Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Nuclear Engineering and Design 295 (2015) 84–95

Contents lists available at ScienceDirect

Nuclear Engineering and Design


journal homepage: www.elsevier.com/locate/nucengdes

Core map generation for the ITU TRIGA Mark II research reactor using
Genetic Algorithm coupled with Monte Carlo method
Mehmet Türkmen a,∗ , Üner Çolak b , Şule Ergün a
a
Nuclear Engineering Department, Hacettepe University, Beytepe Campus, Ankara, Turkey
b
Energy Institute, Istanbul Technical University, Ayazağa Campus, Maslak, Istanbul, Turkey

h i g h l i g h t s

• Optimum core maps were generated for the ITU TRIGA Mark II Research Reactor.
• Calculations were performed using a Monte Carlo based reactor physics code, MCNP.
• Single-Objective and Multi-Objective Genetic Algorithms were used for the optimization.
• keff and ppfmax were considered as the optimization objectives.
• The generated core maps were compared with the fresh core map.

a r t i c l e i n f o a b s t r a c t

Article history: The main purpose of this study is to present the results of Core Map (CM) generation calculations for the
Received 10 April 2015 İstanbul Technical University TRIGA Mark II Research Reactor by using Genetic Algorithms (GA) coupled
Received in revised form 12 August 2015 with a Monte Carlo (MC) based-particle transport code. Optimization problems under consideration
Accepted 15 August 2015
are: (i) maximization of the core excess reactivity (ex ) using Single-Objective GA when the burned fuel
Available online 19 October 2015
elements with no fresh fuel elements are used, (ii) maximization of the ex and minimization of maximum
power peaking factor (ppfmax ) using Multi-Objective GA when the burned fuels with fresh fuels are used.
The results were obtained when all the control rods are fully withdrawn. ex and ppfmax values of the
produced best CMs were provided. Core-averaged neutron spectrum, and variation of neutron fluxes with
respect to radial distance were presented for the best CMs. The results show that it is possible to find an
optimum CM with an excess reactivity of 1.17 when the burned fuels are used. In the case of a mix of
burned fuels and fresh fuels, the best pattern has an excess reactivity of 1.19 with a maximum peaking
factor of 1.4843. In addition, when compared with the fresh CM, the thermal fluxes of the generated CMs
decrease by about 2% while change in the fast fluxes is about 1%.Classification: J. Core physics
© 2015 Elsevier B.V. All rights reserved.

1. Introduction criteria on the maximum power peaking and the maximum excess
reactivity strictly hinder the use of a random Core Map (CM). For
In reactor operations, at the end of each fuel cycle, either some of this reason, all the CMs, formed by whether replacement with the
the burned fuel elements are replaced with fresh fuels (partial core fresh fuels or the shuffling, need optimization.
replacement) or burned fuels are shuffled. For research reactors, The idea of optimization begins with a desire to find the best
fuel shuffling process is very important due to a fuel shortage prob- configuration through minimization (or maximization); however,
lem which is currently emerged and due to the need for sustainable it also comes with many challenges to overcome. The most impor-
operations without long-time interruptions which demand certain tant one is that a typical minimization problem contains many local
requirements (e.g., a definite irradiation period, specific neutron minima and only one global minimum. This implies that there is
flux values at the specific locations and a flexible core map). The always a chance that a solver, which is designed to find the global
most important challenge in performing the fuel shuffling or par- minimum, may fall into a local minimum. Moreover, most engi-
tial core replacement is the preservation of the safety limits. Safety neering problems have more than one (multiple) objective to be
minimized, and these objectives require optimization all at once.
Also, the size of the solution space is a concern on the optimization
∗ Corresponding author. Tel.: +90 312 297 73 00; fax: +90 312 299 21 22. and it is directly related to the count of used variables. More variable
E-mail address: tm@hacettepe.edu.tr (M. Türkmen). means the larger solution space to be examined. Searching large

http://dx.doi.org/10.1016/j.nucengdes.2015.08.022
0029-5493/© 2015 Elsevier B.V. All rights reserved.
M. Türkmen et al. / Nuclear Engineering and Design 295 (2015) 84–95 85

