliance indices. The search results The yield management system aims to enhance product yield.
are the key stages that cause the yield loss or the key variables that
are the root causes of the yield loss. The thin film transistor-liquid
However, the number of workpieces is small at the research-
crystal display (TFT-LCD) process is adopted as the illustrative and-development (RD) and ramp-up phases, and that makes it
example to demonstrate the KSA scheme. hard to find out the root causes of defects among all production
tools. Thus, how to find out the key variables affecting yield
Index Terms—Factory automation, high-dimensional variable
selection, intelligent and flexible manufacturing, key-variable among numerous production tools with limited samples is a
search algorithm (KSA), semiconductor manufacturing. widely concerned issue. This challenge is the so-called high-
dimensional variable selection problem, which is also denoted
as the issue of p  n. “p” in p  n represents the number
I. INTRODUCTION of explanatory variables in the models and means all process-
ITH the advancement of semiconductor and thin film related parameters in semiconductor or TFT-LCD process. “n,”
W transistor-liquid crystal display (TFT-LCD) manufac-
turing technologies, their manufacturing processes are becom-
on the other hand, is the sampling number that is also the amount
of end products in semiconductor or TFT-LCD process.
ing more and more sophisticated. As a result, how to maintain Several approaches have been proposed to solve the high-
feasible production yield of these sophisticated manufacturing dimensional variable selection problem. Among them, Chen
processes becomes an essential issue. As shown in Fig. 1 [1], and Hong proposed sample-efficient regression trees (SERT)
product yield (blue line) will gradually rise up in the ramp-up [2] for semiconductor yield loss analysis, which combined the
phase, and then keep steady in the mass-production phase. On methodologies of forward selection in regression analysis and
the contrary, product cost (red line) will decrease as the phases regression tree. Compared to the conventional decision trees
proceed. Company’s competitiveness would be effectively en- like CART (classification and regression trees), SERT is able
hanced if the blue/red solid lines could be improved into their to handle combination effect in the so-called high-dimensional
corresponding segmented lines. This implies rapidly increasing (p  n) problem. However, in order to handle different types
the yield in the ramp-up phase, and promptly transferring prod- of input variables (continuous, ordinal, binary, and categorical),
ucts into the mass-production phase; then enhancing product SERT needs to transfer all types of input data into binary form
during data preprocessing.
Manuscript received March 01, 2016; accepted June 10, 2016. Date of pub- Data-mining approaches were adopted by Chien et al. [3] to
lication June 23, 2016; date of current version July 28, 2016. This paper was infer possible causes of yield losses and manufacturing process
recommended for publication by Associate Editor N. Huang and Editor J. Li variations. The inputs of this work are chip-probing (CP) yield
upon evaluation of the Reviewers’ comments. This work was supported by the
Ministry of Science and Technology of the Republic of China under Contract results as well as process data, equipment data, and lot history
MOST 104-3011-E-006-001. during wafer fabrication. This work aimed to develop a frame-
The authors are with the Institute of Manufacturing Information and Systems, work for data mining and knowledge discovery from database
National Cheng Kung University, Tainan 701, Taiwan (e-mail: chengft@mail.
ncku.edu.tw; [email protected]; [email protected] that consists of a Kruskal–Wallis test, K-means clustering, and
.ncku.edu.twl; [email protected]; [email protected] the variance reduction splitting criterion to investigate the huge
.edu.tw; [email protected]). amount of semiconductor manufacturing data.
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. Hsu et al. [4] also proposed a data mining framework for find-
Digital Object Identifier 10.1109/LRA.2016.2584143 ing root causes of defects for TFT-LCD manufacturing. Rough
2377-3766 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
180 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 2, NO. 1, JANUARY 2017

