G Model

JIEC-1500; No. of Pages 4

Journal of Industrial and Engineering Chemistry xxx (2013) xxx–xxx

Contents lists available at ScienceDirect

Journal of Industrial and Engineering Chemistry

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

Naphthalene sublimation. Experiment and optimisation based on

neuro-evolutionary methodology
Silvia Curteanu a,*, Mirela Smarandoiu a, Doina Horoba a, Florin Leon b
a
‘‘Gheorghe Asachi’’ Technical University of Iasi, Faculty of Chemical Engineering and Environmental Protection, Str. Prof. Dr. Doc. Dimitrie Mangeron Nr. 73,
700050 Iasi, Romania
b
‘‘Gheorghe Asachi’’ Technical University of Iasi, Faculty of Automatic Control and Computer Engineering, Str. Prof. Dr. Doc. Dimitrie Mangeron Nr. 53A,
700050 Iasi, Romania

A R T I C L E I N F O A B S T R A C T

Article history: This paper presents experimental and simulation studies on spherical samples of naphthalene
Received 16 April 2013 sublimation in the presence of air as driving agent. The mass transfer rates and the influence of air flow
Accepted 1 August 2013 characteristics on mass transfer are investigated. The degree of sublimation and sublimation front
Available online xxx
position as function of time are also determined. A modelling and optimisation strategy based on neural
networks and genetic algorithms, designed in simple and adaptive variants, is developed and applied for
Keywords: determining optimal working conditions which lead to the maximisation or minimisation of sublimation
Mathematical modelling
rate. Accurate results are obtained proving the efficiency of the neuro-evolutionary methodology.
Optimisation
Adaptive genetic algorithm
ß 2013 The Korean Society of Industrial and Engineering Chemistry. Published by Elsevier B.V. All rights
Mass transfer reserved.
Sublimation

1. Introduction naphthalene sublimation process was performed using an ANN–

Naphthalene sublimation technique is one of the most
convenient method to study heat and mass transfer with multiple
applications such as: determining the coefficients of heat and mass
transfer for various configurations of flow space, for various forms
of the surface samples fixed or moving [1–4], estimation of vapour
pressure [5], studying the heat transfer for complex geometric
objects [6–8].
To investigate this type of process by simulation has a series of
advantages related to the possibility of obtaining accurate
predictions or optimal working conditions. With this purpose,
artificial intelligence tools, especially artificial neural networks
(ANN) and genetic algorithms (GA), are often applied [9–13].
The present paper examines the sublimation of naphthalene
spherical samples under atmospheric pressure in air as entrainer,
without recycle. The sample weight was measured continuously as
a function of time, at different air flow characteristics. The
experimental data were used to calculate the mass transfer rate,
the degree of sublimation, and the sublimation front position. The
influence of air flow characteristics has also evaluated in order to
determine optimal process conditions. The optimisation of the
* Corresponding author. Tel.: +40 232 278963; fax: +40 232 271311.
E-mail address: silvia_curteanu@yahoo.com (S. Curteanu).
GA methodology which related the rate of sublimation in the
minimum or maximum values with optimal working conditions
(process time, entrainer temperature, and entrainer flow rate). A
feed-forward neural network with high performance was devel-
oped as a process model to be included into the optimisation
procedure. An efficient GA with adaptive parameters, i.e. adaptive
crossover and mutation rates, was used to solve the optimisation
problem. In an ANN–GA methodology, original adaptive elements,
own implementation, and a new application related to the
naphthalene sublimation define and distinguishe this approach
from those known in the open literature.
2. Experimental
2.1. Materials and methods
The experiments used in the present work consisted in the
sublimation of naphthalene spherical samples under atmospheric
pressure in air as entrainer, without recycle, with the goal to
investigate the mass transfer rates and evaluate the influence of air
flow characteristics on mass transfer.
Fig. 1 shows schematically the experimental equipment
employed in this study. The installation is constituted by a
sublimation chamber (1), a heat exchanger for air (2), a centrifugal
fan (3), an analytical balance (4), and a thermostat (5). The