solution spaces requires a vast amount of time and lots of effort This study is considered as part of an ongoing extensive neu-
even for relatively small problems like Traveling Salesman Prob- tronic research to use the computer model of the research reactor
lem (TSP). In order to decrease the computational time, to maximize in medical/industrial applications. Validity of the reactor model
the solution performance and reliability, and to minimize the cost and burnup calculation of the fuel elements have been successfully
output (function evaluation), so far, many theories, approaches and shown in previous separate studies (Türkmen et al., 2015; Türkmen
analytic methods such as Genetic Algorithm (GA) (Goldberg, 1989), and Çolak, 2013, 2014a,b). In this work, fuel shuffling studies were
Pattern Search (Hooke and Jeeves, 1961) and Simulated Annealing performed for the İstanbul Technical University (ITU) TRIGA Mark
Algorithm (Stevens et al., 1995) have been developed. II Research and Training reactor by using the results of the previous
The type of the problem determines which solution method(s) studies. The considered case studies for the shuffling are: (i) use of
is the most suitable for optimization analyses. For example, the the burned fuels with no fresh fuels and (ii) a mix of burned fuels
stochastic methods are superior, not the best, in the area of in-core and fresh fuels. For this purpose, Single-Objective GA (SOGA) and
nuclear fuel management calculations than the mathematical opti- Multi-Objective GA (MOGA) were employed. Core excess reactivity
mization methods like (non)linear and quadratic, and work well (ex ) and maximum power peaking factor (ppfmax ) were considered
for the pattern search problems (Turinsky et al., 2005). Among the as the optimization objectives. The SOGA is used in the first case to
optimization methods, the GA is very attractive since it imitates the search a maximum value for the ex while the MOGA is utilized
way the nature does (behaves) to get more diverse and better gen- in the second case to find out a maximum ex value considering a
erations. Although the GA concept (Holland, 1975) was emerged safety constraint on the ppfmax . The data (e.g., burnup value, spent
in the late 1960s, Alan Turing (Turing, 1950) lit the GA fire with a fuel content) used in the fuel shuffling calculations were obtained
question of “Can machines think?”. More than half a century, the from the results of the burnup analyses that were presented in a
GA is used to solve the various complex engineering problems. Still, different study (Türkmen et al., 2015). The best CMs, which would
there is a growing endeavor to make the GA better. For this purpose, possibly extend the reactor cycle life-time and maximize the fuel
a number of innovative methods such as VEGA (Schaffer, 1985), utilization, were determined.
NSGA (Srinivas and Deb, 1994), SPEA (Zitzler and Thiele, 1999) and
DMOEA (Yen and Lu, 2003) have been suggested to cope with the
various complicated problems. 2. Shuffling calculations and application of the Genetic
In the literature, the GA is frequently used to optimize CMs for Algorithm
different types of nuclear power reactors such as PWR (Zameer
et al., 2014), AHWR (Thakur et al., 2013), BWR (François et al., 2013), 2.1. Problem description
and VVER (Karahroudi et al., 2013). Yet, several noteworthy stud-
ies outlined below have been published within 10 years to find the The primary purpose of this study is to produce a set of CMs for
optimum CMs for the research reactors. the ITU TRIGA Mark II research reactor which would maximize the
To find the optimum CMs for the research reactors, optimal fuel fuel utilization and extend the fuel cycle length. The burned and
loading strategies were searched for TRIGA Mark II reactor at AERE, fresh fuel elements are considered for the CM generation. Using
Savar (Huda et al., 2008). Although the fuel elements were shuffled these fuel elements, generation of a CM can be achieved in three
according to the out-in fuel management strategy to obtain a better different ways. The first is the use of the burned fuels without using
CM, an optimum CM was not investigated by taking the maximum fresh fuels. The idea stands on changing the locations of the burned
fuel utilization into account. Instead, a simple fuel shuffling was fuel elements inside the core. Typically, the fuel rods located at the
performed after the core reached a nominal condition (decided by rings are shifted to an inner ring while the fuels rods located at the
the researcher) for its life. In the study, the results were obtained most-inner ring are shifted to the most-outer ring. This method,
by using TRIGAP, MVP-BURN, and MCNP4C-ORIGEN2.1. The results called out-in in-core fuel management, is very fast since there
of the study showed that extension of reactor core lifetime to some is only one CM to be prepared and to be searched, and efficient
degree was possible by finding an optimum burnup value for the since a certain extension in the fuel cycle length to some degree
reshuffling. A similar study was carried out by Lyric et al. (2013) for is expected; however, it is not functional to obtain an optimum
the BAEC TRIGA reactor. To obtain the optimum burnup, different CM. Another solution is to shuffle the burned fuel rods in a random
fuel loading schemes were investigated from the viewpoint of core way and then to choose the best one from among them. However,
life extension, neutron economy, and reactor safety. According to this method has some disadvantages. The most crucial one is the
the study, the most preferable scheme, which meets the require- decision on the number of CMs to be searched. Also, this method
ments, was to fill the core with the burned fuel elements along is not efficient as much as out-in strategy due to the fact that all of
with the ten fresh fuel elements in each cycle after an initial cycle the generated CMs have to be examined one by one. Yet, there is a
containing five fresh fuel elements. chance, even if infinitesimal, to discover an optimum CM. Second is
Do and Nguyen (2007) proposed optimum refueling schemes the use of a mix of burned fuels and fresh fuels. In this instance, the
for the TRIGA Mark II reactor of Vietnam by using Single-Objective arrangement of the fuel rods inside the core is very different from
Genetic Algorithm (SOGA) and Multi-Objective Genetic Algorithm the first case described above. As an example, fresh fuels can be
(MOGA). Using the reactor physics codes of CITATION and WIMSD, inserted into the specific locations such as to the most-outer/inner
it was found that most of the fresh fuels had to be located around ring to create a low-leakage core, and close to irradiation channels
the core center for the optimal low-leakage CM. Mazrou and to obtain maximum neutron flux. Third is the use of fresh fuels with
Hamadouche (2006) investigated optimal core configurations for no burned fuels. This is the easiest one if all the fuels to be used are
the 10 MW IAEA LEU research reactor by combining a stochastic completely identical. In all cases except for the last one, it is pos-
method, Simulated Annealing and an adaptive back-propagation sible to generate countless unique CMs using the mentioned fuel
neural network. Hedayat et al. (2009) published the results of types. In addition to enormous number of CMs, the produced CMs
optimum core configuration calculations for a reference research completely depend on the user experiences/knowledge. Yet, still,
reactor by using a hybrid algorithm (MOGA linked with the cascade the main problem is a desire for a particular CM which satisfies all
feed forward artificial neutral networks). Later, a refueling opti- the requirements including safety constraints. In this manner, this
mization problem was solved for the objectives of maximum cycle study handles first two cases and aims to obtain a set of optimum
length and maximum irradiating thermal neutron flux (Hedayat, CMs for the considered research reactor. The case studies consid-
2014). ered in this study for the core shuffling are: (i) use of the burned fuel
86 M. Türkmen et al. / Nuclear Engineering and Design 295 (2015) 84–95

F30 F1 F2
F29 35 34 34
PS
36
F28 F4 Element Number
E24 E1 E2
37 34 A1
E23 25 24 25
F27 E3 F5 MT
38 28 Material Number
D1 25 33
E22 D18 E4
14 D2
30 17 22 F6
NS D17 12
D3 32 GR : Graphite Fuel Element
E21 19 C1 13
RR NS : Neutron Source
29 C12 5 C2 F7
D4 CT : Central Thimble
GR TR 10 4 33
17 PS : Pneumatic Transfer System
E20 C11 B1 C3 E6 RR : Regulating Rod
32 13 2 8 22 F8 SR : Safety Rod
D15 B6 B2 D5
GR 16 34 TR : Transient Rod
22 5 1
E19 C10 C4 E7
CT
33 15 9 26 F9
D14 B5 B3 D6
GR 36
22 7 3 18
E18 C9 B4 C5 E8
34 17 6 11 27
GR D13 C8 C6 D7 GR
25 18 13 20
E17 E9
D12 SR
34 D8 30
GR 25 21 GR
E16 D11 D9 E10
35 25 D10 23 31
E15 22
GR E11 GR
35 E14 32
E12
GR 35 E13
33 GR
34
GR GR
GR GR
GR

Fig. 1. ITU TRIGA Mark II research reactor core map at EOL used for the optimization analyses. (For interpretation of the references to color in this figure legend, the reader
is referred to the web version of this article.)

