Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
1
An integrated model to predict the size of silver nanoparticles in
montmorillonite/chitosan bionanocomposites: a hybrid of data
envelopment analysis and genetic programming approach Elham Sarvestani1; Gholam Reza Khayati*2
1. Department of Materials Science and Engineering, Shahid Bahonar University of Kerman, P. O. Box No. 76135-133, Kerman, Iran;
[email protected] 2. Department of Materials Science and Engineering, Shahid Bahonar University of Kerman, P. O. Box No. 76135-133, Kerman, Iran;
[email protected];098-09151903477
Abstract
Unique chemical and physical properties of silver nanoparticles (AgNPs) enhances its usages in
various categories such as medical utilities. Due to the high dependency of AgNPs properties to
the size, this study is an attempt to employ gene expression programing (GEP) for constructing a
quantitative model for estimating the size of AgNPs in montmorillonite/chitosan
bionanocomposites that prepared by chemical approach. Generalization capabilities, fault
tolerance, noise tolerance, high parallelism, nonlinearity and significant information processing
characteristics are the main advantages of GEP. Accordingly, the practical parameters including
reaction temperature, AgNO3 concentration, weight of montmorillonite in aqueous
AgNO3/chitosan solution (WMMT) and the percentage of chitosan are selected as input parameters
through GEP modeling. The accuracy of proposed models are investigated by statistical
indicators including mean absolute percentage error (MAPE), root relative squared error
(RRSE), root mean square error (RMSE) and correlation coefficient (R2). Finally, the best model
is selected by R2 = 0.987, RMSE = 0.100, RRSE = 0.146 and MAPE = 0.221. The sensitivity
analysis confirmed that the percentage of chitosan, concentration of AgNO3, WMMT and reaction
temperature are the most effecting parameters on the size of AgNPs, respectively.
Keywords
Silver nanoparticle; Gene expression programming; Montmorillonite; Chitosan;
Bionanocomposites; Sensitivity analysis
NOMENCLATURE
MAgNO3 Concentration of silver nitrate (M) MAPE Mean absolute percentage error
mailto:[email protected]
2
T Reaction temperature (C̊) C(i,j) Value returned by the individual
chromosome i for fitness case j
Cts Percentage of chitosan (w/w %) Si Sensitivity level of an input parameter
(%)
MMT montmorillonite N (=
30)
Number of datasets used for sensitivity
test
R2 Correlation coefficient ti The measured values by models
RMSE Root mean square error pi The predicted values by models
RRSE Root relative squared error
1. Introduction
Recently, nanotechnology is a hot issue for many researches in materials science due its
administrated effect on the chemical and physical properties [1]. In this regards, the
preparation of noble metal nanoparticles, e.g. Ag, Au, Pd and Pt, are the subject of various
researches [2]. Generally, there are three methods for the preparation of nanoparticles, i.e.,
chemical, physical and biological approaches. The unique potential of chemical methods for
the preparation of finer particles as well as the narrower distribution size of products
distinguishes this technique from the others. Application of metal nanoparticles play a key
rolls in the enhancement of the performance of nanocomposites with the extensive utilities in
industrials. Silver nanoparticles (AgNPs) have extensive usages in the fields of catalysis,
optics, dentistry, photography, mirrors, clothing, food industries and electronics [3].
Advanced applications of AgNPs cause to the evolution of various techniques by the
preference of the lower size as criterion. Accordingly, the preparation of AgNPs with the
smaller size as well as the lower possibility of aggregation are favorable [1].
Other challenge of the preparation of nanoparticles is their strong tendency for
agglomeration [4]. There are various approaches for decreasing the agglomeration of
nanoparticles such as the usages of layered silicates of montmorillonite (MMT) as matrix.
MMT has two-to-one layered network including one octahedral aluminum layer between the
two layers of silicon tetrahedral. The thickness of this layer is about 1 nm and the lateral
dimension of this layer is in the range of 100–1000 nm [5]. Also, MMT has outstand
3
characteristics including the ion exchange properties, swelling and intercalation. The lamellar
structure of MMT provides the possibility of its delamination to the elemental sheets without
difficulty and usages as substrate for the synthesis of nanoparticles by electroless technique.
For example the structure of MMT employs as an appropriate substrate for the preparation of
biomaterial nanoparticles through the adsorption of cationic ions and anchoring the transition
metal complexes [6].
Chitosan (Cts), i.e., a natural polymers, has unique characteristics including non-toxicity,
solubility in aqueous medium, biodegradability, excellent biocompatibility, multi-functional
groups, artificial skin and bone substitutes [7]. Cts is added to intercalate with WMMT by the
mechanism of hydrogen bonding processes as well as cationic exchange and produces bio
nanocomposites (BNCs) with functional and structural properties [8]. BNCs made of
organic/inorganic nano size filler and polymeric matrix. To the best of our knowledge,
metal/clay/polymer compounds, e.g. BNCs is a hot issues for many researches due to their
excellent properties [7, 9-11].
Gene expression programming (GEP) is used to determine the performance and
properties of engineering materials as a mathematical expression. Respect to the classical
techniques like, artificially neural network and regression analysis GEP provides a capable
environment. Compared to the classical approach, there is not any predefined function
through the GEP modeling [12, 13].