set theory (RST) was applied to generate the candidate rules

and then these rules were utilized to locate the root causes after
several rounds of cross-validation. Decision tree (DT) was also
used to find out the root causes. Different attribute evaluation
criteria were applied by these two methods. RST aims to elim-
inate redundant or surplus attributes in the decision table and
DT explains the data based on the maximum information value
during each splitting tree process.
Ing and Lai [5] proposed to apply orthogonal greedy al-
gorithm (OGA) together with high-dimensional information
criterion (HDIC) and backward trimming (Trim) to solve the Fig. 2. Process flow of TFT-LCD manufacturing.
high-dimensional variable selection problem. OGA is a fast
stepwise regression method which forward selects input vari-
ables to enter a p-dimensional linear regression model (p  n).
Compared to the usual model selection criteria, HDIC+Trim II. INTRODUCING THE TFT-LCD MANUFACTURING PROCESSES
work well in the high-dimensional case, and also resolve the is- Four processes constitute the TFT-LCD manufacturing flow
sue of spurious variable selection when the number of variables as displayed in Fig. 2 [8], which are TFT, CF (color filter), LCD,
is much larger than the sample size. and LCM (liquid crystal module). To start with, the transistors
Unlike stepwise regression, a least absolute shrinkage and are manufactured on a piece of glass substrate in the TFT pro-
selection operator (LASSO) regression [6], [7] was also pro- cess. Next, in the CF process, red, green, and blue color sources
posed to solve the high-dimensional variable selection problem. are fabricated on a glass. And, a TFT substrate and a CF sub-
LASSO is a shrinkage method which minimizes the residual strate are merged in the LCD process with the space between
sum of squares subject to the sum of the absolute value of the these two substrates being filled with liquid crystal. Then, addi-
coefficients being less than a constant. Because of the nature of tional components like driver integrated circuits and backlight
this constraint, it tends to produce some coefficients that are ex- units are connected to comprise the final fabricated products in
actly zero and hence gives interpretable models. The significant the LCM process. Among them, the TFT, CF, and LCD pro-
difference between LASSO and the usual stepwise regression is cesses require high-degree automation and are integrated into
that all the independent variables can be processed at the same the front-end process; while the LCM process, which is mostly
time via LASSO but not the usual stepwise regression. labor-oriented, is denoted as back-end process. This study em-
The approaches surveyed above emphasized on adopting var- phasizes on the front-end process.
ious data-mining and/or regression algorithms to resolve the As shown in Fig. 3, the TFT process consists of five lay-
problem of finding the root causes of yield losses under the high- ers: gate, semiconductor, data, protection, and indium tin oxide
dimensional environment; while the issues of input data quality (ITO) layers. Each layer includes the so-called photo engraving
and reliability level of the search results were not addressed. processes (PEP) such as: film deposition, positive photoresist
The purpose of this paper is to develop a scheme with Key- coating, exposure, developing, etching, and stripping. The coat-
variable Search Algorithms (KSAs) that not only contains new ing, exposure, and developing processes can be combined into
search algorithms but also includes data preprocessing modules the photo step. The process flow of the semiconductor layer, as
for checking the quality of all the input data and reliance-index depicted in Fig. 4, is the most complicated layer among those
module for evaluating the reliability of the search results. More- five ones.
over, a two-phase process is suggested in this paper to apply The first step of the semiconductor process flow is to deposit
the KSA scheme for searching the root causes of yield losses film via CVD tools with the film thickness being the sampling
with Phase I for searching the key stages and Phase II for select- measurement value. In the photo step, before acquiring the sam-
ing the key variables within the most suspicious stage found in pling measurement value (critical dimension, CD of line-width),
Phase I. the workpiece executes coating, exposure, and developing pro-
The KSA scheme is a general one for a wide range of manu- cesses. After that, the etching process follows where the etching
facturing systems, such as semiconductor, TFT-LCD, and bump- depth is taken as the sampling measurement value. Finally, with
ing processes. Due to page limitation, it is impossible to include the metal width being the sampling measurement value, the
all the descriptions of semiconductor, TFT-LCD, and bumping stripper process removes the photoresist on the workpiece.
processes in this paper. Therefore, the typical TFT-LCD man- After going through those five stages shown in Fig. 3, yield
ufacturing process is selected as the illustrative example for tests will be conducted. Yield tests may encounter Type 1 ∼ Type
implementing the KSA scheme. 10 yield losses that are resulted from various defects caused by
The remainder of this paper is organized as follows. Section II electric-test failures, particles, etc.
introduces the TFT-LCD manufacturing process. Section To finish the TFT manufacturing process, PEP should be
III describes the traditional root-cause search procedure of adopted onto each layer, thus there are five PEPs in total in the
a yield loss. Section IV explains the KSA scheme. Section V entire TFT process. This paper will adopt the TFT process for
presents the illustrative example. Finally, summary and conclu- illustration. Due to paper-length limitation, the descriptions of
sion are stated in Section VI. CF and LCD processes are omitted.
CHENG et al.: SCHEME OF HIGH-DIMENSIONAL KSAs ALGORITHMS FOR YIELD IMPROVEMENT 181

Fig. 3. TFT manufacturing process.