elements with no fresh fuels, (ii) use of the burned fuel elements fuel elements used in shuffling are discharged from the reactor
with the fresh fuels. core without a cooling period. The locations in the CM shown in
To obtain optimum CMs free from the user dependency, as a Fig. 1 are randomly filled by the burned fuel elements. All the data
search method, two different methods of the GA, SOGA and MOGA, used in the MC calculations (burnup data and spent fuel content
are utilized. The details on the application of the GA and the case of each fuel element) were obtained from the results of a pre-
studies are given in the following sections. The latest CM of the vious study (Türkmen et al., 2015). The results of another study
ITU TRIGA Mark II reactor used for the optimization analyses is (Türkmen and Çolak, 2014b) show that the ppfmax is less than the
shown in Fig. 1. Initially, all the fuel elements containing 8.5% U safety limit of 1.6 when all the fuel elements are fresh. Using this
enriched with 20% 235 U with no Erbium content are identical. As information, it is assumed that the ppfmax would never exceed
the reactor operates, isotopic content of each fuel element changes the safety limit when only the burned fuels are used to generate
depending on the position inside the core. Material distribution was new CMs. Therefore, in this part of the study, only one CM which
obtained for the End of Life (EOL) condition by performing the bur- gives the highest ex value was investigated by using the SOGA
nup calculation (Türkmen et al., 2015). While at the Beginning of method.
Life (BOL) initial excess reactivity of the given CM was obtained to
be about 1.4 $, at EOL excess reactivity was zero. In the figure, the
burned fuel elements are enumerated from 1 (the most-burned) to
2.2.2. Shuffling the burned fuel elements with fresh fuels
38 (the least-burned) since 38 different spent fuel materials were
This case study focus on the use of burned fuel
obtained in the burnup calculations. The burnup values of the used
s along with the fresh fuels to obtain an optimum CM(s). A
fuel elements are given in Table 1. The fuel elements with the same
certain number of the most-burned fuel elements were removed
material number in the CM are the same materials and completely
from the core and those were replaced with the fresh ones. The
identical from the viewpoint of the burnup data and spent fuel con-
total number of fuel element used in the CM generation is 69
tent. There are 69 locations inside the core to be loaded with the
composed of 59 burned fuels and 10 fresh fuels which are available
fuel elements. During CM generation, all the positions are filled by
for loading in the reactor storage. The fresh and burned fuels
the burned/fresh fuel elements. Fresh fuel elements are numbered
randomly fill the fuel element locations in the CM. With a sound
as “0” in the CM. Fresh CM means that the reactor is fully loaded
judgement, using fresh fuels may cause exceeding safety limits
with the fresh fuels.
at the locations where the fresh fuels are placed. The reason is
that the fresh fuels have higher excess reactivity than the burned
2.2. Case studies for shuffling fuels due to containing higher 235 U content and no fission product
poisoners. Thus, ex and ppfmax parameters were optimized jointly
2.2.1. Shuffling the burned fuel elements with no fresh fuels by taking the safety constraint on the power peaking factor into
In the first case, extension of the core cycle length without using account. A set of CMs which give the highest ex value with ppfmax
the fresh fuels was aimed. The total number of the burned fuel value lower than the limiting value were investigated by using the
elements used in the CM generation is 69. It is assumed that the MOGA.
M. Türkmen et al. / Nuclear Engineering and Design 295 (2015) 84–95 87

Table 1
Burnup values of the used fuel elements.

Material number Burnup (%) Material number Burnup (%) Material number Burnup (%) Material number Burnup (%)

0 0 10 1.34 20 1.11 30 0.90


1 1.51 11 1.28 21 1.06 31 0.88
2 1.50 12 1.25 22 1.02 32 0.83
3 1.45 13 1.25 23 1.02 33 0.79
4 1.42 14 1.23 24 1.01 34 0.78
5 1.40 15 1.22 25 1.00 35 0.76
6 1.39 16 1.20 26 0.98 36 0.72
7 1.39 17 1.20 27 0.96 37 0.63
8 1.36 18 1.19 28 0.95 38 0.63
9 1.36 19 1.14 29 0.92

2.3. Application of the Genetic Algorithm (GA) Sampling Method (SUS) (Baker, 1987) is used in the elitism process
to pass the best solutions to the next generations and to maintain
Genetic Algorithm is an optimization method that uses natural the diversity of the population while the Binary Tournament, two
evolution processes including inheritance, immigration, mutation, individuals are chosen at random and the better of the two individ-
selection and crossover operators. It uses chromosomes to define uals is selected, is used in the selection process to determine the
the candidate solutions of a problem. It evolves the candidate solu- parents of offspring for crossover and mutation operators. In the
tions toward better solutions. Also, the heuristic search in applying case of MOGA, NSGA-II (Deb et al., 2002) with the Binary Tourna-
GA finds the effective solutions for the specific types of the opti- ment is used for the selection and elitism processes. A continuous
mization problems. Although it is an effective method, there is no GA flowchart given in Fig. 2 shows the computational main steps
guarantee to solve every optimization problem. Typically, there are performed. The terms used in the GA are as follows: cost function
two types of GA method. First is the Single-Objective GA (shortly means the generated CM which the MCNP uses to calculate the
SOGA) developed to solve the problems comprising single objec- objectives; (discreet) variables refer to fuel element locations given
tive. Second is the Multi-Objective GA (MOGA) utilized for the in the CM; objectives to be achieved are the ex and ppfmax values
problems having two or more than two objectives. In this study, of a generated CM. The used GA methods and the operators are
both of them are utilized for the case studies mentioned above. described below in detail.
A continuous GA with discreet variables is used since this tech-
nique is easy to use and does not require coding/decoding of 2.3.1. Chromosome representation of the generated core maps
the chromosomes. Selection, Elitism, Crossover, and Mutation are Each CM is represented by a unique chromosome in the GA. A
the primary genetic operators. For the SOGA, Stochastic Uniform generated chromosome with a chromosome length of n is defined

Define the cost functions,


variables, GA parameters

n=0

Initial population generation

SOGA Initial cost calculation MOGA Identify


Sort the population
for each chromosome non-dominated fronts

Convergence Check: Offspring generation by


Max. generation, Crossover & Mutation
Reactor Physics Code:
Stagnant cycle, etc.
MCNP5
Input & output
Convergence Check:
Offspring generation by Max. generation,
Crossover & Mutation Stagnant cycle, etc.
n=n+1
n=n+1

Cost calculation Cost calculation

Combine population
Combine population with offspring population
with offspring population Offspring generation by
Crossover & Mutation

Identify
Fill the new population non-dominated fronts
Fill the new population
- Crowding distance
- Crowded-comparison

Fig. 2. GA flowchart.
88 M. Türkmen et al. / Nuclear Engineering and Design 295 (2015) 84–95

