Professional Documents
Culture Documents
4, NOVEMBER 2016
411
I. I NTRODUCTION
OWADAYS, in order to maintain the competitive
advantage, many companies strive to meet customers
demands in terms of quantity, quality, and delivery accuracy.
To achieve this, low cycle time and high throughput are both
necessary. Unfortunately, these performance goals are known
as conflicting to each other and tradeoffs have to be made
in reality. Product mix, technically defined as the fraction of
each product in a manufacturing process, has large impacts
on both cycle time and throughput due to blockages and idleness of equipments or resources in the production or assembly
lines. When the manufacturing process is complex, it is a difficult task to estimate the cycle time and throughput under
different scenarios of product mix. The determination of the
Manuscript received May 1, 2016; accepted July 9, 2016. Date of publication August 24, 2016; date of current version October 27, 2016. This work was
supported in part by the Advanced Manufacturing and Service Management
Center, National Tsing Hua University and in part by the Ministry of Science
and Technology of Taiwan under Grant MOST101-2628-E-007-010-MY3.
The author is with the Industrial Engineering and Engineering
Manufacturing, National Tsing Hua University, Hsinchu 30013, Taiwan
(e-mail: chang@mx.nthu.edu.tw).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSM.2016.2602388
c 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
0894-6507
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
412
approach based on queueing theory was developed for estimating the relationship between cycle time and throughput [11].
In [9], a simple queueing model was developed to gain an
insight into the relationship among cycle time, lot size, and
the fab technology level. In summary, the analytical models are
good at providing useful information about the manufacturing
system under study and are relatively easy to use.
On the other hand, simulation is a widely-used tool
for estimating cycle time due to the ability to account
for all the technical details essential to the manufacturing process. Applications of simulation for evaluating
cycle time and throughput in semiconductor manufacturing
include [1], [12][16]. In particular, in [12], simulation was
utilized to assess the impact of scheduling on the mean cycle
time of wafer lots in semiconductor manufacturing. The impact
of HLs on the cycle time of other lots via simulation was investigated in [13]. The results showed that as the fraction of HLs
in the WIP increases, both the mean cycle time and the corresponding standard deviation for all other lots increase dramatically. In [15], an aggregate simulation model representation of
workstations with integrated processing equipment was developed. In particular, a curve fitting procedure was introduced to
deal with the typically limited number of arrival and departures
encountered in semiconductor manufacturing practice. In [1], a
simulation-based methodology was proposed for quick development for the relationship between cycle time and product
mix in semiconductor manufacturing. Some empirical analysis has also been conducted for investigating the relationship
between volume and product mix flexibility [17]. Superior to
the analytical models, simulation models can accommodate
all the technical details that are important to the manufacturing process. However, its disadvantage is that it may require a
substantial amount of time to produce statistically valid results.
III. T HE P ROPOSED M ODEL
In semiconductor industry, engineering lots are implemented
not only to enable the detection of problematic equipments,
but also to provide marketing and business with extra flexibility regarding delivery lead times. To enable the control of
the risk associated with the selected product mix, we propose
a quantile-based simulation optimization model. Three types
of products in the manufacturing process are considered, i.e.,
RLs, HLs, and NLs. Let x = [x1 , x2 , x3 ] be the vector of product mix where each component corresponds to the fraction of
RLs, HLs and NLs, respectively. Further, let the response variable of interest be cycle time (CT), so GCT,1 (x, ) be CT of
RLs and GCT,3 (x, ) be CT of NLs, where denotes the randomness involved in the manufacturing system and Cnl is a
prescribed value. Mathematically, the product mix model can
be expressed as follows.
The Quantile-Based Product Mix Model:
Minx 1 GCT,1 (x, )
s.t. 2 GCT,3 (x, ) Cnl
3
xi = 1
i=1
xi 0,
for all i.
(1)
The above model aims at seeking the product mix that can
result in minimized 1 th quantile of CT of RLs subject to
that 2 th CT of NLs is not greater than a prescribed value
Cnl , where the values of 1 and 2 are determined by production managers. Due to the complexity of semiconductor
manufacturing, 2 [GCT,1 (x, )] and 2 [GCT,3 (x, )] are not
analytically available but can be estimated only by simulation, which inevitably contains noise. This in turn imposes
challenges to telling whether or not a solution is feasible and
one solution is superior to another. Seeking the optimal solution(s) of the proposed model is even more challenging. In
Section IV, we will present a solution method that possesses
an efficient search strategy to identify the optimal solution.