In summary, lower particle size of AgNPs induced the synergetic effects on the
biological properties of prepared bio-nanocomposites. However, there is a general trend to
the agglomeration in the range of lower size of AgNPs. Accordingly, the main motivation
behind of this study are: (i) establishing the relationship among the practical variables
including the concentration of AgNO3, chitosan percentage, d-spacing of clay layers and
4
reaction temperature as input values and the size of AgNPs as output value; (ii) use of GEP
for modeling; (iii) determination the effect of each variable quantitatively on the size of
AgNPs; and (iv) employment of MMT as substrate to avoid the agglomeration.
2. Data collection and analysis
2.1. Data collection
Correct selection of practical variables that affected the output is the first step of GEP
modeling. Shabanzadeh et al. [14] investigated the size of AgNPs prepared in
montmorillonite/chitosan bionanocomposites. They showed that the size of AgNPs is
strongly depended on the concentration of AgNO3, chitosan percentage, d-spacing of clay
layers and reaction temperature. Our study uses 30 dataset from literature [14] as shown in
Table 1. The dataset are randomly divided to the training (20 data) and testing (10 data)
datasets [15].
Table 1.
2.2. Data analysis
Obviously, outlier data complicate the clear understanding of AgNPs dependency to the
practical parameters. Accordingly, determination and removing of outlier is beneficially for
providing of higher accuracy through the GEP modeling [16]. Box plot is a data analysis
procedure for determination of outlier during the practical dataset. Fig. 1 shows the box plot
of 4 practical parameters during the synthesis of AgNPs. As shown, the median of Cts and
MAgNO3 data sets are in the box center, indicating the symmetric distribution of these data.
However, the box plot of T and WMMT are skewed to the bottom and top, respectively. In
other words, the box shifted to the bottom whisker and most data are small with some minor
exceptionally large ones. The average of T and WMMT are higher than the median and is close
to the average of T (40) and WMMT (1.99). It is necessary to note that, there aren't any outlier
5
between the selected dataset for all practical parameters. Table 2 summarizes the statistical
explanation of dataset in this study.
Table 2.
Fig. 1.
Another key parameter during the GEP modeling is the independency of practical
parameters to each other. In this study, uses bivariate correlation analysis (BCA) to find the
relationships between the input parameters. BCA, at first, carried out a comprehensive
analysis to find the highly correlated pairs. Low negative/positive correlation between the
pairs is essentially during GEP modeling. In other words, significant interdependency among
the variables may be exaggerated the effect of input data during the modeling as multi
dependency [17]. Table 3 shows the amount of correlation coefficient of experimentally
variables. The main point of Table 3 is the presence of considerable dependency between
MAgNO3 with T, MAgNO3 with Cts, MAgNO3 with WMMT, T with Cts, T with WMMT, Cts with
WMMT.
Table 3.
Principal component analysis (PCA) is suitable approach for investigation the multi
dependency among the experimentally variables. This technique enables us to eliminate the
correlation between variables from a multi-dimensional space to a lower dimension space.
The variables in new space are uncorrelated [18-20]. Since, the presence of a higher
correlation coefficients between the variable is not essentially depicts the multi dependency
of dataset, to assure from possibility of PCA, at first Kaiser Mayer Olkin (KMO) [21] must be
estimated using equation 1 as criterion.
KMO =r
ij
2åår
ij
2åå + aij2åå
(1)
6
In the case that KMO is lower than 0.7, the dependency is unreal and must be neglected
[16]. Since, KMO factor of current study is equal to 0.516, there is no need to use PCA for
removing of multi dependency. In other words, high dependency between some variables in
Table 3 related to the nature of BCA. Accordingly, the selected practical parameters are
independent and as a consequence appropriate for further analysis by GEP.
3. Explanation of predictive equation
3.1. Gene expression programing (GEP)
Ferreira in 2011 proposes a novel approach by consideration of the population based on
the evolutionary strategy to compensate the disadvantageous of genetic algorithm (GA) and
genetic programing (GP) as gene expression programing (GEP) [22, 23]. GEP provides an
advanced strategy for the construction of inferring function among the input parameters with
the capability of forecasting the output by acceptable accuracy as well as minimum
estimation error. The main advantages of GEP are the possibility of finding the best
mathematical equation and its ability for checking the entire search space without trapping in
local minimum. In other words, GEP employed the advantages of GA and GP to provide
better modeling [24, 25]. Like GA, the individuals of GEP have fixed length string characters
(chromosomes) that can be connected each other by Karva language. This language enables
GEP to convert the coded programs to chromosomes. Generally, the coded solution can be
expressed as expression tree (ET) structure. GEP contains of chromosome with various genes
that constructs from 2 sections including tail and head (Fig. 2).
Fig. 2.
The head in Fig. 2 determines the functions and terminals, while the tail just determines
the terminals. It is necessary to note that, the terminals are input variable and constant value
while, the functions include the basic mathematically element including (*, /, +, -), boolean
operators (and, or, nor, not) and nonlinear functions (Exp, 3Rt, sqrt, arctan, tan, sin, -cos).