Genotype= [ 35 24 6 9 … 33 25 33 25 ] 2.3.5. Optimization parameters


Population size depends on the problem range and the con-
Phenotype= [ B1 B2 B3 B4 … F27 F28 F29 F30 ] straints defined. In this study, parent and offspring population sizes
(N) are selected to be 50. A total of population size of 100 is used
Fig. 3. Genotype and phenotype of a sample chromosome.
in the calculations. A uniform fixed crossover of 0.7 and mutation
fraction of 0.3 are used to produce next population. Furthermore,
by a number of genes. Each gene of a chromosome refers to a dif- since exchanging at least 2 different fuel elements yields a mini-
ferent element number (fuel element location) shown in Fig. 1. The mum mutation rate of about 0.03 (=2/69), mutation rate is set to
first gene represents B1 element while the last gene is F30. Further- the value changing from 0.03 to 0.06.
more, each gene carries specific information such as burnup data
and SF content required for the MC calculations. Same materials in 2.3.6. Method of SOGA with stochastic uniform sampling method
different locations are defined as the same gene type. Genotype and Consider the following optimization problem:
phenotype of a produced sample chromosome are given in Fig. 3. As
an illustration, first gene of the sample chromosome is filled with y = f (x1 , . . ., xn ) subject to xn ∈ S
the material 35 and corresponds to the element location B1. Chro-
where y is the objective, f is the cost function to be minimized (or
mosomes are randomly created from among the 69 burned fuel
maximized), xn is the number of variables (may be dependent or
elements for the first case study and, 59 burned and 10 fresh fuel
independent) to be optimized, and S is the one-dimensional search
elements for the second case study.
space with the defined upper/lower boundary limits.
This kind of problem is suitable for the SOGA since the SOGA is
2.3.2. Crossover operator capable of solving the optimization problems which contain only
For the optimization problems described above, all the fuel ele- one objective. Sorting the objective in ascending (or descending)
ments have to be used for once (similar to well-known Traveling order is sufficient to find the next best population. Stochastic Uni-
Salesman Problem – TSP), that is, it is not allowed to use a gene form Sampling (SUS) method is the main selection operator for the
more than once; therefore, Cycle Crossover (CX) (Oliver et al., 1987) SOGA. According to the method, the most probable individuals are
operator is preferred to produce new offspring from the selected selected with a random step size (equally spaced intervals) from
elite parents with Binary Tournament Selection. This method gen- the current population after sorting the cost values in descending
erates single offspring so that each gene comes from one of the order. The elite individuals are stored in the archive to use those
corresponding parents. Thus, half of absolute gene positions of both in the succeeding generation(s). The main steps of the SOGA are
parents are, on the average, preserved. For example, let’s say one outlined below:
parent chromosome is [1 2 3 4 5 6] and the other one is [4 5 6
3 1 2]. Then, the new offspring becomes [4 2 6 3 5 1], and the - For the generation i = 0, create a random initial population (P0 )
complementary offspring is [1 5 3 4 2 6]. with a population size of N.
- Calculate cost functions for each member of the population.
2.3.3. Mutation operator - Sort the population in ascending (or descending) order.
For an elite parent selected with Binary Tournament Selection, - Check the results for the predefined stopping criteria. If the results
a two-point mutation (totally for four fuel elements) is carried out are converged, then stop and provide the solutions. If not, then
using Exchange Mutation (or called Reciprocal Exchange) method go on searching.
(Banzhaf, 1990). According to this method, a randomly selected - Create an offspring population (O0 ) with a size of N by Crossover
gene of an elite parent is swapped with another randomly selected and Mutation operators. Use the Binary Tournament Selection
different gene of the same parent. For instance, suppose that the method to select the parents from the population.
parent chromosome is [1 2 3 4 5 6], and randomly selected genes - Calculate cost functions for each member of the offspring popu-
to be mutated are 2 and 5. In this case, the mutated chromosome lation.
becomes [1 5 3 4 2 6]. If the same materials are chosen for replac- - For i ≥ 1, combine the population (Ci = Pi ∪ Oi ) with the generated
ing, the mutation process is repeated until a different material is offspring population. Now, the combined population size is 2 N.
selected. Further, in order to avoid producing the same offspring, - Fill the new population (Pi + 1 ) from the best solutions of the
all the generated chromosomes are stored in the archive, and only combined population by using the elite selection operator of
different chromosomes are allowed to pass to the next generation. Stochastic Uniform Sampling until the new population is com-
pletely filled.
2.3.4. Convergence criterion - Increase the generation by one (i = i + 1) and go to convergence
For the SOGA, besides a limiting number of sequential stagnant check step.
cycles, standard deviation () of the average population in each
generation is calculated and used to terminate the algorithm. If the 2.3.7. Method of MOGA with NSGA-II
calculated deviation is less than a selected limiting value (de ) or In this case, the following optimization problem was consid-
if the population does not get any contribution from a number of ered:
consecutive cycles, the algorithm stops. Otherwise, the algorithm
y = [f1 (x1 , . . ., xn ) , . . ., fm (x1 , . . ., xn )] subject to xn ∈ S
runs for a number of generations.
In the case of MOGA, convergence criteria are selected to be where y is the objective vector, f is the cost function to be minimized
not only a limiting difference in the average crowding distance (or maximized or both), m is the number of objective functions
between the current population and the previous population but to be evaluated, xn is the number of variables (may be dependent
also a limited number of consecutive stagnant cycles. or independent) to be optimized, and S is the multi-dimensional
Number of the consecutive stagnant cycles and minimum value search space with the defined upper/lower boundary limits.
of the standard deviation of the average population are set to 4 Such problems always have more than one objective to be opti-
and 2/N, respectively. Minimum change in the average crowding mized. Therefore, they require an advanced sorting method which
distance in the current population is limited to 10/N. Maximum is quite different from the method that the SOGA uses. In a multi-
generation is selected to be 200. objective optimization problem, a feasible solution that all of the
M. Türkmen et al. / Nuclear Engineering and Design 295 (2015) 84–95 89

