Accepted Manuscript

On the prediction of viscosity of Newtonian nanofluids

Ali Barati-Harooni, Adel Najafi-Marghmaleki, Armin Mohebbi,

Amir H. Mohammadi

PII: S0167-7322(16)33661-3
DOI: doi: 10.1016/j.molliq.2017.06.088
Reference: MOLLIQ 7534
To appear in: Journal of Molecular Liquids
Received date: 18 November 2016
Revised date: 15 June 2017
Accepted date: 17 June 2017

Please cite this article as: Ali Barati-Harooni, Adel Najafi-Marghmaleki, Armin Mohebbi,
Amir H. Mohammadi , On the prediction of viscosity of Newtonian nanofluids, Journal of
Molecular Liquids (2017), doi: 10.1016/j.molliq.2017.06.088

This is a PDF file of an unedited manuscript that has been accepted for publication. As
a service to our customers we are providing this early version of the manuscript. The
manuscript will undergo copyediting, typesetting, and review of the resulting proof before
it is published in its final form. Please note that during the production process errors may
be discovered which could affect the content, and all legal disclaimers that apply to the
journal pertain.
 ACCEPTED MANUSCRIPT

On the Prediction of Viscosity of Newtonian Nanofluids

Ali Barati-Harooni,a1 Adel Najafi-Marghmaleki,a Armin Mohebbi,b Amir H Mohammadic,d, 2

a
Young Researchers and Elite Club, Ahvaz Branch, Islamic Azad University, Ahvaz, Iran

T
b
Young Researchers and Elite Club, South Tehran Branch, Islamic Azad University, Tehran, Iran
c
Discipline of Chemical Engineering, School of Engineering, University of KwaZulu-Natal, Howard College

IP
Campus, King George V Avenue, Durban 4041, South Africa
d
Institut de Recherche en Gnie Chimique et Ptrolier (IRGCP), Paris Cedex, France

CR
US
AN
M
ED
PT
CE
AC

Corresponding author:
1
1. Email Address: alibarati.2012@yahoo.com (A. Barati-Harooni)
2
2. Email Address: a.h.m@irgcp.fr and amir_h_mohammadi@yahoo.com (A. H. Mohammadi)

1
 ACCEPTED MANUSCRIPT

Abstract - One of the important properties of nanofluids is their viscosity which is required in
various applications. Despite the importance of this property, there is no general model for
prediction of viscosity of nanofluids. Hence, developing accurate and general models for
prediction of this property is of great importance. The purpose of this work is to develop four
accurate and reliable models based on Multilayer Perceptron Neural Networks (MLP-NNs),
Least Square Support Vector Machine (LSSVM), Adaptive Neuro Fuzzy Inference System
(ANFIS) and Gene Expression Programming (GEP) technique for prediction of viscosity of

T
Newtonian nanofluids with different nanoparticles and different base fluids. An extensive

IP
number of 1140 data points of viscosity of nanofluids were utilized for development of the

CR
models. The predictions of the proposed models were examined with the predictions of various
techniques. Moreover, the accuracy of the proposed models was also examined by comparing

US
their predictions with several well-known literature correlations. Results show that the developed
models provide accurate and reliable predictions and outperform literature correlations and
AN
present better results. Moreover, the predictions of CSA-LSSVM model are better and accurate
than the predictions of other developed models.
M

Keywords - Nanofluid; Viscosity; Model; Prediction; Data.

ED
PT

1 Introduction
CE

Nanofluid is a low concentration mixture of small solid particles (10-100nm in size at

concentrations below 4%vol) and a base fluid that normally includes conventional cooling fluids
AC

[1]. Thermal conductivity of these novel fluids were found to be appreciably higher than the base

fluid value, so that even in the low concentration of dispersed particles, anomalous heat transfer

phenomena could be observed [2].

Viscosity is one of the most significant thermo physical properties of nanofluids, especially

in thermal applications that involve heat transfer and fluid flow. The amount of required

2
 ACCEPTED MANUSCRIPT

pumping power and the convective heat transfer coefficient are affected by the variation of

nanofluid viscosity. Therefore, in spite of the complexities in hydrodynamic and particleparticle

interactions of dispersed nanoparticles, precise prediction of the nanofluid viscosity is of great

importance [3]. Over the past two decades, several researches have been conducted on the

properties of nanofluids including synthesis of nanofluid [4, 5], measurement of thermal

T
conductivity [6, 7], improvement of thermal conductivity [8] and estimation of thermal

IP
conductivity [9]. However, despite the importance of nanofluid viscosity in practical usages,

CR
only few researches have been carried out in this area [4, 10]. First, viscosity of nanofluid that

contained Al2O3, SiO2, and TiO2 as nanoparticles dispersed in water was measured by Masuda et

US
al. [11]. In the next studies, the viscosity of different nanofluids was measured by several
AN
researchers [2, 5, 12-14]. In addition, the impact of various factors like shape, particle size,

volume fraction of nanoparticles and temperature over viscosity of nanofluid was studied [15].
M

According to evaluation of effective parameters, a comprehensive report was presented by

ED

Mahbubul et al. [16].

Process models allow generalizing experimental data and formulating results included in
PT

the processes [17]. Einstein model [18] nf = bf (1+2.5)) is the first nanofluid viscosity
CE

prediction model that is generally used at low volume fraction of nanoparticles (normally <0.02)

[19]. On the other hand, power law based models [20, 21] are more applicable for the prediction
AC

of the nanofluid effective viscosity at higher volume fraction (>0.02). Mehrabi [3] summarized

the most significant viscosity correlations of nanoparticles [3]. However, because of applying

some simple assumptions in development of these models, their reliability and validity are quite

limited.

3
 ACCEPTED MANUSCRIPT

Recently, soft computing techniques like neural networks, fuzzy logic, and genetic

programming have been used in various data processing applications without conducting

additional experimental tests [17, 22-24].

Neural networks [25] and Adaptive neuro-fuzzy inference system (ANFIS) [3] have been

employed to model the viscosity of nanofluids. However, the number of fluids and contained

T
nanoparticles and the range of input parameters used to drive former models are limited.

IP
Developing general models for precise prediction of nanofluid viscosity is the chief

CR
objective of the present study. In this work, four models based on MLP-NN, LSSVM, ANFIS

and GEP concepts are presented for prediction of viscosity of Newtonian nanofluids with

US
different nanoparticles and different base fluids by using 1140 data points collected from
AN
literature. The outcomes of the developed models are compared with the literature models and

correlations. Results show the accuracy, superiority and higher generalization ability of the
M

models presented in this work.