7
The head length, is predefined by GEP user while, the tail length, is dependent to h. nmax (the
maximum number of argument) is correlated to the function as equation 2.
t = h(n
max-1) +1
(2)
The chromosome of GEP is made of random; illustrate by a subset of sub-ET that
correlates to the other using linking function (*, /, +, -) [26-29]. As shown in Fig. 3, GEP
starts by generation of chromosomes in a random way and provide the first population as
Karva language (K-Expression). At following, each chromosomes depicted is ET and
mathematical equation, respectively. By consideration of fitness function as criterion,
evaluating the cost of individuals. The best ones selected for the next generation. If f (Si (t))
defined as fitness of individual Si in the population at t generation, by consideration of the
fitness-proportionate selection, the probability of individual Si will be copied to the next
population generation as a result of any one reproduction operation (equation 3) [30]:
f (Si(t))
f (Si(t))
j=1
M
å (3)
After the estimation of probability values of each chromosomes, the selection process
continues by selection as shown in Fig. 3. In this step, by consideration of fitness value as
criterion, the individuals select as candidate for selection. At the next phase, genetic operator
(transposition, recombination and mutation inversion) utilizes to the chromosomes to
generate the better individuals at future generation. The same process is keep on till and
suitable solution is found [31-34]. At following the genetic operators explains in summary:
I. Mutation
The main characteristics of mutation is its intrinsic modification power which are able to
perform anywhere in the chromosome length. Conversion a terminal to the other terminal is
8
performed in the tail by mutation operator while, in the head mutation provide the possibility
of the changes in functions or terminals each other [22, 30, 35].
II. Inversion
This operator is active through the chromosomes head and enables the user to reverse the
length of 1 to 3 in head.
III. Transposition
This operator has three types including: (i) insertion sequence transposition; i.e.,
responsible for the transportation of a fragment to the head of its own or other genes; (ii) root
insertion sequence transposition; i.e., responsible for the transposition of a segment with a
function in the first position of the root of the genes; (iii) gene transposition; i.e. responsible
to transpose all of genes at first chromosomes [22, 30, 35].
Fig. 3.
3.2. Training GEP as an intelligent approach for modeling and obtaining the
mathematical equation
The aim of this study is to propose a mathematical equation as GEP model (GEP-1 to
GEP-10) for prediction of the size of AgNPs. Table 4 summarizes the inferring equation of
GEP-1 to GEP-10 that have the better performance. The number of genes is supposed to eight
(sub-ETs), the multiplication and addition randomly, selects as correlating function. During
the testing and training phases of GEP models uses 10 various sub-ETs MAgNO3, T, Cts and
WMMT as input parameters and AgNPs as output value. Between 30 collected data sets, 20
trails are selected as training data set and 10 trails are chosen for testing phase. The fitness
function, fi, of an individual program, estimates from equation 4:
fi=
1
N
ti- p
i
ti
i=1
i=N
å ´100 (4)
9
In the case of the precision |C(ij)-Tj| ≤ 0.01, the precision = 0 and i = fmax = MCt. While,
for fmax = 1000, M = 100. Such approach provide the possibility of finding the optimum
condition [36, 37].
Afterwards, the set of terminals (T) and the set of functions (F) for generation of
chromosomes are preferred. The former including MAgNO3, T, Cts, and WMMT while, the later
including four basic arithmetic operators (-, +, /, *) and some basic mathematical functions (-,
+, /, *, 3Rt, Inv, Tanh, Arctan, Cos, Sin, Log, Ln, Exp, X2) are utilized. Simultaneously, the
performance of proposed GEP models are monitored during the testing and training phases.
Selection of the chromosomal tree is another step of GEP. In this study at first, initially
considered the single gene and two lengths of heads as first approximation and then the
amount of each ones are increased during each runs. Table 5 abbreviates the investigated
characteristics during the GEP modeling in current study and the size of AgNPs is a function
of MAgNO3, T, Cts and WMMT [36].
Table 4.
Table 5.
3.3. Evaluating the training performance of the model
To find the approximation performance of GEP equations, the real output and predicted
output compare by consideration of statistical performance indices including mean absolute
percentage error (MAPE), root relative squared error (RRSE), root mean square error (RMSE)
and correlation coefficient (R2) as criteria (equations 5-8) [38, 39].
MAPE =1
N
ti- p
i
tii=1
i=N
å ´100 (5)
RRSE =
(ti- p
i)2
i=1
i=N
å
ti- (
1
N) t
ii=1
i=N
å )2å
èç
å
å÷i=1
i=N
å
(6)
10
RMSE =(t
i- p
i)2
Ni=1
i=N
å (7)
R2 = 1-
(ti- p
i)2
i=1
N
å
pi
2
i=1
N
å (8)
Table 6 summarizes the values of these criteria for GEP models.
Table 6.
3.4. Sensitivity analysis
The significance of each practical parameters and relative effects on the amount of
AgNPs size may be investigated by sensitivity analysis (equation 9). Accordingly, each
parameter varies between the maximum and minimum of its level while, the other practical
variable suppose to be constant at their mean values. Then, the changes of output are
measured and uses as threshold for comparing the effect of each parameter. Typically, for
determination of WMMT changes on the size of AgNPs, all practical variables (except WMMT)
are supposed to be fixed in their mean values while, the WMMT changes [39]. To do the
sensitivity analysis a step-by-step methodology is carried out on most appropriate GEP model
by changing of input parameters, one at a time in a constant rates [15].