IV. S OLUTION M ETHOD
AGLS-QC is essentially a direct search method. The advantage of the direct search method is that it does not require
gradient estimation in the search process. Instead, it uses only
the (noisy) objective function values and determines the moving direction by comparing a set of solutions. As discussed
in [18], another advantage of the direct search method is
that it is computationally efficient when the feasible region
is not large,
which is the case in this research due to the
constraint 3i=1 xi = 1. In [3], a direct search method, called
SNM-Q, is proposed to solve the unconstrained quantile-based
simulation optimization problems. In [19], SNM-Q is further
extended, called SNM-QC, to solve the constrained quantilebased simulation optimization problems where the constraints
are linear and analytically known. This contrasts with the
proposed model in Section III where the constraint is quantilebased, not analytically available, and has to be estimated via
simulation.
AGLS-QC incorporates the quantile estimation techniques
and the penalty function method from [19] but further employs
an efficient search strategy based on a novel neighborhood
structure to enable quick identification of the optimal solution. The idea is to sequentially divide the feasible region into
several subregions that encompass solutions that have close
objective function values, and then identify the most promising region where the truly optimal solution is likely to reside
in. Because the algorithm samples a region rather than a
solution in each iteration, the search process for the optimal solution is significantly accelerated. In what follows, we
first present the main framework of AGLS-QC, followed by
the introduction of the quantile estimation technique and the
penalty function method.
A. Main Framework
For each iteration, AGLS-QC performs both global sampling and local sampling for a new solution with a nice balance
between exploring the whole feasible region (i.e., global sampling) and exploiting the local region (i.e., local sampling).
Specifically, for any iteration k, let the whole feasible region
be X. And (x) be the quantile estimate at the solution x.
Further let Q(x) = k (x) + rk P(x), where Q(x) is the Q
value evaluated at x, P(x) is the penalty function that has
a large value when the solution is infeasible, and rk is the
Fig. 1.
413
penalty parameter at iteration k, a positive real number satisfying limk rk = . Let 0 = be an empty set and k the
sampled-solution set that collects all sampled solutions up to
iteration k. Let the most promising region be the neighborhood
of xk , which is the solution that has the minimum Q value in
k .
For any iteration k, AGLS-QC conducts the following four
steps.
Step1. Perform local sampling to select t solutions in the
most promising region and update k .
Step2. Perform global sampling to select s solutions in X
and update k .
Step3. Define/update the neighborhood of each solution
in k . Estimate and compare the Q value of all
solutions in k . Identify the the most promising
region.
Step4. If the termination criterion is satisfied, return the
solution xk . Otherwise, let k = k + 1 and return to
Step 1.
The neighborhood of each sampled solution is defined as
follows. Let Dist(u,v) be the Euclidean distance between
the two solutions u, v. The neighborhood of solution xi is
defined as
(xi ) = x : Dist(x, xi ) min Dist xi , xj , j = i ,
(2)
which represents the region defined in a way that any neighborhoods of two sampled solutions will not overlap with each
other. It is remarkable that the neighborhood of one solution,
say xi , is defined dependent on its neighboring solutions and
their relative locations; both are iteratively changed. The values of t and s are selected by users according to the problem
dimension. For example, for a problem of low dimension, the
value of t and s may not be selected large; on the other hand,
when the problem dimension is large, a large value of t and s
are suggested in order to accelerate the convergence speed of
the algorithm.
As shown in Fig. 1, suppose 10 solutions are sampled
through global and local sampling. The algorithm defines the
neighborhood of each solution and identifies the neighborhood
of the solution with minimum quantile estimate as the most
promising region, for example, the neighborhood of x3 . It can
be easily seen that as the algorithm continues to iterate, more
solutions are sampled, making the the most promising region
smaller, in turn increasing the probability that the solutions in
the most promising region being sampled. In the meantime,
due to more solutions are sampled in the feasible region, the
414
(3)
Let X
= {x X : +
}, which denotes
is set as Nk =
k, rk = 105 k0.1 For the four algorithms
under comparison, we do not adjust their parameter setting in
order to achieve the best performance. Instead, the parameter setting is chosen merely based on an educated guess.
As it is known that the parameter setting can significantly
affect the algorithm performance, this numerical study is not
intended to conclude that one algorithm is dominantly superior to the another. Instead, this numerical study aims to shed
some light into how the algorithm performance might be when
it is applied to solve a set of numerical problems.