Fig. 4. PEP flow of the semiconductor layer.

Step 6: If the root cause is found, jump to Step 7; otherwise,

move back to Step 4.
Step 7: Issue a notice to the relevant departments for fixing
the problem and for continuous improvement.
The major drawbacks of the traditional root-cause search
process are: 1) the information contained in the defect mode
database may not be complete such that associated sampling
rates need to be raised, which will increase the production cost;
2) even the current defect mode is in the database, the search
tool (such as EDA) may not find out the root cause easily.
In fact, this root-cause search issue is a high-dimensional
variable selection problem. To solve this problem economically
and efficiently, the KSA scheme is proposed and will be
presented next.

IV. KSA SCHEME

The KSA scheme contains various data preprocessing mod-
Fig. 5. Traditional root-cause search process of a yield loss. ules, KSA module and reliance index (RIk ) module as shown
in Fig. 6. Besides the fundamental functions of KSA and search
III. TRADITIONAL ROOT-CAUSE SEARCH PROCEDURE quality evaluation, the KSA scheme also possesses the func-
OF A YIELD LOSS tions of automatic data quality evaluation. The inputs of a KSA
scheme include production routes (XR ), process data (XP ),
The traditional root-cause search procedure of a yield loss is and in-line metrology values (y) as well as the final inspection
depicted in Fig. 5 and is described below. data (Y). Note that y may be provided by actual measurement
Step 1: When a yield test failure is encountered, the root-cause and/or virtual metrology [8], [11]. Defects (D) could happen in
search procedure is launched. any production stage, thus they are also imported into the KSA
Step 2: The defect mode database is searched to see if this scheme for analysis. Then, the processed D may join the pro-
defect mode exists in the database or not. duction information group or final inspection group depending
Step 3: If this defect mode exists, then jump to Step 5; other- on the nature of D. The outputs are the search result of Triple
wise, continue to Step 4. Phase Orthogonal Greedy Algorithm (TPOGA): KSO and that
Step 4: Due to the fact that this defect mode is new and its of Automated Least Absolute Shrinkage and Selection Operator,
characteristics are not well known yet, the sampling (ALASSO): KSL , and their accompanying reliance index: RIK .
rate related to the in-line inspections of this defect Those functional modules are described below.
should be increased so as to enhance the defect mode
database.
A. Data Preprocessing
Step 5: Apply a suitable software tool, such as an engineer-
ing data analysis (EDA) [9], to find the root cause of Observing Fig. 6, before performing the KSA module, data
defect. preprocessing must be conducted to assure the data quality of
182 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 2, NO. 1, JANUARY 2017

Fig. 6. KSA Scheme.