1226-086X/$ – see front matter ß 2013 The Korean Society of Industrial and Engineering Chemistry. Published by Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.jiec.2013.08.007

Please cite this article in press as: S. Curteanu, et al., J. Ind. Eng. Chem. (2013), http://dx.doi.org/10.1016/j.jiec.2013.08.007
G Model
JIEC-1500; No. of Pages 4

2 S. Curteanu et al. / Journal of Industrial and Engineering Chemistry xxx (2013) xxx–xxx

purpose was to evaluate the performance of the process, quantified

by the sublimation rate, vs, (output variable) depending on process
2 time, t, entrainer temperature, T, and entrainer flow rate, Mv (the
1 three input variables of the neural model).
The developed neural model was included into an optimisation
6
procedure, designed for obtaining maximum/minimum values for
the sublimation rate, which was solved with GA. This algorithm
8
10 was chosen based on a series of benefits given by the flexibility,
9 ease of operation, minimal requirements, or global perspective
7 offered. In the present work, GA was applied comparatively in two
variants: simple and adaptive.
3 Following the biological analogy, an individual is represented
4
by a chromosome, composed of a number of genes, codified, in this
5 work, by real values. The fundamental operators of a genetic
algorithm are: selection, crossover, and mutation. Roulette-wheel
selection, arithmetic real-values crossover, and mutation by
resetting a gene to a random value in its domain were applied
Fig. 1. Schematic of test installation: (1) sublimation room; (2) heat exchanger; (3) here for optimising the rate of naphthalene sublimation. The stop
centrifugal fan; (4) analytical balance; (5) thermostat; (6) rotameter; (7) valve; (8) condition of the procedure was determined by a pre-established
sample; (9) shaft and (10) digital thermometer. maximum number of generations.
Number of individuals in the population, maximum number of
experiments were carried out in a sublimation chamber, a jacketed generations, crossover and mutation rates are control parameters of
vertical column from thermoresistant glass with 55 mm internal GA which influence the algorithm performance. Generally, it is
diameter, and 600 mm height. A double pipe from glass is used as difficult to choose the most appropriate values for these parameters,
heat exchanger for air. The entrainer heated with electrical the method often applied being trial-and error. A solution to alleviate
resistance in the heat exchanger enters at top of the sublimation this problem is proposed in this work and is consisted in the use of
chamber and passes through the sample. The volumetric flow rate adaptive rates for crossover and mutation. To be able to dynamically
of the air through the sublimation chamber is measured by a adjust these rates, it is important to identify the moment when the
rotameter (6) and controlled with the valve (7). algorithm begins to converge into an optimum, i.e. to observe the
The samples were created by cooling molten naphthalene poured average value of the fitness of the population f , in conjunction with
into stainless steel moulds with about 20 mm diameter. The the maximum fitness fmax. For a population which has converged into
spherical sample (8) was attached to the end of vertical shaft (9) an optimum, f max  f should have a lower value than in the case
placed in the sublimation chamber. During each experiment, the when the population is dispersed in the search space. Thus, the main
temperature in sublimation chamber was kept constant with hot idea of the adaptive genetic algorithm is that the values of the
water circulated in the column jacket. Sample weight was measured crossover rate, pc, and the mutation rate, pm, should be varied
continuously as a function of time with an analytical balance. The depending on the difference f max  f . The rates pc and pm should
sample temperature was measured with a digital thermometer (10) increase when the algorithm tends to converge, and thus increase the
placed in the vicinity of the sample. The duration of the experiments genetic diversity and lower the chances of converging into a local
was measured with a digital chronometer. optimum. However, even if the GA may no longer get stuck into a local
The experimental determinations were made for four values of optimum anymore, its performance, in terms of generations needed
air temperature (50 8C, 60 8C, 65 8C, and 70 8C) and, for each for convergence, may decrease.
temperature, four values of air flow rate (2000 L h1, 3000 L h1, In order to eliminate this problem, good solutions in the
4000 L h1, and 5000 L h1) were considered. The weight of the populations must be kept. This translates into low values for pc and
sample during the sublimation process was measured with an pm for individuals with good fitness functions, and large values for
analytical balance. 150 experimental data resulted from experi- individuals with poor fitness functions. The crossover rate should
ments constituted the data set for modelling and optimisation of be proportional to fmax  fp, where fp is the largest fitness value out
the sublimation process. of the two parents. The mutation rate should be proportional to
To calculate the rate of naphthalene sublimation, the amount of fmax  f, where f is the fitness value of the current individual.
sublimated naphthalene in time, the particle radius, and its surface Therefore, the expressions used to adaptively determine the
area have been determined. crossover and mutation rate, respectively, are:

2.2. Modelling and optimisation strategy k1  ð f max  f p Þ

pc ¼ (1)
f max  f
In this paper, artificial neural networks were chosen to model
the sublimation process of naphthalene based on the fact that they k2  ð f max  f Þ
possess the ability to learn what happens in the process without pm ¼ (2)
f max  f
actually modelling the physical and chemical laws that govern the
system. In the above equations, k1 and k2 are constants, k1,2 2 R+ and
Our experimental data fulfil a necessary condition for ANN k1,2  1.
modelling: a sufficient number of data (150) was obtained which
uniformly cover the investigated domain. Feed-forward multilay- 3. Results and discussion
ered neural network type (multilayer perceptron, MLP) was chosen
and tested for naphthalene sublimation based on its simple 3.1. Processing of the experimental data
structure, ease of designing and training, good predictions it
provides, and the possibility to combine it with other simulation The sublimation rate, the degree of sublimation, and sample
techniques (in this case, genetic algorithms). The modelling radius were calculated as a function of time for various

Please cite this article in press as: S. Curteanu, et al., J. Ind. Eng. Chem. (2013), http://dx.doi.org/10.1016/j.jiec.2013.08.007
G Model
JIEC-1500; No. of Pages 4

S. Curteanu et al. / Journal of Industrial and Engineering Chemistry xxx (2013) xxx–xxx 3

0.4 Table 1
Performance evaluation of SGA and AGA for the minimisation problem.
0.35
Pop. size No. gen. vs103 vs103 vs103 vs103
SGA avg. SGA best AGA avg. AGA best
0.3
Degree of sublimation

20 50 0.0102 0.0002 0.0039 0.0000

0.25 30 50 0.0060 0.0000 0.0031 0.0000
50 50 0.0031 0.0000 0.0011 0.0000
0.2 20 100 0.0119 0.0002 0.0032 0.0000
30 100 0.0089 0.0001 0.0011 0.0000
0.15 50 100 0.0034 0.0000 0.0006 0.0000
20 200 0.0094 0.0000 0.0013 0.0000
0.1 30 200 0.0079 0.0001 0.0005 0.0000
50 200 0.0037 0.0000 0.0002 0.0000
0.05 50 500 0.0047 0.0000 0.0000 0.0000

The values in bold show the GA parameters corresponding to the minimisation of

0 sublimation rate of naphthalene.
0 15 30 45 60 75 90 105 120 135
time, min 3.2. Modelling and optimisation results