ED
PT

2 Theory
CE

2.1 Multilayer Perceptron Neural Networks (MLP-NNs)

AC

MLP-NNs are one of the well-known and applicable types of Artificial Neural Networks

(ANNs) which are applied in nonlinear regression and fitting problems. These networks have a

three layer structure. Each layer has several units known as neurons. These neurons are the main

processing parts of MLP-NN networks. The first layer is the input layer which accepts the input

data points [26]. The number of neurons of this layer is identical to the number of input

parameters of the model. The second layer is known as hidden layer. This layer could have

4
 ACCEPTED MANUSCRIPT

several sub layers. The number of neurons of this layer plays an important role in the prediction

capability of the model. Normally, the number of neurons of this layer is determined by using a

trial and error approach and by minimizing a cost function. The third layer is called output layer

which has a single neuron which delivers the output parameter of the model. The neurons of the

hidden layer contain a specific function which is called activation function or transfer function.

T
Each neuron accepts an input vector and multiplies it by a weight vector and sums the result with

IP
a constant bias term and delivers it to the transfer function [24, 26]. The outputs of transfer

CR
functions are summed up and are delivered to the output layer as the output parameter of the

model. There are several types of activation functions including the linear, tansig and

US
logsig types which are the most commonly applied ones. The weight terms in a MLP-NN
AN
model are optimized by an internal training algorithm such as Levenberg-Marquardt (LM)

algorithm to provide better performance for the model. By using the LM algorithm, the error is
M

back propagated through the layers of model by minimizing of a cost function until the optimum
ED

weight vectors are evaluated and the output parameter is calculated by the model [27].
PT

2.2 Least Square Support Vector Machine (LSSVM)

CE

The LSSVM modeling technique is a machine learning algorithm which has found
AC

extensive applicability in various fields. This algorithm was developed by modifying the original

concept of support vector machine (SVM). In SVM algorithm, the modeling of experimental

data leads to a quadratic programming problem which brings complexities in the optimization

and calculation process of the model [28, 29]. The LSSVM approach overcomes to these

complexities by solving the nonlinear fitting problem by using linear programming approach

which simplifies the calculation process of the model. Modeling a problem by using the LSSVM

5
 ACCEPTED MANUSCRIPT

approach requires the optimization of two tuning parameters namely and 2 . These

parameters are required during solving the linear programming problem by using LSSVM

method. The optimum values of the aforementioned tuning parameters are normally determined

by using an optimization algorithm such as coupled simulated annealing (CSA) or particle

swarm optimization (PSO) [29, 30]. These parameters are optimized by minimizing a cost

T
function between predicted data and target data. In the present work, these parameters were

IP
optimized using the CSA algorithm by minimizing the mean square error (MSE) between

CR
experimental data and predicted values. Further information about the performance of LSSVM

US
algorithm is available in literature.
AN
2.3 Adaptive Neuro Fuzzy Inference System
M

The ANFIS modeling technique is another soft computing approach which is useful for

nonlinear data modeling and fitting problems. This algorithm uses the combined abilities of
ED

Fuzzy Logic (FL) and Artificial Neural Networks (ANNs) [31, 32]. This algorithm has a five
PT

layer structure which is briefly explained herein [31]:

Layer 1: There are certain functions in this layer which are called membership functions
CE

(MFs). The number of these MFs is identical to the number of input parameters. Each MF
AC

accepts the individual inputs and applies a transformation on it. There are various types of MFs

which can be utilized in the nodes of first layer. However, the Gaussian type MFs is more

popular because they provide a smoother behavior for model. These MFs are formulated as

follows:

1 (X i Z j )
2

ij (X i ) exp( ) (1)
2 j2

6
 ACCEPTED MANUSCRIPT

In which ij is the membership degree of j-th MF for i-th input parameter, Xi is the i-th

input parameter, Zj is the center of j-th Gaussian MF and j is the variance parameter of j-th
2

MF.

T
Layer 2: This layer determines a parameter known as firing strength (Wi) by multiplying the

IP
membership degrees of MFs for each input parameter as follows:

CR
m
W j ij (X j ) (2)

US
i 1
Layer 3: In this layer, the calculated firing strengths in the previous layer are normalized
AN
between 0 and 1which is shown by W i as below:

Wi
Wi
M

W i
i
(3)
ED

Layer 4: In this layer, the normalized firing strengths in the previous layer are multiplied by a

first order (linear) Takagi-Sugeno-Kang (TSK) type Fuzzy Inference System (FIS) as follows:
PT

m
W i f i W i ( nij X j rij ) (4)
CE

j 1

In which fi represents the first order TSK type FIS. The nij and rij parameters are adjustable
AC

variables of TSK FIS which should be optimized by a reliable algorithm to provide a better

performance for model.

Layer 5: The sum of all calculated values in the previous layer is the output parameter of the

model (Y ) which is calculated in this layer as below:

Y W i f i (5)
i

7
 ACCEPTED MANUSCRIPT

2.4 Gene Expression Programming (GEP)

Gene expression programming (GEP) is a powerful modeling technique which operates by

using optimization abilities of Genetic Algorithm (GA). The main applications of GEP are for

T
regression and data fitting purposes. GEP is similar to GA in its operation and performance. This

IP
technique was developed by Koza [33] for the first time and was gradually modified after him.

CR
Among various computer-based techniques which are used to solve complex and nonlinear

problems, the GEP is one of the popular ones. In addition, an important drawback of other

US
computer-based approaches such as Artificial Neural Networks (ANNs) is that, the final solution
AN
of these models for a problem is in a black-box form, which brings limitations on their

interpretation and applicability. However, GEP fits the input data and results in a final
M

mathematical model for the problem, which provides higher generalization ability compared to
ED

other models [34, 35].

The symbolic regression term is normally attributed to GEP because it finalizes its
PT

modeling procedure by introducing a mathematical correlation for the problem under

CE

consideration. In classical data regression and fitting techniques, first it is needed to construct a

mathematical formulation and then tune and adjust the parameters and coefficients of the
AC

constructed correlation. However, in symbolic regression which is performed by GEP technique

the mathematical form and the coefficients are constructed and adjusted simultaneously.

GEP conducts the fitting procedure in two stages. First, it combines the functions (the

functions that are characterized and defined by user to be implemented in the fitting process)

based on a random decision process and then it adjusts and optimizes the coefficients of each

8
 ACCEPTED MANUSCRIPT

constructed model by minimizing a cost function which is normally the Mean Square Error

(MSE) between experimental values and estimated data.