objectives are optimized simultaneously may not exist. In such (e.g., peak-to-average axial power density of the hot rod) are almost
cases, there is more than one feasible solution, these solutions are unvarying and within the safety limits defined by Safety Analysis
called Pareto-optimal solutions. Report.
In this study, the selected Multi-Objective GA method used to Each chromosome created by the GA codes (SOGA and MOGA)
find the best CMs is the Non-Dominated Sorting Genetic Algo- is sent (as input) to a prepared script to get the cost values. The
rithm (NSGA-II method) purposed by Deb et al. (2002). It is a script converts chromosomes to the input decks for which the
single parameter (only population size), well-tested and efficient selected physics code understands. Moreover, another generated
Multi-Objective GA. In the algorithm, fitness assignment is based script extracts the ex and ppfs values from the code output.
on the Fast Non-Dominated Sorting method (see Appendix A for MC model of the reactor is based on the previous study
the details) without an external population. Diversity is preserved (Türkmen and Çolak, 2014b). Material compositions of the burned
by using the Crowding Distance Mechanism (see Appendix B). The fuel elements used in the MC model are obtained from the results of
Crowding Distance provides worthwhile benefits to users by com- the previous study (Türkmen et al., 2015). The model includes only
puting the population density close the best solutions. Also, it neutron transport calculations for the particles having energies less
provides a uniform spread of solutions along the best-known Pareto than 20 MeV. The upper and lower energy boundaries for thermal
front without using a fitness sharing value. The selection is carried neutrons are selected to be 10−5 and 0.625 eV, respectively. More,
out by using the Crowded-Comparison operator (see Appendix C). it is assumed that energy of a fast neutron must be in the range
NSGA-II algorithm is briefly defined in the following main steps: of 0.1–20 MeV. Therefore, epithermal energy range is in between
0.625 eV and 100 keV. KCODE criticality calculations use 5000 neu-
trons per cycle with 50 passive and 250 active cycles. The used
- For the generation i = 0, create a random initial population (P0 )
history yields a maximum relative error of 0.00070 for the mul-
with a population size of N.
tiplication factor and 0.0040 for the power peaking factor. F4 cell
- Calculate cost functions for each member of the population.
flux tally (particles/cm2 s) is used to obtain the power peaking fac-
- Identify the non-dominated fronts (F1 , F2 , . . .) using the fast non-
tor values of the fuel elements. FMESH averaged over a mesh cell
dominated sorting method.
(particles/cm2 ) is used to evaluate neutron fluxes with respect to
- Create an offspring population (O0 ) with a population size of N
radial distance and core-average neutron spectrum. The neutron
by the Crossover and Mutation operators. Use the Binary Tourna-
spectrum calculations are carried out using the group bounds based
ment Selection method to select the parents from the population.
on 238-group ENDF/B-V.
- Check the results for the predefined stopping criteria. If the results
A full core calculation is performed since the reactor core is
are converged, then stop and provide the Pareto-front solutions.
asymmetric. To remain within the defined safety limits of the reac-
If not, then go on searching.
tor, all the control rods are set to their fully-withdrawn positions.
- Calculate cost functions for each member of the offspring popu-
It is assumed that the reactor operates at the steady-state and hot
lation.
full-power conditions. Neutron cross-section libraries at the mean
- For i ≥ 1, combine the population (Ci = Pi ∪ Oi ) with the generated
temperatures of the materials are used.
offspring population. Now, the combined population size is 2 N.
- Identify the non-dominated fronts (F1 , F2 , . . .) of Ci using the fast
non-dominated sorting method.
- Calculate crowding distance value of each solution in each front
3. Results
using the Crowding Distance method.
- Fill the new population (Pi + 1 ) from the subsequent non-
3.1. Shuffling the burned fuels without fresh fuels
dominated fronts with respect to their rank (from the best
solutions of F1 in the Ci ) until the new population is filled.
Change of several SOGA results with the generation is plot-
- If the size of the last non-dominated front Fn causes exceed-
ted in Fig. 4. The figure shows that the number of contributions
ing the size of the new population, select the best solutions by
decrease as the generation increases and, no contribution at some
sorting the solution of the last front according to the Crowded-
generations is observed. According to the results, the best pat-
Comparison operator in descending order and then fill the
tern has a multiplication factor of 1.00867 (1.173 $) ± 0.00068 with
population with the least-crowded solutions until no more
a maximum power peaking of 1.4826 ± 0.0041. Maximum power
position remains.
production, located at the node B1 (instrumented fuel element
- Create an offspring population with a population size of N by the
location) in fresh core map, is observed in node B1. The same fig-
Crossover and Mutation operators. Use the Binary Tournament
ure also includes change of the average population and the best
Selection method by selecting the parents from the population
solution as the number of generations increase. The average pop-
based on the Crowded-Comparison operator.
ulation first increases rapidly and then levels off at around 1.05 $.
- Increase the generation by one (i = i + 1) and go to convergence
After more than 30 generations (1500 iterations), the SOGA gives
check step.
the best chromosome. Standard deviation fluctuates around 0.069
and seemingly remains constant even if the generation increases.
2.3.8. Cost calculation: Reactor physics code However, to increase the population size and number of genera-
MCNP5 (X-5 Monte Carlo Team, 2005) is selected as a Monte tion would lead to decrease the standard deviation. As seen in the
Carlo based-reactor physics code for the cost calculations (function figure, there is almost no change in the average population and
evaluations). The code is capable of solving the neutron transport its corresponding standard deviation after about 50th generation.
problems including calculation of multiplication factor and cell- In addition to these, contributions decrease significantly. Thus, it
average flux values in a critical system. In this study, the code is can be deduced that the solution is converged. At this point, the
used to calculate the ex and ppfmax for each generated CM. algorithm stopped after the four stagnant cycles.
Excess reactivity is calculated by using the formula (keff − 1)/keff . The best pattern (upper values) obtained using SOGA is illus-
For ppfmax calculations, pin-power peaking factor, the peak rod trated in Fig. 5. The results are provided in terms of fractional
power to core-average rod power ratio, namely hot rod peak- burnup. The fuel element marked with vertical lines indicates the
ing factor, is used as the safety limit since it is reported in the location of the ppfmax . When compared with the out-in fuel man-
previous study (Türkmen and Çolak, 2014b) that other safety limits agement, the fuel elements with higher burnup are not located in
90 M. Türkmen et al. / Nuclear Engineering and Design 295 (2015) 84–95

1.40 0.10

1.20
0.09
1.00

0.08
0.80

ρex ($)

σunb.
0.60
0.07

0.40 Offspring contributions


Best solution 0.06
0.20 Average population
Standard deviation
0.00 0.05
0 10 20 30 40 50 60
Generation

Fig. 4. Change of SOGA results with the generation.

Fig. 5. Best patterns obtained using SOGA (upper) and MOGA (lower). (For interpretation of the references to color in this figure legend, the reader is referred to the web
version of this article.)

the inner rings. Instead, they are dispersed more or less evenly generation due to stagnant cycle criterion. The best pattern (lower
throughout the core. values) obtained using MOGA is illustrated in Fig. 5. The pattern
is selected from the pareto-front solutions in accordance with the
3.2. Shuffling the burned fuels with the fresh fuels maximum ex and the ppfmax lower than 1.5. Noting that “zero” bur-
nup values in the CM indicate locations of fresh fuel elements. The
Fig. 6 shows the average population and offspring contributions pattern has a multiplication factor of 1.00880 (1.195 $) ± 0.00068
to the elite population. It is clear that as the number of gen- with a maximum peaking factor of 1.4843 ± 0.0041. Maximum
erations increase, the number of contributions from the current power production is observed in node B1. Initial generation, mid-
population to the next population decrease. Average crowding dis- point (50th) and pareto-front (final) solutions on the solution space
tance changes with the generation; however, it becomes constant are plotted in Fig. 7. The solutions approach the pareto-front line as
after about 60th generation. The algorithm stopped at the 103th the generation increases. According to the results, it appears that
M. Türkmen et al. / Nuclear Engineering and Design 295 (2015) 84–95 91

1.40 0.7
Offspring contributions
Average population
1.20 Average crowding distance
0.6

Average Crowding Distance


1.00

0.5
0.80

ρex ($) 0.60


0.4

0.40
0.3
0.20

0.00 0.2
0 20 40 60 80 100
Generation

Fig. 6. Change of MOGA results with the generation.