Si(%) =
1
N
%change inoutput
%change in input
å
èçå
å÷j=1
N
åj
(9)
4. Results and discussion
In current study, R2 (which closer to 1 is better), RMSE (which closer to 0 is better),
RRSE (which closer to 0 is better) and MAPE (which closer to 0 is better) criteria are utilized
to evaluate the accuracy of proposed GEP models. As shown in Table 6, 0.945 < R2 < 0.995;
0.073 < RMSE < 0.319; 0.039 < RRSE < 0.219 and 0.221 < MAPE < 3.039. Fig. 4 shows
the changes of selected statistical indices for 10 most appropriate models. Accordingly, the
high value of R2 belong to GEP-5 model and equal to the 0.995 and 0.985 for the training and
11
testing phases, respectively. If uses RMSE as metric for GEP validation, the lower RMSE
belongs to GEP-3 due to the closer value to zero equal to 0.067 and 0.100 during the training
and testing phases. Similar selection GEP model by utilization of RRSE reveals the GEP-3
with RRSE equal to 0.125 for training and 0.146 for testing phases has the best performances
and in the case of MAPE uses as criteria, GEP-3 selected as the most appropriate model.
Consequently, using of one statistical approach can't be satisfied simultaneously the lower
values of errors (RMSE, RRSE, MAPE) and higher values of R2. Since, to contribute all
statistical indices for the selection of most appropriate GEP models defined Fitness value as
equation 10.
Fitness - value =1/ R2 + RMSE + RRSE + MAPE (10)
Fig. 4.
In this approach, the most appropriate model has the lower fitness value (Eq. 10). Fig. 5
shows the change of fitness value for various GEP models. Accordingly, the best structure
belongs to GEP-3 with 3 genes, 8 head and 30 chromosomes. In which, the genes linked each
other using the addition function.
Fig. 5.
Figure 6 shows the expression tree of GEP-3. The inferring equation between the
practical parameters and Ag particles size of GEP-3 model in Table 6 reveals the complexity
of practical parameters on output.
Fig. 6.
Figure 7 compares the practical and the predicted values of the size of AgNPs. In which,
the experiment number is defined by the number in the circle and hexagonal and the number
in vertical axis are the Ag nanoparticles size. The values of statistical indices as well as
fitness value are added to Fig. 7. As shown, GEP-3 provides the acceptable accuracy during
the testing and training set, respectively.
12
Fig. 7.
Figure 8 shows the effect of experimentally parameters on outputs as 3D diagrams. Fig.
8(a) in which the variation of AgNPs size drawn as a function of MAgNO3 and Cts confirm the
significant effect of these parameters. Fig. 8(b), in which MAgNO3 and T select as variables,
shows the relatively higher effect of MAgNO3 respect to T on the output. The yellow and red
points inside of Fig. 8(c) shows MAgNO3 is higher effective respect to WMMT. Also, the output
value changes directly by increasing of MAgNO3 and WMMT. Analysis of Fig. 8(d) reveals that
the effect of Cts on the output is higher than T. Also, WMMT factor has higher effect on
AgNPs size respect to the T (Fig. 8(e)). Simultaneously, increasing of WMMT and Cts could
be increased the size of output (Fig. 8(f)).
Fig. 8.
Generally, the prediction of AgNPs by consideration of the effect of four variables,
including Cts, MAgNO3, WMMT and T are studied. As expected, the size of AgNPs increases by
increasing of AgNO3 concentration, temperature of reaction and amount of MMT. However,
the influence of temperature of reaction is very low on the enhancement of the size of AgNPs
and with the increases of chitosan percentages, the AgNPs size decreases. However, with the
simultaneous increases of chitosan with other parameters, the size of the AgNPs increases. As
shown in Fig. 9, the important factors in the size of AgNPs are the Cts, MAgNO3, WMMT and T
of the reaction, respectively.
Fig. 9.
5. Conclusion
In the present study, a gene expression programming is utilized to forecast the size of AgNPs
by consideration of the effect of MAgNO3, T, Cts and the WMMT as input variables. The results
confirm the GEP structure with 3 genes, 8 head and 30 chromosomes has the better behavior
during the training algorithm with acceptable performance during GEP molding of prepared
13
bionanocomposites. As a results, GEP-3 model with R2=0.977, RRSE=0.146, RMSE=0.100
and MAPE=0.221 has been suggested for the estimation of the size of AgNPs prepare at
various practical parameter. The results confirm that by increasing of the amount of Ag+ in
concentration and temperature, the size of AgNPs would be increased. While, at a higher
chitosan percentage, the size of output decreased. Also, GEP models provide a suitable
approach for the prediction of the size of AgNPs.
6. References
1. Shabanzadeh, P., Yusof, R. and Shameli, K. "Artificial neural network for modeling the
size of silver nanoparticles’ prepared in montmorillonite/starch bionanocomposites",
Journal of Industrial and Engineering Chemistry, 24, pp. 42-50 (2013).
2. Veisi, H., Azizi, S. and Mohammadi, P., "Green synthesis of the silver nanoparticles
mediated by Thymbra spicata extract and its application as a heterogeneous and
recyclable nanocatalyst for catalytic reduction of a variety of dyes in water", Journal of
Cleaner Production, 170:, pp. 1536-1543 (2016).
3. Rai, M., Yadav, A. and Gade, A., "Silver nanoparticles as a new generation of
antimicrobials". Biotechnology advances, 27(1): pp. 76-83 (2009).
4. Shameli, K., et al., "Green biosynthesis of silver nanoparticles using Callicarpa maingayi
stem bark extraction", Molecules, 17(7), pp. 8506-8517 (2012).
5. Shameli, K., et al., "Synthesis of silver/montmorillonite nanocomposites using γ-
irradiation", International journal of nanomedicine, 5, pp. 1067 (2010).
6. Makwana, D., et al., "Characterization of Agar-CMC/Ag-MMT nanocomposite and
evaluation of antibacterial and mechanical properties for packaging applications",
Arabian Journal of Chemistry, 13(1), pp. 3092-3099 (2018).