In the next step, GEP uses its operators such as cross over and mutation to construct a new

population of better and fittest solutions. Mutation operator replaces a bad performance sub-tree

with a new sub-tree with better performance. Cross over is also the process of combining two

T
better and fittest expressions to generate a new offspring.

IP
Generally, GEP uses two approaches to process a fitting problem called Standard Genetic

CR
Programming (SGP) and Multi-gene Genetic Programming (MGP). In SGP technique, just one

gene is created and used, while in MGP technique, several genes are created to handle the fitting

US
problem. Every gene in MGP method is generated by weighted linear combinations of results of
AN
several SGP trees and each tree is considered as a gene in the total fitting process [36].

The mathematical formulation of MGP fitting procedure can be represented as below [37]:
M

N
Y P b0 w i G i (6)
ED

i 1

where YP is the estimated value by the GEP model, Gi is the ith gene in model, N represents
PT

the number of genes utilized in GEP fitting process, wi denotes the weight term of the ith gene
CE

and b0 is the bias term. The optimum values of weight terms for each gene are evaluated by least

square fitting of estimations of the proposed GEP model versus experimental data. For instance,
AC

the MGP structure that is shown in Figure 1 is formulated as below:

Y d 0 d1 (a b Sin(c )) d 2 (d e f g ) (7)

This figure represents a symbolic regression which composes of two genes and a

maximum depth equal to 3 for each tree.

9
 ACCEPTED MANUSCRIPT

3 Data acquisition

Experimental measurement of viscosity of nanofluids was the subject of many researches

in recent years [2, 3, 10, 38-41]. According to these works, it can be concluded that the viscosity

of nanofluid is dependent on the temperature, size of nanoparticle, volume fraction of

T
nanoparticle, bulk density of nanoparticle and viscosity of the base fluid. Hence, these

IP
parameters were utilized as input parameters for development of the models. Moreover, as the

CR
viscosity of base fluid was not reported in some data points, this parameter was calculated by

using an appropriate correlation for each system. These correlations are summarized in Table 1.

US
In the present study, an extensive number of 1140 experimental data of viscosity of
AN
different nanofluids with various mixtures of nanoparticles and base fluids were collected from

several published works in the literature [2, 3, 13, 38-49]. The statistical details of the gathered
M

data points which were used in the present study are listed in Table 2. Moreover, the name and
ED

type of base fluid and nanoparticle as well as the number of data points gathered from each

reference are listed in Table 3.

PT
CE

4 Model development
AC

4.1 MLP-NN model

An important parameter which greatly affects the performance of the MLP-NN model is

the number of neurons in the hidden layer of this network. In the present study, a single hidden

layer network with tansig type transfer function in the hidden layer and purelin transfer function

in the output layers is utilized. The number of neurons in the hidden layer was changed from 4 to

10
 ACCEPTED MANUSCRIPT

25 and the MSE value of the model was monitored for different models. Figure 2 shows the plot

of the MSE value of different models against the number of neurons of the hidden layer. As is

clear from this figure the network with 18 neurons in the hidden layer exhibits the lowest MSE

value and is the most appropriate model.

T
4.2 CSA-LSSVM model

IP
CR
As mentioned earlier, modeling a non-linear data fitting problem by LSSVM model

requires optimization of two tuning parameters namely and 2 which are related to structure

US
of these models. In the present study, the CSA algorithm was utilized to obtain the optimum
AN
values of these parameters. Using this algorithm, the optimum values of and 2 were

obtained to be 104285427.15 and 0.53, respectively.

M
ED

4.3 Hybrid-ANFIS model

PT

In order to evaluate the performance of ANFIS model for prediction of experimental

CE

viscosity data first a TSK type FIS was constructed by using the training data. The TSK type FIS
AC

was constructed using the genfis 2 function of MATLAB. This function employs the subtractive

clustering method for development of the initial FIS. After construction of the initial FIS it

should be trained by using an appropriate algorithm. In the present work, the hybrid method was

utilized to train the initial constructed FIS. The parameters of the Hybrid-ANFIS model are listed

in Table 4.

11
 ACCEPTED MANUSCRIPT

4.4 GEP model

To develop an effective and dependable model for prediction of viscosity of nanofluids,

great effort was put forward to apply different GEP based correlations using trial and error

process. Throughout these trial and error approaches for development of GEP based model, it

T
IP
was found that an increase in the number of genes for a model leads to an increase in the depth of

CR
expression trees (ETs) and as a result leads to complexities in solution domain of problem and

final correlation and increases the run time of the model. As a result, in the present study, three

US
genes were used for the final form of the GEP based model. Moreover, the Mean Square Error

(MSE) was used as a cost function to optimize the coefficients of final model. The operators and
AN
functions of , +, -, / and power ( ab ) were used for development of GEP based model. The input
M

parameters were temperature (T(K)), volume fraction of nanoparticle ( (%)), size of

ED

nanoparticle ( S (nm)), viscosity of base fluid ( bf (cp (centipoise))) and density of nanoparticle

( (g/cm3)). The output parameter was the viscosity of nanofluid ( nf (cp)). The optimum
PT

mathematical model which was achieved using GEP model for prediction of viscosity of
CE

nanofluid is as follows:
AC

nf ( A B C )10 (8)

2.708472(10 T 10 )
A 1.461570 10 bf (9)
16.484304
0.112396(10 10 bf ) (10)
B 10
S 10
S 4.872
C 31.49091107 (11.208771
10
)

12
 ACCEPTED MANUSCRIPT

(11)

5 Results and discussion

T
5.1 Evaluation of the developed models

IP
CR
Different statistical quality measure techniques were utilized to evaluate the accuracy of

predictions of the developed models. First, the estimated viscosity data were plotted versus the

US
target viscosity data as shown in Figure 3. According to this figure, the experimental data and
AN
estimated data are acceptably distributed around the unit slope line (Y=X line) for all models.

This indicates that there is an acceptable agreement between predictions of the developed models
M

and target viscosity data. The correlation coefficient of estimations of the developed models are
ED

0.9942, 0.9991, 0.9892 and 0.9505 for MLP-NN, CSA-LSSVM, Hybrid-ANFIS and GEP

models, respectively, which is another indication for the accuracy of the developed models.
PT

Figure 4 shows the relative deviation plot of outcomes of the proposed models. This figure also
CE

reveals that the relative deviations of the models are mostly concentrated in the region between

relative deviations of 20%. This indicates that the developed models are able to estimate most
AC

of target viscosity data with low relative errors. This figure also shows that the CSA-LSSVM

model exhibits better predictions in comparison with other models as the relative error in

predictions of this model are lower than other models.