1.40
Initial generation
Mid-point (50th)
1.20
Pareto-front (final)

1.00

0.80
ρex ($)

0.60

0.40

0.20

0.00
1.44 1.45 1.46 1.47 1.48 1.49 1.5 1.51
ppfmax

Fig. 7. Initial, mid-point and pareto-front solutions on the solution space.

the Pareto-front solutions converge a semi-concave curved line. variation of the fast flux. Also, fractional changes of the obtained
All the solutions in the pareto-front solutions can be used for the best patterns from the fresh CM are given in the same figures in
specific purposes. units of percentage. In these figures, “zero” radial distance means
core center; negative and positive values illustrate right and left
hand sides of the CM. Flux calculations are extended through the
3.3. Comparison of the generated best core maps with the fresh graphite and water blocks.
core map From the figures, although the shape of the thermal/fast neutron
fluxes of the generated CMs fits very well with the shape of the
The result of the previous study (Türkmen et al., 2015) shows fresh CM, there is an absolute decrease in the thermal fluxes at the
that the reactor loaded with the fresh fuels has an initial excess center of the reactor core. All the fractional changes appear almost
reactivity of about 1.4 $ (keff ≈ 1.0103) at the hot full power condi- symmetric from the core center. It means that total reactivity of
tion, fission product equilibrium and steady-state condition. Using the right side of the CM nearly equals to that of the left side. It is
this information, the reactivity difference between the SOGA and calculated that there is a certain decrease in the thermal flux up
fresh core is found to be 23 ± 10 cent while the reactivity difference to 3.0%1 for the SOGA and 2.0% for the MOGA inside the reactor
between the MOGA and fresh core is 21 ± 10 cent. It is clear that core. When compared with the fresh CM, a significant thermal flux
excess reactivity of any CM to be generated using the GAs cannot increase in the SOGA is observed inside the graphite and water
be higher than that of the fresh CM. That is why, it seems that the blocks. An interesting point is that the thermal flux makes a peak at
calculated best core maps are quite sufficient to use as the reloading periphery of the core when the SOGA is used; but, not for the MOGA.
patterns.
Variation of thermal neutron flux with the radial distance for
the fresh CM and the best CMs obtained using the SOGA and MOGA
is shown in Fig. 8. A similar figure, Fig. 9, is presented for the 1
Computational error of the calculated values is no more than 0.3%.
92 M. Türkmen et al. / Nuclear Engineering and Design 295 (2015) 84–95

7.0 8
Fresh Core Map
SOGA
6.0 MOGA 6
SOGA/Fresh

Neutron Flux (1012 n cm-2 s-1)


MOGA/Fresh
5.0 4

Fractional Change (%)


4.0 2

3.0 0

2.0 -2

1.0 Graphite Reactor Core Graphite -4


Water Water

0.0 -6
-70 -60 -50 -40 -30 -20 -10 0 10 20 30 40 50 60 70
Radial Distance (cm)

Fig. 8. Change of thermal flux with the radial distance for the generated core maps. (For interpretation of the references to color in this figure legend, the reader is referred
to the web version of this article.)

6.0 6
Fresh Core Map
SOGA 5
MOGA
5.0
SOGA/Fresh
4
Neutron Flux (1012 n cm-2 s-1)

MOGA/Fresh

Fractional Change (%)


4.0 3

2
3.0
1

2.0 0

-1
1.0
Water Graphite Reactor Core Graphite Water -2

0.0 -3
-70 -60 -50 -40 -30 -20 -10 0 10 20 30 40 50 60 70
Radial Distance (cm)

Fig. 9. Change of fast flux with the radial distance for the generated core maps. (For interpretation of the references to color in this figure legend, the reader is referred to
the web version of this article.)

This situation can be attributed to the positional arrangement of of higher than 10 MeV and less than 10−4 eV is meaningless due to
the fuel elements inside the core, referring to Fig. 5. Fast flux inside huge computational error.
the core is higher than (not more than 1%) the fresh core when Removal of a certain quantity of neutron poisons (e.g., Xe and
the burned fuels with fresh fuels are used. Similar to thermal flux Sm) from the reactor core causes significant decrease in the num-
distribution of the SOGA, fast flux of the SOGA is lower than (about ber of neutrons absorbed in the thermal energy range (as seen in
1% less) the fresh core at the core center and then rapidly increases thermal spectrum of the MOGA after 0.04 eV), whereas the pres-
(by about 1%) at the most outer rings, in particular in the Ring E. ence of the poisoners leads to a significant reduction in the number
This is because most of the least-burned fuel elements are located of neutrons less than 0.07 eV (as in the thermal spectrum of the
at that ring. SOGA). In addition, small changes in 238 U content of the core yield
Core-averaged neutron spectrum for the fresh CM and the best higher/lower resonance absorptions around neutron energy of 1 eV.
CMs optimized using the SOGA and MOGA is plotted in Fig. 10.
For the tally purposes, the mesh geometry is limited to the fuel
element length (including top/bottom graphite reflectors) in the 4. Discussion
axial direction and to the inner surface of the graphite reflector in
the radial direction. The results show that when the fuel elements Besides the objectives (ex and ppfmax ) optimized in this study,
are arranged as suggested in the SOGA and/or MOGA, no signif- neutron group fluxes in the central thimble, pneumatic system,
icant change is observed in the range of about 1 eV and 10 MeV. beam ports and thermal column should be optimized simulta-
However, spectra are affected considerably by the core arrange- neously. This is because these irradiation channels are frequently
ments, after the particular energy less than 1 eV. According to the used in the critical experiments, some of which require certain neu-
figure, replacing a number of burned fuels with the fresh fuels in tron flux values. This study shows that the CM produced by the
the MOGA results in about 10% flux collapse near 0.2 eV. In the case MOGA just meets the requirements for the central thimble irradi-
of SOGA, there is more than 20% increase in the same energy. This ation channel by providing the thermal/fast flux at the core center
is because 235 U has a peak absorption cross section at that energy. (from Fig. 8) as much as the fresh CM. In addition, of course, other
It should be noted that the fractional change in the energy intervals irradiation channels must be analyzed carefully. However, for now,
M. Türkmen et al. / Nuclear Engineering and Design 295 (2015) 84–95 93

1.0E+12 40
Fresh Core Map
1.0E+11 SOGA
MOGA 30
1.0E+10 SOGA/Fresh

Neutron Flux (n cm-2 s-1)


MOGA/Fresh

Fractional Change (%)


1.0E+09 20

1.0E+08
10
1.0E+07

1.0E+06 0

1.0E+05
-10
1.0E+04

1.0E+03 -20
1.0E-11 1.0E-09 1.0E-07 1.0E-05 1.0E-03 1.0E-01 1.0E+01
Energy (MeV)

Fig. 10. Core-averaged neutron spectrum for the generated core maps. (For interpretation of the references to color in this figure legend, the reader is referred to the web
version of this article.)