The performance measure is defined by
xk
,
(4)
(x0 )
where xk and correspond to the best solution found when
the algorithm is terminated and the true globally optimal function value, respectively. This quantity measures how much the
objective function value has been reduced by the algorithm
when it is terminated. Therefore, when it is close to zero, it
means the algorithm is around the optimum; on the other hand,
when it is close to one, the algorithm is far from the optimum
or even diverges.
Due to the consideration that users may have limited knowledge about the optimal solution, the initial solution x0 is
415
TABLE I
P ERFORMANCE C OMPARISON W HEN X N ORMAL(0, g(X)), = 0.9
TABLE II
P ERFORMANCE C OMPARISON W HEN X N ORMAL(0, g(X)), = 0.8
TABLE III
P ERFORMANCE C OMPARISON W HEN X E XPONENTIAL( = g(X)), = 0.9
TABLE IV
P ERFORMANCE C OMPARISON W HEN X E XPONENTIAL( = g(X)), = 0.8
416
TABLE VI
O PTIMAL P RODUCT M IX BY AGLS-QC
TABLE V
E XAMPLES OF ROUTE S HEET
xi = 1.
(5)
i=1
The above model is to seek the product mix that can achieve
0.9-quantile of CT of RLs, while ensuring that the 0.8-quantile
of CT of NLs can be less than 65. Table VI provides the
optimal product mix solved by AGLS-QC. It is suggested that
the fractions of RLs, HLs and NLs, should 0.12, 0.2 and 0.68,
respectively.
This optimal product mix is then implemented in the wafer
fab as follows. Based on the wafer start in the fab, the optimal product mix can be converted into the quantity of wafers
of RLs, HLs, and NLs that would be released into the manufacturing system. It is remarkable that since the cycle times
of RLs, HLs and NLs are different, the fraction of RL, HL
and NL in the WIP may change with time. Therefore, production engineers typically monitor the product mix in the WIP
and determine how many more wafers of RLs, HLs and NLs
should be released into the manufacturing system on a daily
basis in order to stick to the optimal product mix.
After the suggested product mix is implemented, it is found
that, compared to the current practice, the 0.9-quantile of CT
of RLs is decrased by 28% and the constraint of CT of NLs
is also satisfied. This result shows that the proposed model
and the solution method can truly result in the product mix
that has satisfactory performance in practice, thus validating
the viability of this research in practical settings.
VII. C ONCLUSION
In this paper, we proposed a quantile-based simulation optimization model to characterize the risk-controlled product mix
planning in semiconductor manufacturing and applied a solution method, AGLS-QC, to obtain the optimal product mix in
semiconductor manufacturing with risks taken into consideration. As opposed to the existing analytical models that require
assumptions to obtain mathematical tractability, the proposed
model and the solution method use simulation as a tool to
evaluate the system performance where all the important manufacturing details are allowed to be taken into consideration.
Moreover, the novel quantile-based formulation allows for the
control of the upside risk associated with the selection of
the product mix. An extensive numerical study shows that
the performance AGLS-QC is promising and is more efficient
compared to three existing heuristics. Finally, an empirical
study conducted in collaboration with a leading semiconductor
min 50
x
+ 47.5x5 + 17 + x
s.t. [20x1 + 12x2 + 11x3 + 7x4 + 4x5 + x ] 50
5
xi = 1
i=1
xi 0, i = 1, . . . , 5.
5
(6)
i=1
xi 0, i = 1, . . . , 6
(7)
Problem 3:
2
2
min
10 xi+1 xi2
+ (1 xi )2
2
2
90 xi+3 xi+2
+ (1 xi+2 )2
2
+
10(xi+1 + xi+3 2)
2
+ 1/ 10(xi+1 xi+3 ) + x
+
Wi G(i)
(10)
xi = 1
i=1
xi 0, i = 1, . . . , 6.
(11)
s.t. 127 + 2x12 + 3x24 + x3 + 4x44 + 5x5 + x 0
5
n
i=1
i=1
i=1
(x) = N 1
For any fixed x, let G(1) G(2) . . . G(n) be order statistics obtained by sorting the observations {Gi , i = 1, 2, . . . , n}
in an ascending order. One traditional quantile estimator is:
G[n]
if = 0
1 =
(9)
if > 0
G[n]+1
i=1
i=1
min 0.5
B. Quantile Estimation
2 =
Problem 1:
Problem 2:
417
(8)
418
[5] Y. Narahari and L. M. Khan, Modeling the effect of hot lots in semiconductor manufacturing systems, IEEE Trans. Semicond. Manuf., vol. 10,
no. 1, pp. 185188, Feb. 1997.