The statistical evaluation of estimations of the developed models were examined by

utilizing three statistical parameters including correlation factor (R2), Average Absolute Relative

13
 ACCEPTED MANUSCRIPT

Deviation (AARD%) and Root Mean Square Error (RMSE). The formulation of these

parameters is represented as follows:

( Pr ed (i ) Exp (i )) 2
R2 1 i 1
N (12)
(
i 1
Pr ed (i ) Exp ) 2

T
100 N ( Pr ed (i) Exp (i ))
AARD % (13)

IP
N i 1 Exp (i)

CR
N

( Pr ed (i ) Exp (i )) 2
(14)
RMSE ( i 1
) 0.5
N

US
In which Pred denotes the estimated viscosity of nanofluid, Exp is the experimental value
AN
of viscosity, N denotes the number of data points and Exp is the average value of experimental

viscosity. The statistical parameters of the developed models are listed in Table 5. This table also
M

confirms the accuracy of the developed models thanks to low values of AARD% and RMSE
ED

values and high R2 value.

At this part, the predictions of the developed models will be compared with outcomes of
PT

the literature correlations for prediction of viscosity of nanofluids. Table 6 presents the
CE

formulation of these correlations. The cross plot of these correlations are shown in Figure 5.

Comparison between this figure and Figure 3 reveals that the estimations of these correlations
AC

diverge from unit slope line more than the developed models as well as lower R2 values which

confirm that the estimations of the developed models are more accurate and dependable than

these correlations. Moreover, Figure 6 shows the relative deviation plot of literature correlations.

According to this figure, the relative errors of these correlations are higher than the proposed

models. In addition, the cumulative absolute relative error plot of the proposed models and the

14
 ACCEPTED MANUSCRIPT

literature correlations is shown in Figure 7. According to this figure, the predictions of the

models are accurate and dependable than other methods because most of data points predicted by

these models (about 80%) exhibit an error less than 20%. However, the literature correlations of

Batchelor [50], Wang et al. [51], Chen et al. [4], Abedian et al. [52], Ward [53], Brinkman [54],

Einstein [18] and Krieger and Dougherty [49,50] are capable of predicting about 36%, 57%, 58%,

T
24%, 24%, 24% and 24% of data points, respectively, with an error less than 20%. This Figure

IP
also shows that the predictions of the CSA-LSSVM model are better than other developed

CR
models and literature correlations as this model is able to predict all of data points with relative

errors less than 15%.

US
The statistical parameters of these literature correlations are listed in Table 7. According to
AN
this table, these correlations exhibit higher AARD% and RMSE values and lower R2 values

compared to the proposed models which reveal that the developed models effectively
M

outperforms literature correlations for prediction of experimental data.

ED

In addition to these correlations, Meybodi et al. [55] recently developed a correlation for

prediction of viscosity of water based nanofluids. This correlation is also provided in Table 6. A
PT

comparison was also conducted between the predictions of the developed models and this
CE

correlation. However, as the correlation developed by Meybodi et al. [55] is just valid for water

based nanofluids, it was applied to data points in which the base fluid was water (907 data out of
AC

1140 data points). Figure 8 shows a graphical comparison between R2, AARD% and RMSE

values of the developed models in the present work and the correlation developed by Meybodi et

al. [55]. This figure also shows the accuracy of the proposed models because they exhibit higher

R2 values and lower AARD% and RMSE values compared to Meybodi et al. [55] correlation.

Another advantage of the developed models in the present work over Meybodi et al. [55]

15
 ACCEPTED MANUSCRIPT

correlation is the higher generalization ability of the proposed models in this study. The newly

introduced models are not limited to water based nanofluids and in addition to water and

deionized water based fluids, are valid for other base fluids including mixture of ethylene glycol

and water, pure ethanol, pure ethylene glycol, mixture of propylene glycol and water, pure R11

refrigerant and toluene. Moreover, it should be mentioned that the developed correlation in the

T
present work is just applicable for prediction of viscosity of nanofluid with Newtonian behavior

IP
(i.e. the viscosity of nanofluid is independent of shear rate).

CR
5.2 Sensitivity analysis of the input and output data
US
AN
To understand the relation between individual input parameters and the value of viscosity
M

of nanofluid, a sensitivity analysis was performed on the effect of input data on the target
ED

viscosity value. For this purpose, the relevancy factor (r) was utilized. An input parameter with a

higher r value has greater influence on the value of output parameter (viscosity in the present
PT

work). However, it is better to use directional form of relevancy factor which contains a value
CE

with a certain direction and provides a better judgment about the influence of an input parameter

on target value. The mathematical expression for directional r value is as follows:

AC

(Inp k ,i Inp k )( nf ,Pr ed (i ) nf ,Pr ed )

r (Inp k , nf ,Pr ed ) i 1 (15)
n n

(Inp
i 1
k ,i Inp k ) 2 ( nf ,Pred (i ) nf ,Pred ) 2
i 1

In which nf ,Pred (i ) and nf ,Pred are the value of i-th predicted nanofluid viscosity and

average value of predicted nanofluid viscosity, respectively. Inp k ,i represents the i-th value of k-

16
 ACCEPTED MANUSCRIPT

th input parameter and Inp k is the average value of k-th input parameter. Figure 9 depicts the

influence of individual inputs on viscosity of nanofluid for different nanoparticles. This figure

shows that the temperature has positive effect on the viscosity value of Al2O3, CuO, carbon nano

tube (CNT) and graphene based nanofluids and a negative effect on the viscosity of SiO2 and

TiO2 based nanofluids. This figure shows that the particle volume fraction has the largest effect

T
IP
on the viscosity of Al2O3 based nanofluids but for CuO, SiO2, TiO2 and graphene based

nanofluids the viscosity of base fluid has the largest impact on the viscosity of nanofluid. For

CR
CNT based nano fluids the size of nano particle has largest impact on the viscosity value.