all the inputs: production information (XR , XP , and y), defects high-dimensional linear regression [5]. The steps of OGA are
(D), and final inspections (Y). The characteristics of these inputs briefly descried below.
are described below. Step 1: Define R0 = Y = (y1 , y2 , . . . , yn )T . Choose the
XR needs to be discretized into 1 or 0, which indicates that variable defined as xŜ 1 that is most correlated with
the workpiece getting through this stage or not. XP contains R0 in {X = x1 , x2 , . . . , xp }. Then, the correspond-
tool process data (such as voltage, pressure, temperature, etc.) ing residual will be
which need to be centralized. y stands for inline inspection data
R̂1 = R0 − β̂ŝ11 xŝ 1 (1)
(such as critical dimension, thickness, etc.) which need to be
centralized. As for D, different companies have different defini- where
tions of defects, thus discussion with domain experts is required p number of parameter;
before executing data-preprocessing and quality check. Finally, n sample size;
Y stands for the yield test results that should be centralized. xŜ 1 highest correlation variable with R0 in X;
The data quality evaluation algorithm of XR , DQIX R , eval- 1
β̂ŝ 1 regression coefficient of R0 for xŜ 1.
uates the following four facts: 1) a stage may contain several Step 2: Choose another variable, xŜ 2 , which is most corre-
devices of the same type, while a stage utilizes only one device;
if a process should get through three devices, it then has three lated with R̂1 in X. Then, the corresponding residual
stages; 2) if a device is used in different processes, the same will be
device in a different process would be considered as a different R̂2 = R0 − β̂ŝ11 xŝ 1 − β̂ŝ22 xŝ 2 (2)
stage; 3) there are only two possibilities for a workpiece passing
through the device: get through (1) or not (0); 4) a workpiece where β̂ŝ22 is the regression coefficient of R0 for xŜ 2 .
cannot get through any device that doesn’t belong to that stage. Step 3: Return to Step 2 and repeat m times so as to stepwise
Similarly, the data quality evaluation algorithms of XP and y choose xŜ 1 , xŜ 2 , . . . , xŜ m and calculate the corre-
are denoted as DQIX P and DQIX y , respectively. Both DQIX P , sponding regression coefficients (β̂ŝ11 , β̂ŝ22 , . . . , β̂ŝmm ).
and DQIX y adopt the algorithm similar to the process data Then, the corresponding residual will be
quality evaluation scheme utilized in AVM [10], [11].
Finally, the data quality evaluation algorithm of Y is denoted R̂m = R0 − β̂ŝ11 xŝ 1 − β̂ŝ22 xŝ 2 − . . . −β̂ŝ2m xŝ m . (3)
as DQIY . DQIY applies the algorithm similar to the metrology However, it is hard to select exact m features to recover the
data quality evaluation scheme used in AVM [10], [11]. entire original signal. Therefore, Ing and Lai proposed a termi-
nation condition, HDIC [5], to choose along the OGA path that
B. KSA Module has the smallest value of a suitably chosen criterion. Let
To double check the reliability of the search results, the KSA HDIC(J) = nlogϑ̂2J + #(J)wlogp (4)
module contains two algorithms: TPOGA and ALASSO. They
with
are described below.
1
n
1) Triple Phase Orthogonal Greedy Algorithm, TPOGA:
The greedy algorithm is a stepwise regression method that con- ϑ̂2J = (yi − ŷi;J )2
n i=1
siders the correlation between all the causing parameters (X)
and the results (Y). In this study, X includes all the related where
variables of production: XR , XP , and y; while Y represents the J set of variables selected in the model (xŝ 1, xŝ 2 , . . . ,
final inspection. xŝ m );
Pure greedy algorithm (PGA) and OGA are commonly ϑ̂2J mean square error of the corresponding model;
used in the literature for solving the high-dimensional regres- w general constant penalties > 0;
sion problem. In general, OGA performs better than PGA in yi ith sample of actual value of final inspection;
CHENG et al.: SCHEME OF HIGH-DIMENSIONAL KSAs ALGORITHMS FOR YIELD IMPROVEMENT 183