these channels are beyond the interest of this study and will be Table 2
A summary of the results of the optimum CMs.
investigated as a future study.
Recalling the parameters used in the GA such as population Optimization methods keff ex ($) ppfmax
size, crossover/mutation fractions, mutation rate and convergence SOGA:SUS 1.00867 ± 0.00068 1.173 1.4826 ± 0.0041
criteria, they were adjusted using the user experiences and the MOGA:NSGA-II 1.00880 ± 0.00068 1.195 1.4843 ± 0.0041
outcomes of the previous studies described in the literature. For
instance, the total population size of 100 was used in this study
while the previously published articles were used the population the fission products are not taken into account. Still, it is possible
sizes of 50 to 200 depending on the problem size. Similarly, most of to accelerate the computations by eliminating the isotopes at very
the GA methods imply to use a mutation rate of 0.01 in order not to low amount.
turn the solver into classic search event. However, a mutation rate Compared with the result of the fresh CM having the initial
changing from 0.03 (min) to 0.06 (max) was preferred in this study excess reactivity of about 1.4 $, the maximum excess reactivity is
as the exchange mutation method was adjusted to swap up to 4 found to be 1.19 $ when the MOGA is used. It can be deduced that
fuel elements at once. In brief, small changes in these parameters shuffling the fuels may lead to an extension in the fuel cycle length
significantly affects the convergence, speed and, thus, the obtained as high as the fuel cycle length of the fresh CM. It, however, needs a
optimum solutions. Therefore, the users must be aware of which separate burnup calculation to get the exact value of the fuel cycle
values of the parameters are better for their problems. extension. For this purpose, for the best patterns provided in this
Although the solutions provided in this study are stated as the study, discharge burnup values of the fuel elements are going to be
best solutions, it should be noted that those, in fact, may not be presented as a separate study.
the best. The reason is the stopping criteria defined by a num- Concisely, when compared with the fresh core, it is observed
ber of external parameters. In the study, the best patterns were that the spectrum is remarkably affected by not only fissile content
obtained when the stagnant cycle criterion specified was success- but also amount of the thermal neutron absorber. Here, fissile ura-
fully achieved. Therefore, one thing should be noted that there is, nium tends to increase some part of the neutron energy spectrum
however, a possibility to find a better CM(s) if the algorithms go on while the absorbers tend to decrease thermal flux. Furthermore, it
searching through the search space. should be kept in mind that different core arrangement with the
The most important issue in an optimization study is the time same fuel elements means different neutron leakage characteristic
required to compute a cost value. Although the Monte Carlo-based and, therefore, a different spectrum for the core.
codes yield more reliable results than the deterministic codes
which use discrete ordinate methods or other methods, the MC 5. Conclusion
codes naturally consume more time since each particle is simu-
lated. In this study, one cost calculation takes more or less eleven In this study, a set of optimal CMs were searched by coupling
minutes to obtain less than a standard deviation of 0.07% when the GA and MC method. Optimization was performed for the prob-
Intel Xeon E5 Processors (12 × 4) with a core frequency of 3.10 GHz lems: (i) maximization of the ex using SOGA when the burned fuel
are used. For over 2000 iterations, the necessary time for the full elements with no fresh fuels are used, (ii) maximization of the ex
computation would be at least two weeks. and minimization of ppfmax using MOGA when the burned fuel ele-
Besides longer computational time originated from the nature ments with the fresh fuels are used. Best CMs obtained using the
of the MC method, extensive geometry and material details used GA were compared with the fresh core from the viewpoint of core-
in the reactor model also result in increasing the computational averaged neutron spectrum and change of the neutron group fluxes
time. Simple geometry and/or less material mean less time for with the radial distance. A summary of the results is presented in
computing. For example, the reactor model used in this study cov- Table 2.
ers all of the reactor in/out-core equipment as well as the tally There is no observed case that exceeds the safety limits (defined
calculations are performed using a total of 250 isotopes for each by the safety report2 ) when the burned fuels with no fresh fuels are
fuel material. It is important to note in here that, it is observed used. However, when the mix of burned and fresh fuels are used, it is
from the previous study that using only important fission prod-
uct/actinide isotopes in the burnup computations changes the ex
results by around 200 pcm. It is clear that the magnitude of this 2
The safety limits for the excess reactivity and the ppfmax are, in turn, 3 $ and
computational error would swamp the entire calculations if all of 1.60.
94 M. Türkmen et al. / Nuclear Engineering and Design 295 (2015) 84–95

found that some generated CMs have higher ppfmax value than 1.50 - Set the upper
  boundary
and lower  values as infinite distance,
(but not the limiting value) at the locations where the fresh fuels are Im x(1,m) = ∞ and Im x(l,m) = ∞.
   
located. Therefore, the best CMs are selected within the generated - For i = 2,. . .,l − 1, calculate Im x(i,m) = Im x(i,m) +
CMs in such a way that the excess reactivity and peaking factor     
max − I min

Im x(i+1,m) − Im x(i−1,m) /Im m .
values are sufficiently below the safety limits. It is also understood
from the results that the SOGA method instead of the MOGA in the
optimization of the second case is sufficient in generating the CMs. Appendix C. Details of crowded-comparison operator
If the problems considered in this study are solved using the
classic search method, there would be about 4 × 1019 (a combina- After the fast non-dominated sorting method and the crowding
tion of 69 fuel elements using 38 different materials) possible CMs distance calculation are performed, Crowded-Comparison opera-
to be searched when the burned fuels with no fresh fuels are used tor is used to select the elite individuals for the offspring creation
for the shuffling. The number would be about 3 × 1019 (a combina- by using the binary tournament selection technique. Crowded-
tion of 69 fuel elements using 30 different materials) for the mix Comparison operator (≺) is defined as follows:
of burned and fresh fuels. On the other hand, in this study, total
number of core maps used to find the best patterns for the SOGA
and MOGA are 2950 and 5100, respectively. In short, the GA meth- - Assume that each solution p in front Fi has its non-domination
ods used in this study provide worthwhile benefits to the users by rank (prank ) and crowding distance value (xdistance ).
shortening the search time significantly. - Prefer the solution with the lower (better) rank if two solutions
To sum up, the optimal CM with the SOGA may extend the reac- have different non-domination ranks. If the selected solutions are
tor cycle life to some degree. However, with the fresh fuels, the located at the same front then, the solution that is located in a
reactor cycle life can be extended to a large period by choosing a lesser crowded region.
proper CM, which ensures the safety limits, as well. - If (prank < qrank ) then (p ≺n q).
- Or if (prank = qrank ) and (xdistance > ydistance ) then (p ≺n q).