US
AN
4. Conclusion
M

In order to estimate the viscosity of Newtonian nanofluids, four models namely MLP-NN,

CSA-LSSVM, Hybrid-ANFIS and GEP were developed in the present study. The main objective
ED

of this work was to provide accurate and general models to improve the generalization ability
PT

and accuracy of previous literature correlations. The error analysis of predictions of the

developed models shows their acceptable accuracy and reliability of them. The comparison
CE

between statistical quality measure parameters of the developed models and the literature
AC

correlations also shows the superiority and higher accuracy of the models developed in the

present work. Moreover, the predictions of CSA-LSSVM model are more accurate and reliable

than the predictions of other developed models. As the developed models exhibit acceptable

generalization ability and accuracy, they could be effectively utilized for prediction of viscosity

of Newtonian nanofluids.

17
 ACCEPTED MANUSCRIPT

References

[1] M. Kamiski, R.L. Ossowski, Prediction of the effective parameters of the nanofluids using
the generalized stochastic perturbation method, Physica A: Statistical Mechanics and its
Applications, 393 (2014) 10-22.
[2] S. Murshed, K. Leong, C. Yang, Investigations of thermal conductivity and viscosity of
nanofluids, International Journal of Thermal Sciences, 47 (2008) 560-568.
[3] M. Mehrabi, M. Sharifpur, J.P. Meyer, Viscosity of nanofluids based on an artificial

T
intelligence model, International Communications in Heat and Mass Transfer, 43 (2013) 16-21.
[4] H. Chen, Y. Ding, C. Tan, Rheological behaviour of nanofluids, New journal of physics, 9

IP
(2007) 367.
[5] L. Chen, H. Xie, Y. Li, W. Yu, Nanofluids containing carbon nanotubes treated by

CR
mechanochemical reaction, Thermochimica Acta, 477 (2008) 21-24.
[6] J.-Y. Jung, C. Cho, W.H. Lee, Y.T. Kang, Thermal conductivity measurement and
characterization of binary nanofluids, International Journal of Heat and Mass Transfer, 54 (2011)

US
1728-1733.
[7] S. Lee, S.-S. Choi, S. Li, and, J. Eastman, Measuring thermal conductivity of fluids
containing oxide nanoparticles, Journal of Heat transfer, 121 (1999) 280-289.
AN
[8] S. Shaikh, K. Lafdi, R. Ponnappan, Thermal conductivity improvement in carbon
nanoparticle doped PAO oil: An experimental study, Journal of Applied Physics, 101 (2007)
064302.
[9] H.U. Kang, S.H. Kim, J.M. Oh, Estimation of thermal conductivity of nanofluid using
M

experimental effective particle volume, Experimental Heat Transfer, 19 (2006) 181-191.

[10] W. Duangthongsuk, S. Wongwises, Measurement of temperature-dependent thermal
ED

conductivity and viscosity of TiO 2-water nanofluids, Experimental thermal and fluid science, 33
(2009) 706-714.
[11] H. Masuda, A. Ebata, K. Teramae, Alteration of thermal conductivity and viscosity of liquid
PT

by dispersing ultra-fine particles. Dispersion of Al2O3, SiO2 and TiO2 ultra-fine particles, DOI
(1993).
[12] B.C. Pak, Y.I. Cho, Hydrodynamic and heat transfer study of dispersed fluids with
CE

submicron metallic oxide particles, Experimental Heat Transfer an International Journal, 11

(1998) 151-170.
[13] C. Nguyen, F. Desgranges, G. Roy, N. Galanis, T. Mar, S. Boucher, H.A. Mintsa,
Temperature and particle-size dependent viscosity data for water-based nanofluidshysteresis
AC

phenomenon, International Journal of Heat and Fluid Flow, 28 (2007) 1492-1506.

[14] T.X. Phuoc, M. Massoudi, R.-H. Chen, Viscosity and thermal conductivity of nanofluids
containing multi-walled carbon nanotubes stabilized by chitosan, International Journal of
Thermal Sciences, 50 (2011) 12-18.
[15] V.Y. Rudyak, S. Dimov, V. Kuznetsov, On the dependence of the viscosity coefficient of
nanofluids on particle size and temperature, Technical Physics Letters, 39 (2013) 779-782.
[16] I. Mahbubul, R. Saidur, M. Amalina, Latest developments on the viscosity of nanofluids,
International Journal of Heat and Mass Transfer, 55 (2012) 874-885.
[17] E. Heidari, S. Ghoreishi, Prediction of supercritical extraction recovery of EGCG using
hybrid of Adaptive Neuro-Fuzzy Inference System and mathematical model, The Journal of
Supercritical Fluids, 82 (2013) 158-167.

18
 ACCEPTED MANUSCRIPT

[18] A. Einstein, Eine neue bestimmung der molekldimensionen, Annalen der Physik, 324
(1906) 289-306.
[19] J.A. Eastman, S. Phillpot, S. Choi, P. Keblinski, Thermal transport in nanofluids 1, Annu.
Rev. Mater. Res., 34 (2004) 219-246.
[20] I.M. Krieger, T.J. Dougherty, A mechanism for non Newtonian flow in suspensions of
rigid spheres, Transactions of The Society of Rheology (1957-1977), 3 (1959) 137-152.
[21] L.E. Nielsen, Generalized equation for the elastic moduli of composite materials, Journal of
Applied Physics, 41 (1970) 4626-4627.
[22] A. Najafi-Marghmaleki, M.R. Khosravi-Nikou, A. Barati-Harooni, A new model for
prediction of binary mixture of ionic liquids+ water density using artificial neural network,

T
Journal of Molecular Liquids, 220 (2016) 232-237.

IP
[23] A. Hemmati-Sarapardeh, B. Aminshahidy, A. Pajouhandeh, S.H. Yousefi, S.A. Hosseini-
Kaldozakh, A soft computing approach for the determination of crude oil viscosity: Light and

CR
intermediate crude oil systems, Journal of the Taiwan Institute of Chemical Engineers, 59 (2016)
1-10.
[24] S. Sabzevari, M. Moosavi, Density prediction of liquid alkali metals and their mixtures
using an artificial neural network method over the whole liquid range, Fluid Phase Equilibria,

US
361 (2014) 135-142.
[25] F. Yousefi, H. Karimi, M.M. Papari, Modeling viscosity of nanofluids using diffusional
neural networks, Journal of Molecular Liquids, 175 (2012) 85-90.
AN
[26] E. Heidari, M.A. Sobati, S. Movahedirad, Accurate prediction of nanofluid viscosity using a
multilayer perceptron artificial neural network (MLP-ANN), Chemometrics and Intelligent
Laboratory Systems, 155 (2016) 73-85.
M

[27] M.A. Ahmadi, Developing a robust surrogate model of chemical flooding based on the
artificial neural network for enhanced oil recovery implications, Mathematical Problems in
ED

Engineering, 2015 (2015).