ŷi;J ith sample of predictive value of final inspection corre-

sponding to J.
Define Jˆm to be the set of all the variables selected after
executing OGA+HDIC via (4), where m represents the number
of variables selected in the model. However, there are still some
irrelevant variables in Jˆm . The method of removing irrelevant
variables is needed.
Trimming is the method of rechecking all the variables se-
lected in the model so as to remove those irrelevant ones. The
trimming steps [5] are described below.
Step 1: Based on the concept of OGA+HDIC, define a subset
N̂ of Jˆm , where N̂ is the real relevant variables in the
model.
Step 2: N̂ = {Jˆl : HDIC(Jˆm − {Jˆl }) > HDIC(Jˆm ), 1 ≤
l ≤ m} if m > 1. N̂ = {Jˆ1 } if m = 1.
Step 3: Repeat Step 2 for m − 1 times.
Finally, combining the processes of OGA, HDIC, and Trim- Fig. 7. Flowchart of ALASSO with automated λ adjusting. λ: penalty; KV:
key variables; #KV: number of KV; Target#KV: target number of KV.
ming, the so-called TPOGA is formed.
2) Automated Least Absolute Shrinkage and Selection Op-
The value of λ should be set before solving (7). The larger
erator, ALASSO: In general regression models, coefficients are
value of λ is set, the less number of key variables will be gener-
obtained with a least-squares method. Set the loss function as
ated. If the value of a regression coefficient is zero shown in the
||Y − XB||2 . The penalty function f(λ, β) is added into the reg-
solution, then it means that its corresponding variable is deleted.
ularization regression, where λ being defined as penalty and
To reduce the puzzle of setting λ and obtain proper results,
β as regression coefficient. When the penalty function equals
the so-called Automated LASSO (ALASSO), which contains
λ||β||2 , this method is called ridge regression; on the other
the automated λ adjusting method, is proposed in this paper.
hand, when the penalty function equals λ||β||, then the method
The flowchart of ALASSO is shown in Fig. 7 and described
is known as LASSO regression [6], [7]. In LASSO regression,
below.
the value of λ may affect the coefficients of the variables. In
Step 1: The initial value of λ is set to be 5 by experience.
addition, if the coefficients equal 0, it means that these variables
Step 2: Execute LASSO.
are not suggested of being selected into the model. The LASSO
Step 3: Select a proper target number of key variables
formulation is stated as follows [6], [7]:
(Target#KV); for example: 10. Then, check that the
⎛ ⎞2
number of key variables (#KV) selected by LASSO
 n p
β̂ lasso = argmin ⎝yi − β0 − xij βj ⎠ (5) equals Target#KV or not. If no, go to Step 4.1 to recal-
β i=1 j =1 culate λ; otherwise go to Step 4.2. The root causes of
a yield loss should be focused. Therefore, Target#KV

p
is set to avoid picking up too many unnecessary vari-
subject to |βj | ≤ t (6)
ables.
j =1
Step 4.1: With Lowerbound being defined as the lower bound
where of λ. The initial value of Lowerbound is 0 (be-
yi final inspection result of the ith sample; cause λ ≥ 0). If #KV < Target#KV, λnew = (λold +
xij jth parameter of the ith sample; Lowerbound)/2. On the other hand, if #KV >
n sample size; Target#KV, λnew = λold + (λold + Lowerbound)/2
p number of parameter; and re-assign Lowerbound = λold . Finally, return to
t tuning parameter which restricts the sum of |βj |; Step 2.
βj regression coefficient of jth parameter; Step 4.2: Sort the KV by importance in descending order.
β̂ lasso optimized coefficients. Step 5: Output the sorted KV as the result.
Equation (6) is the constraint that sets the sum of β absolute
values to be less than t. With this constraint, residual sum of C. RIK Module
squares are minimized so as to make some regression coeffi- RIK is designed to evaluate the reliability of KSA results with
cients be equal to 0 for obtaining an explanatory model. reference to the reliance index (RI) concept of the AVM system
To solve (5), combine (5) and (6) such that the LASSO prob- [11], [12]. By comparing the results of TPOGA and ALASSO
lem is written in the equivalent Lagrangian form: and taking overlapping and weights into considerations, RIK is
⎧
2 ⎫ re-calculated and set between 0 ∼ 1. With RIKT = 0.7 being the
⎨1 
n 
p

p ⎬
β̂ la sso = argmin yi − β0 − xij βj +λ |βj | . threshold, if RIK is greater than RIKT , good search results are
β ⎩2 ⎭ obtained; otherwise, the search results need to be re-examined.
i= 1 j=1 j=1
(7) The calculation of RIK is described below.
184 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 2, NO. 1, JANUARY 2017