Fig. 2. Degree of sublimation of naphthalene vs. time at entrainer temperature
T = 70 8C and various entrainer flow rates: (&) 2000 L h1; (*) 3000 L h1; (&)
4000 L h1 and (*) 5000 L h1.
temperatures and flow rates of entrainer, using experimental
data gathered. Fig. 2 presents an example for the variation of
degree of sublimation versus time, considering various entrainer
temperatures. Air temperature enhances greater sublimation
rate than flow rate of the entrainer. As expected, increasing
temperature and air flow causes a marked reduction of the
sample radius.
An example for the variation of sublimation rate of naphthalene
is given in Fig. 3 where two stages of the process is evidenced. In
the first stage, the sublimation rate increases in time because the
temperature of the surface sample rises and affects both
desorption of molecules from the solid and movement of the
vapour formed by the diffusion layer (the sublimation heat, the
naphthalene vapour pressure at the interface gas–solid, and the
diffusion coefficient of the naphthalene vapours in air are functions
of temperature). The duration of the first stage depends on the air
characteristics and is 15–30 min. The effect of temperature is
greater than that of the flow rate. In the second stage, sublimation
takes place in stationary conditions (air temperature and flow are
constant) and the sublimation rate is approximately constant. The
average rate of sublimation is influenced less by increasing air flow
than increasing its temperature.
0.14
The available dataset was divided into 2/3 for training and 1/3
for testing phases. A trial and error method, combined with the
well known back-propagation algorithm, is applied to train
different neural networks models. The best network was selected
based on a balance between performance (best generalisation
capability) and size. This is realised by MLP (3:15:5:1), a multilayer
perceptron with two intermediate layers with 15 and 5 neurons,
respectively, having mean square error MSE = 0.003135 and
correlation r = 0.9934 in the training phase and MSE = 0.265847
and r = 0.9299 in the validation phase.
The optimisation procedure proposed in this paper is executed
in two stages: process modelling with artificial neural networks
and the actual optimisation based on genetic algorithms developed
in two variants, simple and adaptive.
In order to evaluate the performance of the adaptive genetic
algorithm for the minimisation and maximisation of vs modelled
with a neural network, a comparison was made with a simple GA,
using stochastic universal sampling, a fixed crossover rate of 90%, a
fixed mutation rate of 2%, and elitism: the best individual in a
population is directly copied into the next one.
The adaptive version uses the same configuration, excepting for
the crossover and mutation rates which are dynamically deter-
mined during the run of the algorithm. The values for the constants
required in Eqs. (1) and (2) are k1 = 1 and k2 = 0.5 [14].
Fifty trials were performed for each configuration, while
varying the population size and the number of generations used
as termination criterion. In each setting, the average fitness of the
50 trials was recorded as well as the best fitness observed. Tables 1
and 2 displays these performance metrics (as average and best
values) for the simple genetic algorithm (SGA) and adaptive
genetic algorithm (AGA) applied with the goal of minimising or
maximising vs.

0.12

0.1 Table 2
-1

Performance evaluation of SGA and AGA for the maximisation problem.

vs 10 , kg m s
-2

0.08 Pop. size No. gen. vs103 vs103 vs103 vs103

SGA avg. SGA best AGA avg. AGA best
-3

0.06 20 50 0.1897 0.2048 0.2056 0.2093

30 50 0.1920 0.2060 0.2065 0.2098
0.04 50 50 0.1922 0.2060 0.2081 0.2094
20 100 0.1967 0.2048 0.2061 0.2093
0.02 30 100 0.1966 0.2063 0.2076 0.2095
50 100 0.1977 0.2068 0.2086 0.2097
20 200 0.2021 0.2072 0.2056 0.2092
0
30 200 0.2012 0.2066 0.2076 0.2096
0 15 30 45 60 75 90 105 120 135
50 200 0.1995 0.2049 0.2089 0.2098
time, min 50 500 0.2029 0.2082 0.2091 0.2098

Fig. 3. Sublimation rate of naphthalene vs. time at entrainer flow rate 2000 L h1 The values in bold show the GA parameters corresponding to the maximisation of
and various air temperatures (&) 50 8C; (&) 60 8C; (*) 65 8C and (*) 70 8C). sublimation rate of naphthalene.

Please cite this article in press as: S. Curteanu, et al., J. Ind. Eng. Chem. (2013), http://dx.doi.org/10.1016/j.jiec.2013.08.007
G Model
JIEC-1500; No. of Pages 4