[28] M.A. Ahmadi, M. Ebadi, S.M. Hosseini, Prediction breakthrough time of water coning in
the fractured reservoirs by implementing low parameter support vector machine approach, Fuel,
117 (2014) 579-589.
PT

[29] M.-A. Ahmadi, A. Bahadori, A LSSVM approach for determining well placement and
conning phenomena in horizontal wells, Fuel, 153 (2015) 276-283.
[30] M.-A. Ahmadi, B. Pouladi, Y. Javvi, S. Alfkhani, R. Soleimani, Connectionist technique
CE

estimates H 2 S solubility in ionic liquids through a low parameter approach, The Journal of
Supercritical Fluids, 97 (2015) 81-87.
[31] J.-S. Jang, ANFIS: adaptive-network-based fuzzy inference system, IEEE transactions on
AC

systems, man, and cybernetics, 23 (1993) 665-685.

[32] L.A. Zadeh, Fuzzy sets, Information and control, 8 (1965) 338-353.
[33] J.R. Koza, Genetic programming: on the programming of computers by means of natural
selection, MIT press1992.
[34] C. Preechakul, S. Kheawhom, Modified genetic algorithm with sampling techniques for
chemical engineering optimization, Journal of Industrial and Engineering Chemistry, 15 (2009)
110-118.
[35] M. Vatani, M. Asghari, G. Vakili-Nezhaad, Application of Genetic Algorithm to the
calculation of parameters for NRTL and Two-Suffix Margules models in ternary extraction ionic
liquid systems, Journal of Industrial and Engineering Chemistry, 18 (2012) 1715-1720.

19
 ACCEPTED MANUSCRIPT

[36] M. Izadmehr, R. Shams, M.H. Ghazanfari, New correlations for predicting pure and impure
natural gas viscosity, Journal of Natural Gas Science and Engineering, 30 (2016) 364-378.
[37] C. Hii, D.P. Searson, M.J. Willis, Evolving toxicity models using multigene symbolic
regression and multiple objectives, International Journal of Machine Learning and Computing, 1
(2011) 30.
[38] J.-H. Lee, K.S. Hwang, S.P. Jang, B.H. Lee, J.H. Kim, S.U. Choi, C.J. Choi, Effective
viscosities and thermal conductivities of aqueous nanofluids containing low volume
concentrations of Al 2 O 3 nanoparticles, International Journal of Heat and Mass Transfer, 51
(2008) 2651-2656.
[39] Z. Jia-Fei, L. Zhong-Yang, N. Ming-Jiang, C. Ke-Fa, Dependence of nanofluid viscosity on

T
particle size and pH value, Chinese Physics Letters, 26 (2009) 066202.

IP
[40] S.W. Lee, S.D. Park, S. Kang, I.C. Bang, J.H. Kim, Investigation of viscosity and thermal
conductivity of SiC nanofluids for heat transfer applications, International Journal of Heat and

CR
Mass Transfer, 54 (2011) 433-438.
[41] M. Singh, L. Kundan, Experimental study on thermal conductivity and viscosity of Al2O3
nanotransformer oil, Int. J. Theo. App. Res. Mech. Engg, 2 (2013) 125-130.
[42] I. Tavman, A. Turgut, M. Chirtoc, H. Schuchmann, S. Tavman, Experimental investigation

US
of viscosity and thermal conductivity of suspensions containing nanosized ceramic particles,
Archives of Materials Science, 100 (2008).
[43] M. Chandrasekar, S. Suresh, A.C. Bose, Experimental investigations and theoretical
AN
determination of thermal conductivity and viscosity of Al 2 O 3/water nanofluid, Experimental
Thermal and Fluid Science, 34 (2010) 210-216.
[44] P.K. Namburu, D.P. Kulkarni, D. Misra, D.K. Das, Viscosity of copper oxide nanoparticles
M

dispersed in ethylene glycol and water mixture, Experimental Thermal and Fluid Science, 32
(2007) 397-402.
ED

[45] D.P. Kulkarni, D.K. Das, R.S. Vajjha, Application of nanofluids in heating buildings and
reducing pollution, Applied Energy, 86 (2009) 2566-2573.
[46] M. Naik, L.S. Sundar, Investigation into thermophysical properties of glycol based CuO
nanofluid for heat transfer applications, World Academy of Science, Engineering and
PT

Technology, 59 (2011) 440-446.

[47] M. Pastoriza-Gallego, C. Casanova, J.a. Legido, M. Pieiro, CuO in water nanofluid:
influence of particle size and polydispersity on volumetric behaviour and viscosity, Fluid Phase
CE

Equilibria, 300 (2011) 188-196.

[48] R. Singh, O. Sanchez, S. Ghosh, N. Kadimcherla, S. Sen, G. Balasubramanian, Viscosity of
magnetitetoluene nanofluids: Dependence on temperature and nanoparticle concentration,
AC

Physics Letters A, 379 (2015) 2641-2644.

[49] N. Jamshidi, M. Farhadi, D. Ganji, K. Sedighi, Experimental investigation on viscosity of
nanofluids, International Journal of Engineering, 25 (2012) 201-209.
[50] G. Batchelor, The effect of Brownian motion on the bulk stress in a suspension of spherical
particles, Journal of Fluid Mechanics, 83 (1977) 97-117.
[51] X. Wang, X. Xu, S.U. S. Choi, Thermal conductivity of nanoparticle-fluid mixture, Journal
of thermophysics and heat transfer, 13 (1999) 474-480.
[52] B. Abedian, M. Kachanov, On the effective viscosity of suspensions, International Journal
of Engineering Science, 48 (2010) 962-965.
[53] S. Ward, Properties of well-defined suspensions of solids in liquids, Journal of Oil & colour
Chemists Association, 38 (1955).

20
 ACCEPTED MANUSCRIPT

[54] H. Brinkman, The viscosity of concentrated suspensions and solutions, The Journal of
Chemical Physics, 20 (1952) 571-571.
[55] M.K. Meybodi, A. Daryasafar, M.M. Koochi, J. Moghadasi, R.B. Meybodi, A.K.
Ghahfarokhi, A novel correlation approach for viscosity prediction of water based nanofluids of
Al 2 O 3, TiO 2, SiO 2 and CuO, Journal of the Taiwan Institute of Chemical Engineers, 58
(2016) 19-27.
[56] I.M. Krieger, T.J. Dougherty, A mechanism for non Newtonian flow in suspensions of
rigid spheres, Transactions of the Society of Rheology, 3 (1959) 137-152.
[57] S. Halelfadl, P. Estell, B. Aladag, N. Doner, T. Mar, Viscosity of carbon nanotubes water-
based nanofluids: Influence of concentration and temperature, International Journal of Thermal

T
Sciences, 71 (2013) 111-117.