Appendix A. Details of fast non-dominated sorting method


References
- According to the method, the population is sorted according to
Banzhaf, W., 1990. The “molecular” traveling salesman. Biol. Cybern. 64, 7–14.
non-domination sets. For any solution of p and q in the main Baker, J.E., 1987. Reducing bias and inefficiency in the selection algorithm. In:
population (P), the algorithm follows below steps: Proceedings of the Second International Conference on Genetic Algorithms and
- Set Sp =∅. If p dominates q, then insert solution q into Sp = Sp ∪ {q} their Application. L. Erlbaum Associates, Hillsdale, NJ, pp. 14–21.
Deb, K., et al., 2002. A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE
where Sp is the solution set which contains all the individuals Trans. Evol. Comput. 6 (2), 182–197.
dominated by p. Do, B.Q., Nguyen, P.L., 2007. Application of a genetic algorithm to the fuel reload
optimization for a research reactor. Appl. Math. Comput. 187, 977–988.
 np = 0. If q dominates p, then increase the np by one
- Set
François, J.-L., et al., 2013. Comparison of metaheuristic optimization techniques for
np = np + 1 where np is the number of solutions which dom- BWR fuel reloads pattern design. Ann. Nucl. Energy 51, 189–195.
inate p. Goldberg, D.E., 1989. Genetic Algorithms in Search Optimization and Machine Learn-
- When no solution (if np equals to zero) dominates p, the solution ing. Addison-Wesley, Reading, NY.
Hedayat, A., et al., 2009. Optimization of the core configuration design using a
p belongs to first front F1 , F1 = F1 ∪ {p}.
hybrid artificial intelligence algorithm for research reactors. Nucl. Eng. Des. 239,
- Set the front counter to one (i = 1). While the front Fi is not empty 2786–2799.
Fi =/ ∅, set Q =∅ where Q is used to store the solutions for the next Hedayat, A., 2014. Developing a practical optimization of the refueling program for
ordinary research reactors using a modified simulated annealing method. Prog.
front Fi + 1 .
Nucl. Energy 76, 191–205.
- For each solution p in the front Fi , and each solution q in the set Holland, J.H., 1975. Adaptation in Natural and Artificial Systems. University of Michi-
Sp , follow the below rule; gan Press, Ann Arbor, MI.
- Visit each member q in the set Sp and decrease nq by one, Hooke, R., Jeeves, T.A., 1961. Direct search solution of numerical and statistical prob-
lems. J. Assoc. Comput. Mach. 8 (2), 212–229.
nq = nq − 1.   Huda, M.Q., Bhuiyan, S.I., Obara, T., 2008. Burnup analysis and in-core fuel manage-
- When nq becomes zero nq = 0 , put the solution q in a separate ment study of the 3 MW TRIGA Mark II research reactor. Ann. Nucl. Energy 35,
list Q, Q = Q ∪ {q}. 141–147.
Karahroudi, M.R., Shirazi, M.S.A., Sepanloo, K., 2013. Optimization of designing the
- Increase the front counter by one (i = i + 1) to find the members core fuel loading pattern in a VVER-1000 nuclear power reactor using the genetic
of the next fronts. algorithm. Ann. Nucl. Energy 57, 142–150.
- Now, identify the members of the current front Fi by setting Lyric, Z.I., et al., 2013. Optimum burnup of BAEC TRIGA research reactor. Ann. Nucl.
Energy 55, 225–229.
Fi = Q. Mazrou, H., Hamadouche, M., 2006. Development of a supporting tool for optimal
fuel management in research reactors using artificial neural networks. Nucl. Eng.
Des. 236, 255–266.
Appendix B. Details of crowding distance Oliver, I.M., Smith, D.J., Holland, J.R.C., 1987. A study of permutation crossover oper-
ators on the traveling salesman problem. In: Genetic Algorithms and Their
Applications: Proc. Second Int. Conf. on Genetic Algorithms. Lawrence Erlbaum
After the non-dominated sorting step is completed, the density Associates, Cambridge, MA, pp. 224–230.
calculation of the solutions around a particular solution needs to Schaffer, J.D., 1985. Multiobjective optimization with vector evaluated genetic algo-
be carried out. This is because the selection process is performed rithms. In: Proc. Int. Conf. on Genetic Algorithm and their Applications.
Srinivas, N., Deb, K., 1994. Multiobjective optimization using nondominated sorting
over the rank and crowding distance values of the individuals. The in genetic algorithms. J. Evol. Comput. 2 (3), 221–248.
Crowding Distance stands on the idea calculating the Euclidian dis- Stevens, J.G., Smith, K.S., Rempe, K.R., 1995. Optimization of pressurized water reac-
 
tance between each solution in a front according to m objectives. tor shuffling by simulated annealing with heuristics. Nucl. Sci. Eng. 121, 67–88.
For each front Fi containing the number of solutions l = Fi , the
Thakur, A., Singh, B., Krishnani, P.D., 2013. In-core fuel management for AHWR. Ann.
Nucl. Energy 57, 47–58.
algorithm can be expressed as below: Turing, A.M., 1950. Computing machinery and intelligence. Mind LIX 238, 433–460.
Turinsky, P.J., Keller, P.M., Abdel-Khalik, H.S., 2005. Evolution of nuclear fuel man-
agement and reactor operational aid tools. Nucl. Eng. Technol. 37 (1), 79–90.
- Define the crowding-distance set I as I(i) = 0 for each solution, i. Türkmen, M., Çolak, Ü., 2013. ITU TRIGA Mark II research reactor: a benchmark
- For each objective m, sort the solutions of Fi in the ascending analysis with various codes. In: The European Research Reactor Confer-
ence: RRFM2013, April 21–25, RRFM2013 Trans., Saint Petersburg, Russia, pp.
order. Assign each solution as x(i,m) in the sorted listed. 395–400.
M. Türkmen et al. / Nuclear Engineering and Design 295 (2015) 84–95 95

Türkmen, M., Çolak, Ü., 2014a. Fuel burnup calculation in ITU TRIGA Mark II Research Yen, G.G., Lu, H., 2003. Dynamic multiobjective evolutionary algorithm: adaptive
Reactor by using Monte Carlo method. In: NENE2014: 23rd Int. Conf. Nuclear cell-based rank and density estimation. IEEE Trans. Evol. Comput. 7 (3), 253–274.
Energy for New Europe, September 8–11, 2014, Portorož, Slovenia. Zameer, A., Mirza, S.M., Mirza, N.M., 2014. Core loading pattern optimization of a typ-
Türkmen, M., Çolak, Ü., 2014b. Analysis of ITU TRIGA Mark II Research Reactor using ical two-loop 300 MWe PWR using Simulated Annealing (SA), novel crossover
Monte Carlo method. Prog. Nucl. Energy 77, 152–159. Genetic Algorithms (GA) and hybrid GA(SA) schemes. Ann. Nucl. Energy 65,
Türkmen, M., Çolak, Ü., Ergün, Ş., 2015. Effect of burnup on the neutronic parameters 122–131.
of ITU TRIGA Mark II Research Reactor. Prog. Nucl. Energy 83, 26–34. Zitzler, E., Thiele, L., 1999. Multiobjective evolutionary algorithms: a comparative
X-5 Monte Carlo Team, 2005. MCNP—a general Monte Carlo N-particle transport case study and the strength Pareto approach. IEEE Trans. Evol. Comput. 3 (4),
code, version 5 , LA-UR-03-1987, April 24, 2003 (Revised October 3rd, 2005). 257–271.

You might also like