TABLE I
TOTAL NUMBER OF DEVICES IN THE TFT PROCESS

Layer Process Total no. of devices

PEP 1: Gate Film Deposition (CVD) 34

Lithography 25
Etching 14
Stripping 11
Measurement 43
Others 13
PEP2: Semiconductor Film Deposition (CVD) 70
Lithography 25
Etching 38
Stripping 11
Measurement 33
Fig. 8. Procedure for finding the root causes of a yield loss by applying the
PEP3: Data Film Deposition (CVD) 34
KSA scheme.
Lithography 25
Etching 45
Suppose that up to ten key variables are selected by both Stripping 11
Measurement 43
TPOGA and ALASSO. The order of each key variable being PEP4: Protection Film Deposition (CVD) 70
picked out is meaningful, thus the original scores of the first to Lithography 25
tenth variables descend from 1 ∼ 0.1 sequentially. Etching 38
Stripping 11
Since the sequential order of key variables searched by Measurement 33
TPOGA and ALASSO is important as the variable chosen ear- PEP5: ITO Film Deposition (CVD) 17
lier matters more, weights are assigned not only according to Lithography 25
Etching 4
the sequential order but also the 80-20 principle [13], [14], to Measurement 22
ensure higher score goes to the crucial minority. As a result, the Others Annealing 4
final scores become Electrical Test 13
⎧ Final Repair 26
⎪ OSl Stocker 26
⎪
⎪  × 0.8 , when pick order is 1, 2, 3.
⎨ 3 OSl
l=1
FSl = (8)
⎪
⎪ OSl
⎪
⎩ 10 × 0.2, when pick order is 4, 5, . . . , 10 Step 2: The data of yield test result (Y) as well as inline data
l=4 OSl
(y) and production routes (XR ) are fed into the KSA
where OSl is the original score and FSl is the final score with scheme to search for the Top N (=3 for example)
l = 1, 2, . . . , 10 being the pick order. devices that cause the yield loss.
The rule of thumb of calculating RIK is explained below. If a Step 3: The same data of yield test result (Y) and the process
certain variable has been chosen by both TPOGA and ALASSO data ((XP )) of all the devices of the stage to which the
with the same sequential order, the score of this variable is most suspicious device belongs are fed into the KSA
counted. Then, if a certain variable has been chosen by both scheme to search for the key variables that cause the
TPOGA and ALASSO with different sequential orders, the score yield loss.
of this variable is also counted. Finally, if TPOGA and ALASSO Step 4: Issue a notice to the relevant departments for fixing
pick out different variables without overlapping, the score of this the problem and for continuous improvement.
variable is not counted. Therefore, RIK of the search results is
calculated by V. ILLUSTRATIVE EXAMPLES
10 10  
F SO i + F SL j The TFT array process data for TFT-LCD manufacturing at
RIK = if Oi = Lj (9) AUO Industries, Inc., Taiwan, are adopted as the illustrative ex-
i=1 j =1
2
ample. Lot is the basic tracking unit of TFT production process.
where There are a total of 28 sheets in a full lot, and each sheet can be
FSOi final score of Oi cut into six chips. There are five PEP layers in the TFT process
FSLj final score of Lj and each layer consists of different number of devices. The pro-
Oi ith pick variable of TPOGA, i = 1, 2, 3, . . . , 10 cess route descriptions, including layers, processes, and number
Lj jth pick variable of ALASSO, j = 1, 2, 3, . . . , 10. of devices are shown in Table 1.
An example of RIK calculation will be presented in Section V. Both etching and film deposition machines contain multiple
Upon finishing description of the entire functions of the KSA chambers. Each chamber is considered as a different device in
scheme, the procedure for finding the root causes of a yield loss this study, and each device in the algorithm is treated as a vari-
by applying the KSA scheme is shown in Fig. 8 and explained able in XR . By summing up all the numbers of devices shown
below. in Table I, there are a total of 789 variables in this example. Note
Step 1: When a yield test failure is encountered, the root-cause that all kinds of variables in XR are denoted as devices in this
search procedure is launched. paper. Therefore, a chamber in a production tool, a metrology
CHENG et al.: SCHEME OF HIGH-DIMENSIONAL KSAs ALGORITHMS FOR YIELD IMPROVEMENT 185

Fig. 9. Accumulated Type 2 Loss results.