[6] S.-H. Chung and H.-W. Huang, Cycle time estimation for wafer fab
with engineering lots, IIE Trans., vol. 34, no. 2, pp. 105118, 2002.
[7] S. H. Chung, W. L. Pearn, A. H. I. Lee, and W. T. Ke, Job order releasing and throughput planning for multi-priority orders in wafer fabs, Int.
J. Prod. Res., vol. 41, no. 8, pp. 17651784, 2003.
[8] J. G. Shanthikumar, S. Ding, and M. T. Zhang, Queueing theory for
semiconductor manufacturing systems: A survey and open problems,
IEEE Trans. Autom. Sci. Eng., vol. 4, no. 4, pp. 513522, Oct. 2007.
[9] E. Zarifoglu, J. J. Hasenbein, and E. Kutanoglu, Lot size management
in the semiconductor industry: Queueing analysis for cycle time optimization, IEEE Trans. Semicond. Manuf., vol. 26, no. 1, pp. 9299,
Feb. 2013.
[10] A. A. Kalir, Segregating preventive maintenance work for cycle
time optimization, IEEE Trans. Semicond. Manuf., vol. 26, no. 1,
pp. 125131, Feb. 2013.
[11] F. Yang, B. Ankenman, and B. L. Nelson, Efficient generation of cycle
time-throughput curves through simulation and metamodeling, Naval
Res. Logistics, vol. 54, no. 1, pp. 7893, 2007.
[12] L. M. Wein, Scheduling semiconductor wafer fabrication, IEEE Trans.
Semicond. Manuf., vol. 1, no. 3, pp. 115130, Aug. 1988.
[13] B. Ehteshami, R. G. Petrakian, and P. M. Shabe, Trade-offs in cycle
time management: Hot lots, IEEE Trans. Semicond. Manuf., vol. 5,
no. 2, pp. 101106, May 1992.
[14] A. I. Sivakumar and C. S. Chong, A simulation based analysis
of cycle time distribution, and throughput in semiconductor backend
manufacturing, Comput. Ind., vol. 45, no. 1, pp. 5978, 2001.
[15] C. P. L. Veeger, L. F. P. Etman, J. V. Herk, and J. E. Rooda,
Generating cycle time-throughput curves using effective process time
based aggregate modeling, IEEE Trans. Semicond. Manuf., vol. 23,
no. 4, pp. 517526, Nov. 2010.
[16] K. Crist and R. Uzsoy, Prioritising production and engineering lots
in wafer fabrication facilities: A simulation study, Int. J. Prod. Res.,
vol. 49, no. 11, pp. 31053125, 2011.
[17] M. Hallgren and J. Olhager, Flexibility configurations: Empirical analysis of volume and product mix flexibility, Omega, vol. 37, no. 4,
pp. 746756, 2009.
[18] K.-H. Chang, Stochastic NelderMead simplex methodA new globally convergent direct search method for simulation optimization, Eur.
J. Oper. Res., vol. 220, no. 3, pp. 680694, 2012.
[19] K.-H. Chang and H.-K. Lu, Quantile-based simulation optimization
with inequality constraints: Methodology and applications, IEEE Trans.
Autom. Sci. Eng., vol. 13, no. 2, pp. 701708, Apr. 2016.
[20] K.-H. Chang, A quantile-based simulation optimization model for sizing hybrid renewable energy systems, Simulat. Model. Pract. Theory,
vol. 66, pp. 94103, Aug. 2016.
[21] C. A. Floudas and P. M. Pardalos, A Collection of Test Problems
for Constrained Global Optimization Algorithms, vol. 455. Berlin,
Germany: Springer-Verlag, 1990.
[22] H.-P. Schwefel, Evolution and Optimum Seeking: The Sixth Generation.
New York, NY, USA: Wiley, 1993.
[23] R. J. Serfling, Approximation Theorems of Mathematical Statistics.
New York, NY, USA: Wiley, 1980.
[24] F. E. Harrell and C. E. Davis, A new distribution-free quantile
estimator, Biometrika, vol. 69, no. 3, pp. 635640, 1982.
[25] T. Dielman, C. Lowry, and R. Pfaffenberger, A comparison of quantile estimators, Commun. Stat. Simulat. Comput., vol. 23, no. 2,
pp. 355371, 1994.