4 S. Curteanu et al. / Journal of Industrial and Engineering Chemistry xxx (2013) xxx–xxx

Table 3 the sublimation rate than temperature. After establishing station-

Best results obtained by SGA and AGA.
ary conditions of sublimation, the sublimation rate remains
Optimisation Algorithm t Mv T vs103 constant.
problem The simulation study was performed using a hybrid tool
Maximisation SGA 115.589 4907.882 69.956 0.2082 composed of feed-forward neural networks and genetic algorithms
Maximisation AGA 119.498 4997.412 69.999 0.2098 designed in simple and adaptive variants. The best modelling
Minimisation SGA 0.716 3030.908 53.728 0.0000 results are provided by a network MLP (3:15:5:1), which balanced
Minimisation AGA 0.176 4541.858 67.425 0.0000
the size and performance. ANN-AGA optimisation calculated
successfully the operating conditions which lead to a minimum
or maximum sublimation rate for naphthalene spherical samples
The best values for population size and number of generations (relative error less 2%).
are relatively small, meaning a reduced execution time: Pop. An important aspect to be emphasised is related to the general
size = 20 or 30 and No. gen = 50. character of this procedure which contains neural network models
As it can be seen from these tables, the adaptive version and two variants of genetic algorithms, providing opportunities to
constantly outperforms the classic genetic algorithm in terms of make predictions and calculate optimal working conditions for
solution quality. Since a fixed number of generations is used as the other different chemical engineering processes.
termination criterion, the execution times are approximately the
same in the two situations. Acknowledgements
Table 3 presents the actual best results for each algorithm.
Some experiments were performed to check the validity of the This work was supported by the ‘‘Partnership in priority areas –
optimisation results. Establishing the working conditions: PN-II’’ program, financed by ANCS, CNDI – UEFISCDI, project PN-II-
t = 119.5 min, Mv = 4997 L h1 and T = 70 8C, it was obtained PT-PCCA-2011-3.2-0732, No. 23/2012.
vs = 0.2075103 kg m2 s1 (compared to the theoretical values of
0.2098103) and for t = 116 min, Mv = 4910 L h1 and T = 70 8C,
vs = 0.2057103 kg m2 s1 (compared to the theoretical values of References
0.2082103). These experimental values are in a very good
concordance with simulation resulted for maximisation of [1] E.L. Andreas, Icarus 186 (2007) 24.
[2] T.A. Neumann, M.R. Albert, C. Engel, Z. Courville, F. Perron, Int. J. Heat Mass Trans.
sublimation rate (relative error about 1.2%). 52 (2009) 309.
[3] I.V. I.V. Shevchuk, Lecture Notes in Appl. Comput. Mech. 45 (2009) 193.
4. Conclusions [4] H.G. Kwon, S.D. Hwang, H.H. Cho, Int. J. Heat Mass Trans. 54 (2011) 1071.
[5] P.H. Chen, JrM. Miao, C.S. Jian, Rev. Sci. Instrum. 67 (1996) 2831.
[6] S.C. Lau, J. Cervantes, J.C. Han, R.J. Rudolph, K. Flannery, Int. J. Heat Mass Trans. 46
A number of mass transfer experiments have been planned and (2003) 3991.
performed in order to obtain sublimation rate for naphthalene [7] J.H. Park, S.Y. Yoo, KSME Int. J. 18 (2004) 1258.
[8] K. Hong, T.H. Song, Int. J. Heat Mass Trans. 50 (2007) 3890.
spherical samples at various operating conditions - temperature
[9] P. Zamankhan, P. Malinen, H. Lepomäki, AIChE J. 43 (1997) 1684.
and air flow rate. Analysing the experimental data, it was observed [10] J.O. Valderrama, J. Zavaleta, Ind. Eng. Chem. Res. 44 (2005) 4824.
that by increasing air flow rate and temperature, an increase in the [11] J.O. Valderrama, C. Faúndez, Inform. Technol. 17 (2006) 119.
rate of sublimation occurs, in both stages. In the first stage, the [12] F. Gharagheizi, Thermochim. Acta 469 (2008) 8.
[13] M. Bagheri, M. Bagheri, A.H. Gandomi, A. Golbraikh, Thermochim. Acta 543 (2012)
naphthalene sublimation rate increases in time due to increased 96.
temperature of solid–gas interface; the air flow rate enhance less [14] M. Srinivas, L.M. Patnaik, IEEE Trans. Syst. Man Cybernet. 24 (1994) 656.

Please cite this article in press as: S. Curteanu, et al., J. Ind. Eng. Chem. (2013), http://dx.doi.org/10.1016/j.jiec.2013.08.007