7. Lavorgna, M., et al., "MMT-supported Ag nanoparticles for chitosan nanocomposites:
structural properties and antibacterial activity", Carbohydrate polymers, 102: pp. 385-392
(2014).
8. Usman, M.S., et al., "Copper nanoparticles mediated by chitosan: synthesis and
characterization via chemical methods", Molecules, 17(12): pp. 14928-14936 (2012).
9. Shameli, K., et al., "Synthesis and characterization of silver/montmorillonite/chitosan
bionanocomposites by chemical reduction method and their antibacterial activity".
International journal of nanomedicine, 6: pp. 271 (2011).
10. Ahmad, M.B., et al., "Antibacterial effect of silver nanoparticles prepared in bipolymers
at moderate temperature", Research on Chemical Intermediates, 40(2): pp. 817-832
(2014).
https://www.sciencedirect.com/science/journal/18785352/13/1
14
11. Shameli, K., et al., "Green synthesis of silver/montmorillonite/chitosan
bionanocomposites using the UV irradiation method and evaluation of antibacterial
activity", International journal of nanomedicine, 5, pp. 5:875 (2010).
12. Jafari, M.M. and Khayati, G.R., "Prediction of hydroxyapatite crystallite size prepared by
sol–gel route: gene expression programming approach", Journal of Sol-Gel Science and
Technology, 86(1), pp. 112-125 (2018).
Ebrahimzade, H., Khayati, G.R. and Schaffie, M., "A novel predictive model for
estimation of cobalt leaching from waste Li-ion batteries: Application of genetic
programming for design", Journal of Environmental Chemical Engineering, 6(4), pp. 3999-4007 (2018)
13. Shabanzadeh, P., et al., "Prediction of silver nanoparticles’ diameter in
montmorillonite/chitosan bionanocomposites by using artificial neural networks",
Research on Chemical Intermediates, 41(5), pp. 3275-3287 (2015).
14. Khayati, G.R., "Adaptive neuro-fuzzy inference system and neural network in predicting
the size of monodisperse silica and process optimization via simulated annealing
algorithm", Journal of Ultrafine Grained and Nanostructured Materials, 51(1): pp. 43-52
(2018).
15. Faradonbeh, R.S. and Monjezi, M., "Prediction and minimization of blast-induced ground
vibration using two robust meta-heuristic algorithms", Engineering with Computers,
33(4): pp. 835-851 (2017).
16. Sun, M., "Prediction of viscosity of branched alkanes using gene expression
programing."Petroleum Science and Technology, 36(23): pp. 2049-2056 (2018).
17. Tiryaki, B., "Predicting intact rock strength for mechanical excavation using multivariate
statistics, artificial neural networks, and regression trees", Engineering Geology, 99(1-2):
pp. 51-60 (2008).
18. Sayadi, A.R., Khalesi, M.R. and Borji, M.K., "A parametric cost model for mineral
grinding mills". Minerals Engineering, 55, pp. 96-102 (2014).
19. Sayadi, A.R., Lashgari, A. and Paraszczak, J.J., "Hard-rock LHD cost estimation using
single and multiple regressions based on principal component analysis", Tunnelling and
underground space technology, 27(1), pp. 133-141 (2012).
20. Kaiser, H.F., "An index of factorial simplicity", Psychometrika, 39(1), pp. 31-36 (1974).
21. Alanni, R., et al., "New gene selection method using gene expression programing
approach on microarray data sets." International Conference on Computer and
Information Science. Springer, Cham, (2018).
22. Dikmen, E., "Gene expression programming strategy for estimation performance of LiBr–
H 2 O absorption cooling system." Neural Computing and Applications, 26(2), pp.409-
415 (2015).
23. Steeb, W.-H., "The Nonlinear Workbook: Chaos, Fractals, Cellular Automata, Neural
Networks, Genetic Algorithms, Fuzzy Logic with C++, Java, Symbolicc++ and Reduce
Programs". World Scientific, (2001).
https://www.sciencedirect.com/science/journal/22133437/6/4
15
24. Brownlee, J., "Clever algorithms: nature-inspired programming recipes", Jason Brownlee,
2011.
25. Monjezi, M., et al., "Modification and prediction of blast-induced ground vibrations
based on both empirical and computational techniques". Engineering with Computers,
32(4), pp. 717-728 (2016).
26. Faradonbeh, R.S., et al., "Prediction of ground vibration due to quarry blasting based on
gene expression programming: a new model for peak particle velocity prediction".
International journal of environmental science and technology, 16(6), pp. 1453-1464
(2016).
27. Khandelwal, M., et al., "A new model based on gene expression programming to estimate
air flow in a single rock joint", Environmental Earth Sciences, 75(9), pp. 739 (2016).
28. Aval, S. B., Ketabdari, H. and Gharebaghi, S. A., "Estimating shear strength of short
rectangular reinforced concrete columns using nonlinear regression and gene expression
programming", Structures, Elsevier, 12: (2017).
29. Ferreira, C., "Gene expression programming: mathematical modeling by an artificial
intelligence", Springer, 21, (2006).
30. Jafari, M. M., and Khayati, G. R., "Prediction of hydroxyapatite crystallite size prepared
by sol–gel route: gene expression programming approach." Journal of Sol-Gel Science
and Technology, 86(1), pp. 112-125 (2018).