IP
[58] J. Chevalier, O. Tillement, F. Ayela, Rheological properties of nanofluids flowing through
microchannels, Applied physics letters, 91 (2007) 3103.

CR
[59] I.S. Khattab, F. Bandarkar, M. Khoubnasabjafari, A. Jouyban, Density, viscosity, surface
tension, and molar volume of propylene glycol+ water mixtures from 293 to 323K and
correlations by the JouybanAcree model, Arabian Journal of Chemistry, DOI (2012).
[60] P. Garg, J.L. Alvarado, C. Marsh, T.A. Carlson, D.A. Kessler, K. Annamalai, An

US
experimental study on the effect of ultrasonication on viscosity and heat transfer performance of
multi-wall carbon nanotube-based aqueous nanofluids, International Journal of Heat and Mass
Transfer, 52 (2009) 5090-5101.
AN
[61] M. Kole, T. Dey, Investigation of thermal conductivity, viscosity, and electrical
conductivity of graphene based nanofluids, Journal of Applied Physics, 113 (2013) 084307.
[62] N. Singh, G. Chand, S. Kanagaraj, Investigation of thermal conductivity and viscosity of
M

carbon nanotubesethylene glycol nanofluids, Heat Transfer Engineering, 33 (2012) 821-827.

ED
PT
CE
AC

21
 ACCEPTED MANUSCRIPT

T
IP
Figure 1. Schematic structure of a MGP structure with two genes.

CR
US
AN
M
ED
PT
CE
AC

22
 ACCEPTED MANUSCRIPT

T
IP
CR
US
AN
M

Figure 2. The change in MSE value of different MLP-NN models against number of hidden
ED

neurons.
PT
CE
AC

23
 ACCEPTED MANUSCRIPT

T
IP
CR
US
AN
M
ED

(a)
PT
CE
AC

24
 ACCEPTED MANUSCRIPT

T
IP
CR
US
AN
M
ED

(b)
PT
CE
AC

25
 ACCEPTED MANUSCRIPT

T
IP
CR
US
AN
M
ED

(c)
PT
CE
AC

26
 ACCEPTED MANUSCRIPT

T
IP
CR
US
AN
M
ED

(d)
PT

Figure 3. Cross plot of predicted data against target data: (a) MLP-NN, (b) CSA-LSSVM, (c)
CE

Hybrid-ANFIS, (d) GEP

AC

27
 ACCEPTED MANUSCRIPT

T
IP
CR
US
(a) (b)
AN
M
ED
PT
CE
AC

(c) (d)

Figure 4. Relative error plot of the developed models: (a) MLP-NN, (b) CSA-LSSVM, (c)

Hybrid-ANFIS, (d) GEP

28
 ACCEPTED MANUSCRIPT

T
IP
CR
US
(a) AN (b)
M
ED
PT
CE
AC

(c) (d)

29
 ACCEPTED MANUSCRIPT

T
IP
CR
US
(e) AN (f)
M
ED
PT
CE
AC

(g) (h)

Figure 5. Cross plot of predictions of the literature correlations: (a) Batchelor [50], (b) Wang et

al. [51], (c) Chen et al. [4], (d) Abedian et al. [52], (e) Ward [53], (f) Brinkman [54], (g) Einstein

[18], and (h) Krieger and Dougherty [56, 57].

30
 ACCEPTED MANUSCRIPT

T
IP
CR
US
(a) AN (b)
M
ED
PT
CE
AC

(c) (d)

31
 ACCEPTED MANUSCRIPT

T
IP
CR
US
(e) AN (f)
M
ED
PT
CE
AC

(g) (h)

Figure 6. Relative deviation plot of the literature correlations: (a) Batchelor [50], (b) Wang et al.

[51], (c) Chen et al. [4], (d) Abedian et al. [52], (e) Ward [53], (f) Brinkman [54], (g) Einstein

[18], and (h) Krieger and Dougherty [56, 57].

32
 ACCEPTED MANUSCRIPT

100
Einstein
90 Batchelor
Brinkman
80
Ward
70 Wang et al.
Chen et al.
Frequency (%)

60
Abedian et al.
50 Krieger and Dougherty

T
GEP

IP
40
MLP-NN
30 CSA-LSSVM

CR
Hybrid-ANFIS
20

10

US
0
0 5 10 15 20
Absolute Relative Error (%)
AN
Figure 7. Cumulative frequency of data points versus absolute relative error.
M
ED
PT
CE
AC

33
 ACCEPTED MANUSCRIPT

1.2 14

1 12

10
0.8

AARD%
8
R2

0.6
6
0.4

T
4

IP
0.2 2

CR
0 0

CSA-LSSVM MLP-NN Hybrid-ANFIS GEP Meybodi et al. CSA-LSSVM MLP-NN Hybrid-ANFIS GEP Meybodi et al.

US
(a) (b)

0.45
AN
0.4

0.35
M

0.3
ED
RMSE

0.25

0.2

0.15
PT

0.1

0.05
CE

CSA-LSSVM MLP-NN Hybrid-ANFIS GEP Meybodi et al.

AC

(c)

Figure 8. Comparison between the developed models and Meybodi et al. [55] correlation: (a) R2,

(b) AARD%, (c) RMSE.

34
 ACCEPTED MANUSCRIPT

T
IP
CR
US
(a) (b)
AN
M
ED
PT
CE

(c) (d)
AC

35
 ACCEPTED MANUSCRIPT

T
IP
CR
US
(e) (f)
Figure 9. Sensitivity analysis of effect of different parameters on the viscosity of nanofluids
AN
containing different nanoparticles: (a) Al2O3, (b) CuO, (c) SiO2, (d) TiO2, (e) CNT, (f) Graphene
M
ED
PT
CE
AC

36
 ACCEPTED MANUSCRIPT

Table 1. Literature correlations used to determine the viscosity of base fluid in each system.