station, or a stocker, and so on is treated as a device as shown in

Table I. Fig. 10. RIK result of XR search.
The count of defects is by chip, along with 10 different yield-
test types. Defects will be allocated in percentage into these
ten types according to their causes. For example, a certain chip
has defects of Types 1 & 3 with percentages of 66% and 33%
respectively, then the yield-test result of this chip will be [0.66
0 0.33 0 0 0 0 0 0 0]. If there are no defects detected in this
chip, the result is [0 0 0 0 0 0 0 0 0 0]. Each sheet contains
six chips. Therefore, the counts of defects of a sheet are to add
up the yield-test results of all six chips. After discussing with
domain experts, Type 2 Loss is the most critical defect, thus the
analysis of this defect is conducted first within the time period
of 9 days.
A total of 113 lots were processed during this 9-days period.
However, only 104 lots are kept after filtering through DQIXR .
The accumulated Type 2 Loss results of these 104 lots are shown
in Fig. 9 with the largest value (6.9) being occurred at Lot 49.
Thus, the root causes of Type 2 Loss at Lot 49 are searched as
follows.
As indicated in Step 2 of Fig. 8, besides Y, both y and XR are
required to be the inputs for searching the Top N devices. To do
so, total inspection is required for inline metrology. However, the
inline data, y, have merely sampling inspection in this example;
therefore, only XR is adopted as input in this case.
Lot 49 contains 28 sheets while there are a total of 789 devices
in this TFT process. Therefore, this high dimensional variable
selection problem has (p=) 789 variables of XR and (n=) 28
samples. The yield test result (Y) is Type 2 Loss. By applying the
Fig. 11. KSA search results of Type 2 Loss on Lot 49. (a) Top 1 Device:
KSA scheme, the Top 10 devices searched by both TPOGA and Chamber B of Equipment 7 of Film Deposition process in Protection layer.
ALASSO are shown in Fig. 10. A blue bar indicates that a certain (b) Top 2 Device: Chamber A of Equipment A of Film Deposition process in
variable has been chosen by both TPOGA and ALASSO with Semiconductor layer.
the same sequential order, the score of this variable is counted;
a brown bar means that a certain variable has been chosen by the results are reliable. The top two devices are investigated as
both TPOGA and ALASSO with different sequential orders, follows.
the score of this variable is also counted; and a white bar shows As shown in Fig. 11, the Top 1 device is Chamber B of
that TPOGA and ALASSO pick out different variables without Equipment 7 of Film Deposition process in Protection layer;
overlapping, the score of this variable is not counted. As such, while Top 2 device is Chamber A of Equipment A of Film
RIK is calculated by (8) and (9). The result of RIK is 0.932 that Deposition process in Semiconductor layer. Fig. 11(a) shows
is larger than the threshold (0.7). This implies that the search that among those 28 sheets, 3 out of 8 Type 2 Loss samples
results of TPOGA and ALASSO are almost the same, and thus were processed by the Top 1 device. Also, Fig. 11(b) indicates
186 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 2, NO. 1, JANUARY 2017

to the KSA scheme to find Top N devices that cause the

yield losses.
2) Phase II: After obtaining the most suspicious device from
Phase I, input yield test results (Y) as well as the process
data (XP ) of all the devices of the stage to which the
most suspicious device belongs to the KSA scheme to
drill down and find out the root causes.
The TFT-LCD process is adopted as the illustrative example
Fig. 12. RIK result of Xp search. for demonstration. The test results reveal that the KSA scheme
is a promising tool to solve the high dimensional variable search
problem so as to find out the root causes of yield losses.
However, if a specific stage contains merely one device for
processing, the OGA and LASSO regression models will treat
this stage as a constant term. As a result, this stage will not be
selected even though the root cause of yield losses is in the stage.
This is the limit of the KSA scheme. How to handle this blind-
stage issue and include the resolution into the KSA scheme will
be the future work.