31. Zhong, J., Feng, L., and Ong, Y. S., "Gene expression programming: A survey." IEEE
Computational Intelligence Magazine, 12(3), pp. 54-72 (2017).
32. Faradonbeh, R.S., "Monjezi, M. and Armaghani, D.J., Genetic programing and non-linear
multiple regression techniques to predict backbreak in blasting operation", Engineering
with Computers, 32(1), pp. 123-133 (2016).
33. Baykasoğlu, A., et al., "Prediction of compressive and tensile strength of limestone via
genetic programming". Expert Systems with Applications, 35(1-2), pp. 111-123 (2008).
34. Ferreira, C. and Gepsoft, U., "What is gene expression programming". (2008).
35. Sarıdemir, M., "Effect of specimen size and shape on compressive strength of concrete
containing fly ash: Application of genetic programming for design", Materials & Design
(1980-2015), 56, pp. 297-304 (2014).
36. Jafari, S., and Mahini, S. S., "Lightweight concrete design using gene expression
programing", Construction and Building Materials 139, pp. 93-100 (2017).
37. Fallahpour, A., et al., "An integrated model for green supplier selection under fuzzy
environment: application of data envelopment analysis and genetic programming
approach", Neural Computing and Applications, 27(3), pp. 707-725 (2016).
38. Khozani, Z. S., Bonakdari, S., and Ebtehaj, I., "An analysis of shear stress distribution in
circular channels with sediment deposition based on Gene Expression
Programming." International Journal of Sediment Research, 32(4), pp. 575-584 (2017).
16
39. Hoseinian, F.S., et al., "Semi-autogenous mill power model development using gene
expression programming", Powder Technology, 308, pp. 61-69 (2017).
40.
41. Elham Sarvestani is received his M.Sc. in Materials Science Engineering from Shahid Bahonar
University of Kerman (Iran) in 2020 and B.Sc. Materials Science Engineering from Shahid
Bahonar University of Kerman (Iran) in 2017.
42. Gholam Reza Khayati is currently a faculty of Materials Science and Engineering at Shahid
Bahonar University of Kerman. He received his Ph.D. in materials science from Shiraz University
of Shiraz (Iran) and M.Sc. (honor student) in Materials Science and Engineering at Department of
Materials Science & Engineering, Tehran University (Iran). He received his B.Sc. in Materials
Science from Shahid Bahonar University of Kerman, Iran in 2004. He has published more than 40
scientific papers in the field of synthesis and characterization of nanostructures materials.
Figures caption:
Fig. 1. Boxplot for MAgNO3, T, Cts and WMMT variables during the green synthesis of AgNPs.
Fig. 2. Typically, illustration of Karva language and expression tree for a chromosome with
two genes.
Fig. 3. Flowchart of GEP strategy.
Fig. 4. Comparison of validation criteria for GEP models (a) R2, (b) RMSE, (c) RRSE and (d)
MAPE for GEP-3 structure.
Fig. 5. The changes of fitness value for 10 GEP models.
Fig. 6. Expression tree of GEP-3 with 3 genes to forecast the average particle size of AgNPs.
Fig. 7. Comparison of predicted values by GEP-3 and actual data, (a) training data sets and
(b) testing data sets.
Fig. 8. Three-dimensional surfaces plots: (a) the effects of MAgNO3 and Cts (w/w%), (b) the
effects of MAgNO3 and T, (c) the effects of MAgNO3 and WMMT, (d) the effects of Cts (w/w%)
and T, (e) the effects of T and WMMT, (f) the effects of Cts (w/w%) and WMMT on size of
AgNPs.
Fig. 9. Sensitivity analysis of the practical parameter during the preparation of AgNPs.
17
Tables caption:
Table 1. Illustration of practical data for the preparation of AgNPs by green synthesis.
Table 2. Statistical distribution of MAgNO3, T, Cts and WMMT in development of GEP models.
Table 3. Correlation coefficients among MAgNO3, T, Cts and WMMT for preparation of AgNPs
by green synthesis.
Table 4. Representation of inferring equations for different GEP models.
Table 5. The most appropriate characteristics of GEP algorithm in current study.
Table 6. Various statistical criteria for most 10 appropriate GEP models.
18
Table 1. Illustration of practical data for the preparation of AgNPs by green synthesis
No. Factors Response
M AgNO3
(M)
T
(̊C)
Cts
(w/w %)
WMMT
(gr) AgNPs size (nm)
1 0.5 25 0.25 1.34 3.51
2 0.5 30 0.25 1.35 3.54
3 0.5 40 0.25 1.39 3.58
4 0.5 50 0.25 1.40 3.62
5 0.5 60 0.25 1.41 3.63
6 1.0 25 0.5 1.55 3.85
7 1.0 30 0.5 1.56 3.87
8 1.0 40 0.5 1.60 3.94
9 1.0 60 0.5 1.69 4.08
10 1.5 30 1 1.93 4.14
11 1.5 35 1 1.97 4.15
12 1.5 40 1 2.11 4.18
13 1.5 50 1 2.16 4.21
14 2.0 25 1 2.29 4.25
15 2.0 30 1.5 2.08 4.52
16 2.0 60 1.5 2.33 5.14
17 5.0 30 2 2.17 4.87
18 5.0 40 2 2.25 5.32
19 5.0 50 2 2.35 5.59
20 5.0 60 2 2.38 5.632
21 1.5 60 1 2.26 4.22
22 2 35 1.5 2.12 4.60
23 2 40 1.5 2.18 4.64
24 2 50 1.5 2.29 4.69
25 5 35 2 2.20 4.99
26 0.5 35 0.25 1.37 3.55
27 1 35 0.5 1.58 3.89
28 1 50 0.5 1.64 3.97
29 1.5 25 1 1.88 4.11
30 5 25 2 2.14 4.76
19
Table 2. Statistical distribution of MAgNO3, T, Cts and WMMT in development of GEP models
Variable Min. Max. Mean Median Std.