Correlation Type of base fluid Reference

1.12646 0.039638T
exp( )
bf (cp ) 1 0.00729769T Deionized Water (DIW) [40]
1000

247.8
2
bf (cp ) 2.414 10 T 140 Water [49]

T
IP
Mixture of ethylene glycol and
bf (cp ) exp(0.01797T 1.0305) [49]

CR
water (25:75)

Mixture of ethylene glycol and

bf (cp ) exp(0.02045T 1.7203) [49]

US
water (50:50)

bf (cp ) exp(0.04452T 4.0316) Ethylene glycol [49]

AN
bf (cp ) 1.3 Ethanol [58]
M

x 1x 2 x x (x x 2 ) Mixture of propylene glycol and

ln bf (cp ) x 1 ln 1 x 2 ln 2 926.206( ) 606.41( 1 2 1 )
T T [59]
ED

x 1 is the weight fraction of water water

Mixture of ethylene glycol and

273.15 273.15 2
PT

ln bf (cp ) 1000 (5.919142 9.457844( ) 10.98990( ) ) [44]

T T
water (60:40) (weight)
CE
AC

37
 ACCEPTED MANUSCRIPT

Table 2. Statistical details about gathered data bank in the present work.

Parameter Min. Max. Average Standard Deviation

Temperature (K) 257.94 344.35 309.66 15.31

Size of nanoparticle (nm) 7.00 220 33.73 17.17

Volume fraction of nanoparticle (%) 0.01 9 2.86 3.01

T
Density of nanoparticle (g/cm3) 0.2 6.31 4.40 1.44

IP
Viscosity of base fluid (cp) 0.39 6.03 1.08 0.96

CR
Viscosity of nanofluid (cp) 0.44 6.48 1.63 1.16

US
AN
M
ED
PT
CE
AC

38
 ACCEPTED MANUSCRIPT

Table 3. The detailed analysis of the present data set.

Nanoparticle Base fluid Number of data Reference

Al2O3 DI Water 4 [2]

Al2O3 R11 refrigerant 21 [41]

[3, 13, 38,

T
Al2O3 Water 506 42, 43]

IP
CuO Ethylene glycol and water mixture 60:40 weight 11 [44, 45]

CR
Propylene glycol and water mixture, 30:70 [46]

US
CuO volume 53

CuO Water 286 [3, 13, 47]

AN
Fe3O4 Toluene 2 [48]

SiC DI Water 2 [40]

M

SiO2 DI water 64 [39]

ED

SiO2 Ethylene glycol and water mixture, 25:75 volume 11 [49]

PT

SiO2 Ethylene glycol and water mixture, 50:50 volume 30 [49]

SiO2 Ethanol 18 [58]

CE

SiO2 Water 32 [3, 42]

AC

TiO2 DI water 3 [2]

TiO2 Ethylene glycol 35 [4]

TiO2 Water 8 [10]

Carbon nanotube
DI water 2 [60]
(CNT)

CNT DI water +Sodium dodecylbenzenesulfonate 20 [57]

39
 ACCEPTED MANUSCRIPT

(SDBS)

Graphene DI water + Ethylene Glycol 17 [61]

CNT Ethylene Gycol 15 [62]

T
IP
CR
US
AN
M
ED
PT
CE
AC

40
 ACCEPTED MANUSCRIPT

Table 4. Detailed analysis of the developed ANFIS model.

Parameter Description/Value

Structure of FIS Takagi-Sugeno

Type of initial FIS genfis 2

Type of MFs Gaussian

T
Number of inputs 5

IP
Number of outputs 1

CR
Number of fuzzy rules 15

US
Radius of influence/Radius of cluster 0.4

Training method Hybrid (Least square and back propagation techniques)

AN
M
ED
PT
CE
AC

41
 ACCEPTED MANUSCRIPT

Table 5. Statistical parameters of the developed models.

Parameter Data set MLP-ANN CSA-LSSVM Hybrid-ANFIS GEP

Test data 0.9939 0.9989 0.9884 0.9480

R2 Train data 0.9943 0.9994 0.9918 0.9608

0.9942 0.9990 0.9892 0.9505

T
All data
4.28 1.46 5.05 9.92

IP
Test data
AARD% Train data 3.82 1.36 4.55 9.58

CR
All data 3.92 1.44 4.95 9.85

Test data 0.090 0.037 0.123 0.27

RMSE Train data 0.088 0.030 0.109 0.23

US
All data 0.088 AN 0.035 0.121 0.26
M
ED
PT
CE
AC

42
 ACCEPTED MANUSCRIPT

Table 6. Literature correlations for prediction of viscosity of nanofluids.

Reference Correlation

Einstein [18] nf bf (1 2.5 )

bf
nf
(1 )2.5
Brinkman [54]

T
Ward [53] nf bf (1 2.5 (2.5 )2 (2.5 )3 (2.5 ) 4 )

IP
nf bf (1 2.5 6.2 2 )

CR
Batchelor [50]

nf bf (1 7.3 123 2 )
Wang et al.

US
[51]

Chen et al. [4] nf bf (1 10.6 (10.6 )2 )

AN
bf
nf
Abedian et al.

[52] (1 2.5 )
M

Krieger and

nf bf (1 ) 1.625
ED

Dougherty [56,
0.65
57]

PT

133.54064976-343.82413843 exp( ) 290.11804759 (exp( )) 2 78.993120761 (exp( ))3

Meybodi et al.
nf bf S S S
2
ln(S ) (ln(S ))
[55] 0.91161630781 32.330142333 11.732514460
CE

T T
AC

43
 ACCEPTED MANUSCRIPT

Table 7. Statistical parameters of the literature correlations.

Method R2 AARD% RMSE

Einstein [18] 0.5345 25.64 0.9409

Brinkman [54] 0.5408 25.37 0.9325

Ward [53] 0.5443 25.23 0.9279

T
Batchelor [50] 0.4589 30.38 1.9455

IP
Wang et al. [51] 0.6084 14.89 0.8387

CR
Chen et al. [4] 0.5992 14.93 0.8714

Abedian et al. [52] 0.5444 25.22 0.9278

US
Krieger and Dougherty
0.5421 25.32 0.9309
[56, 57]
AN
M
ED
PT
CE
AC

44
 ACCEPTED MANUSCRIPT

Research Highlights

Four viscosity models based on MLP-NN, LSSVM, ANFIS and GEP techniques have

T
been developed for Newtonian nanofluids.

IP
1140 literature data were utilized for development of the models.

CR
The accuracy of the models has been examined by comparing their predictions with
several literature correlations.

US
The models provides accurate predictions and outperforms literature correlations and
AN
presents better results.

The predictions of CSA-LSSVM model are better and accurate than the predictions of
M

other developed models.

ED
PT
CE
AC

45