ACKNOWLEDGMENT
Fig. 13. Root cause analysis of control voltage on chamber A of equipment A. The authors would like to thank AUO in Taiwan for providing
the raw data used in the illustrative example. This work has also
filed a U.S. provisional patent application under application no.:
that another 3 out of 8 Type 2 Loss samples were processed 62/260,656 on November 30, 2015.
by the Top 2 device. To find out the root causes, Step 3 of
Fig. 8 should be performed by inputting Y and XP into the REFERENCES
KSA scheme.
The Top 2 device is selected for illustration. The process data [1] R. Vernon, “International investment and international trade in the product
cycle,” Quart. J. Econ., vol. 80, pp. 190–207, 1966.
(XP ) of the Top 2 device has 27 variables. After conducting [2] A. Chen and A. Hong, “Sample-efficient regression trees (SERT) for
the KSA analysis on all the devices of the same stage to which semiconductor yield loss analysis,” IEEE Trans. Semicond. Manuf.,
the Top 2 device belongs, its RIK value of XP search is 0.864 vol. 23, no. 3, pp. 358–369, Aug. 2010.
[3] C. F. Chien, W. C. Wang, and J. C. Cheng, “Data mining for yield en-
(>0.7) as shown in Fig. 12. Therefore, the search result is reli- hancement in semiconductor manufacturing and an empirical study,” Exp.
able with the Top 1 variable being Control Voltage. Syst. Appl., vol. 33, pp. 192–198, 2007.
To confirm that “Control Voltage” is the root cause, the “Con- [4] C. Y. Hsu, C. F. Chien, K. Y. Lin, and C. Y. Chien, “Data mining for
yield enhancement in TFT-LCD manufacturing: An empirical study,”
trol Voltage” values of all the five chambers (A-E) of Equipment J. Chinese Inst. Ind. Eng., vol. 27, no. 2, pp. 140–156, 2010.
A are drawn and verified by a box plot chart and the hypothe- [5] C.-K. Ing and T. L. Lai, “A stepwise regression method and consistent
sis test is executed as shown in Fig. 13. It shows that p-value model selection for high-dimensional sparse linear models,” Statistica
Sinica, vol. 21, pp. 1473–1513, 2011.
(=0.002) is less than 0.05, which indicates that Chamber A’s [6] R. Tibshirani, “Regression shrinkage and selection via the LASSO,”
“Control Voltage” value in red circle is indeed less than those J. Roy. Statist. Soc. B, vol. 58, no. 1, pp. 267–288, 1996.
of the other chambers. As such, the root cause of this Type 2 [7] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical
Learning: Data Mining, Inference, and Prediction. New York, NY, USA:
Loss is due to the abnormality found in Chamber A’s “Control Springer, 2009.
Voltage”. [8] F.-T. Cheng, C.-A. Kao, C.-F. Chen, and W.-H. Tsai, “Tutorial on applying
The KSA scheme was implemented in a computer with i5- the VM technology for TFT-LCD manufacturing,” IEEE Trans. Semicond.
Manuf., vol. 28, no. 1, pp. 55–69, Feb. 2015.
[email protected] GHz CPU, 6.0 GB RAM, and 1 TB hard disk. With [9] L. Fowler and T. Davis, “Engineering data analysis using discovery,” in
the illustrative example mentioned above, the execution times Proc. IEEE/SEMl Adv. Semicond. Manuf. Conf., Nov. 1996, pp 416–422.
of the Steps 2 (for searching the key stages) and 3 (for selecting [10] Y.-T. Huang and F.-T. Cheng, “Automatic data quality evaluation for the
AVM system,” IEEE Trans. Semicond. Manuf., vol. 24, no. 3, pp. 445–454,
the key variables in the most suspicious stage) shown in Fig. 8 Aug. 2011.
are 1.22 and 1.36 sec, respectively. [11] F.-T. Cheng, H.-C. Huang, and C.-A. Kao, “Developing an automatic
virtual metrology system,” IEEE Trans. Autom. Sci. Eng., vol. 9, no. 1,
pp. 181–188, Jan. 2012.
[12] F.-T. Cheng, Y.-T. Chen, Y.-C. Su, and D.-L. Zeng, “Evaluating reliance
VI. SUMMARY AND CONCLUSION level of a virtual metrology system,” IEEE Trans. Semicond. Manuf.,
A high-dimensional Key-variable Search Algorithm (KSA) vol. 21, no. 1, pp. 92–103, Feb. 2008.
[13] D. Port, R. Kazman, H. Nakao, and M. Katahira, “Practicing
for yield improvement is proposed in this paper. A two-phase what is preached: 80/20 rules for strategic IV & V assessment,”
process is suggested to apply the KSA scheme for searching the in Proc. IEEE Int. Conf. Exploring Quantifiable IT Yields, 2007,
root causes of yield losses: pp. 45–54.
[14] M. Singson and P. Hangsing, “Implication of 80/20 rule in electronic
1) Phase I: Feed yield test results (Y) as well as total- journal usage of UGC-infonet consortia,” J. Academic Librarianship,
inspection inline data (y) and/or production routes (XR ) vol. 41, no. 2, pp. 207–219, 2015.