MAgNO3 0.50 5.00 2.00 1.50 1.61
T 25.00 60.00 40.00 37.50 12.11
Cts 0.25 2.00 1.04 1.00 0.65
WMMT 1.34 2.38 1.99 21.02 0.36
Table 3. Correlation coefficients among MAgNO3, T, Cts and WMMT for preparation of AgNPs
by green synthesis
Variable MAgNO3 T Cts WMMT
MAgNO3 1 0 0.921 0.723
T 0 1 0.033 0.193
Cts 0.921 0.033 1 0.891
WMMT 0.723 0.193 0.891 1
20
Table 4. Representation of inferring equations for different GEP models
Models Predicted equation
GEP-1
Y = ((((WMMT* MAgNO3)* T)*( 0.034*0.034))+(( 0.034*WMMT)*(
MAgNO3+Cts)))+((((Cts -WMMT )*(WMMT* 0.074))*((WMMT*WMMT )-
MAgNO3))- T)+((MAgNO3/(((-2.33)+Cts)* T))-(((-2.33)- WMMT )- T))
GEP-2
Y = ((((MAgNO3-Cts)*(0.006*T))+((0.006/Cts )+ WMMT))/WMMT
)*((((WMMT +WMMT )*Cts)/((-0.446)-WMMT))+((WMMT+MAgNO3)/(WMMT*(-
0.446))))*((((0.587-Cts)*(0.587/T))/((MAgNO3/0.587)-WMMT))+0.587) *
((WMMT + WMMT)/(((WMMT /T)-(-0.220))-(MAgNO3+WMMT)))
GEP-3
Y = (((((((((((-0.126)-Cts)+WMMT)*MAgNO3)*
MAgNO3)))^(1/3))))^(1/3))+(1.0/((((WMMT+(-0.830))-
tanh(MAgNO3))+(WMMT+Cts)/(((T))^(1/3))))))+(d(4)+((((-0.834)*Cts)-
tanh((-0.834)))*(tanh(MAgNO3)-WMMT)))
GEP-4
Y = ((Atan((T*(-0.432)))+Atan(WMMT))*((MAgNO3*(-0.432))+ (-
0.432)))+exp(Atan(((exp(MAgNO3)+(Cts/WMMT))+(0.150/Cts))))+(((-
0.007)+ MAgNO3)/(((T+WMMT)*(-0.007))*(WMMT+T)))
GEP-5
Y = (Log((sin(sin(sin(MAgNO3)))+(Log(((MAgNO3*(-
1.020))+T))/Log(10))))/Log(10))+sin((((WMMT*Cts)-((-
0.304)/WMMT))*(((-0.304)+WMMT)*sin((-
0.304)))))+(T+((sin(WMMT)+(WMMT-T))+((0.891*Cts)+0.891)))
GEP-6
Y = ((cos((WMMT+WMMT))^2)-((Cts*WMMT)-(0.006+0.006))) +
((WMMT^2)+(((MAgNO3/0.321)*(Cts-MAgNO3))/(T*Cts)))+(cos((cos(WMMT)-
((-0.530)^2)))-((Cts^2)*((-0.530)+(-0.530)))) +
(((cos(WMMT)/Cts)+(Cts/T))*(( WMMT-MAgNO3)*(0.200*WMMT)))
GEP-7
Y = ((((((((T))^(1/3))+(1.176-WMMT))*((1.176+1.176)+(Cts*
MAgNO3)))))^(1/3))+((((((0.607+Cts)+(MAgNO3+MAgNO3))+((0.607+
MAgNO3)+(0.607+MAgNO3)))))^(1/3))+((((MAgNO3-
WMMT)*Cts)*MAgNO3)/(((((-0.682)))^(1/3))*(T+MAgNO3)))
GEP-8
Y = ((0.520*Cts)+((((0.520-MAgNO3*Cts))))^(1/3)))+((((((((-
1.110)))^(1/3))^3)+(((MAgNO3))^(1/3)))*((((((T))^(1/3))))^(1/3)))^3)+((((((
Cts))^(1/3))/T)+( 0.333*Cts))+((((0.333))^(1/3))+(Cts+WMMT)))+(0.391-
((( MAgNO3^3)/(WMMT*T))/((Cts^3)- MAgNO3)))
GEP-9
Y = ((1.0/((((Cts-WMMT)*(T/(-0.333)))+((-0.333)-
T))))+WMMT)+(1.0/((WMMT+(((WMMT+(-0.970))/((-
0.970)+Cts))*(WMMT/T)))))+((tanh((WMMT*MAgNO3))-((Cts*(-0.317))/(
MAgNO3+(-0.317))))-(-0.317))+(((-0.185)/((T/ MAgNO3)-WMMT))-((MAgNO3*(-
0.185))-(T-T)))
GEP-10 Y = ((((((Cts*WMMT)/(((0.369))^(1/3)))+(
MAgNO3*Ln(T)))))^(1/3))+Ln(Ln((T+(((1.685^2)^2)-(
MAgNO3^2)))))+(((((((((Cts^2)-(WMMT-(-
21
0.209)))+(MAgNO3/Cts))))^(1/3))))^(1/3))
Table 5. The most appropriate characteristics of GEP algorithm in current study
GEP parameters GEP-1 to GEP-10
Number of runs 100
Parameters Value
Used functions *, /, +, -, 3Rt, Inv, Tanh, Arctan, Cos,
Sin, Log, Ln, Exp, X2
chromosomes Number 34 (GEP-2), 32 (GEP-8), 30 for other
models
size of Head 8
genes Number 4 (GEP-2), (GEP-8) and (GEP-9), 3
for other models
Linking function Addition , Multiplication
Fitness function error type Root relative squared error
Constant per gene 1
Mutation rate 0.0044
Inversion rate 0.1
One-point recombination rate 0.2
Two-point recombination rate 0.2
Gene recombination rate 0.1
Gene transportation rate 0.1
22
Table 6. Various statistical criteria for most 10 appropriate GEP models
Models Performance
indices R2 RMSE RRSE MAPE
GEP-1 Training set 0.992 0.046 0.086 0.989
Testing set 0.954 0.153 0.222 2.134
GEP-2 Training set 0.995 0.039 0.073 0.628
Testing set 0.945 0.160 0.233 2.215
GEP-3 Training set 0.985 0.067 0.125 1.238
Testing set 0.987 0.100 0.146 0.221
GEP-4 Training set 0.980 0.076 0.141 1.212
Testing set 0.953 0.168 0.244 2.092
GEP-5 Training set 0.995 0.040 0.074 0.564
Testing set 0.985 0.102 0.149 1.274
GEP-6 Training set 0.991 0.050 0.094 1.034
Testing set 0.969 0.129 0.188 2.089
GEP-7 Training set 0.990 0.052 0.097 0.916
Testing set 0.963 0.143 0.208 1.827
GEP-8 Training set 0.988 0.058 0.107 1.095
Testing set 0.982 0.113 0.164 1.554
GEP-9 Training set 0.987 0.062 0.114 1.090
Testing set 0.973 0.219 0.319 3.039
GEP-10 Training set 0.986 0.064 0.119 0.957
Testing set 0.980 0.129 0.188 1.714
23
Fig. 1. Boxplot for MAgNO3, T, Cts and WMMT variables during the green synthesis of AgNPs.
WMMT (gr)Cts (w/w%)T (̊C)MAgNO3 (M)
24
Fig. 2. Typically, illustration of Karva language and expression tree for a chromosome with
two genes.
Fig. 3. Flowchart of GEP strategy.
Sub-ET1: Sub-ET2:
(𝑎 + 𝑏) × 𝑎
𝑏
(𝑎 − 𝑏)
Gene1Gene2
Head Tail
Position
number
25
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
1 2 3 4 5 6 7 8 9 10
R2
GEP Model
Train Testa)
0
0.05
0.1
0.15
0.2
0.25
0.3
1 2 3 4 5 6 7 8 9 10
RM
SE
GEP Model
Train Testb)
0
0.1
0.2
0.3
0.4
0.5
1 2 3 4 5 6 7 8 9 10
RR
SE
GEP Model
Train Testc)
26
Fig. 4. Comparison of validation criteria for GEP models (a) R2, (b) RMSE, (c) RRSE and (d)
MAPE for GEP-3 structure.
Fig. 5. The changes of fitness value for 10 GEP models.
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
1 2 3 4 5 6 7 8 9 10
MA
PE
GEP Model
Train Testd)
0
1
2
3
4
5
6
7
8
1 2 3 4 5 6 7 8 9 10
Fit
nes
s V
alue
GEP Model
Train Test
27
Fig. 6. Expression tree of GEP-3 with 3 genes to forecast the average particle size of AgNPs.
Fig. 7. Comparison of predicted values by GEP-3 and actual data, (a) training data sets and
(b) testing data sets.
Sub-ET 1: Sub-ET 2: Sub-ET 3:
a) GEP- Training data
0123456
12
3
4
5
6
7
8
910
1112
13
14
15
16
17
18
1920
Target
Model
R2=0.985
RRSE=0.125
RMSE=0.067
MAPE=1.238
Fitness Value=2.445
b) GEP- Testing data
0123456
1
2
3
4
5
6
7
8
9
10
Target
Model
R2=0.987
RRSE=0.146
RMSE=0.100
MAPE=0.221
Fitness Value=1.48
28
Cts (w/w %)MAgNO3
T (̊C)MAgNO3
WMMT (gr) MAgNO3
29
Fig. 8. Three-dimensional surfaces plots: (a) the effects of MAgNO3 and Cts (w/w%), (b) the
effects of MAgNO3 and T, (c) the effects of MAgNO3 and WMMT, (d) the effects of Cts (w/w%)
T (̊C)Cts (w/w %)
WMMT (gr) T (̊C)
WMMT (gr) Cts (w/w %)
30
and T, (e) the effects of T and WMMT, (f) the effects of Cts (w/w%) and WMMT on size of
AgNPs.
Fig. 9. Sensitivity analysis of the practical parameter during the preparation of AgNPs.
0
20
40
60
80
100
120
MAgNO3 (M) T ( ̊C) Cts (w/w%) WMMT (gr)
Sen
siti
vit
y le
vel
(%
)
MAgNO3 (M)
5%-5%10%
-10%15%
-15%
WMMT (gr)