Prediction of viscosity of several alumina-based nanofluids using various
artificial intelligence paradigms - Comparison with experimental data and
empirical correlations

Ehsan Gholami, Behzad Vaferi, Mohammad Amin Ariana

PII: S0032-5910(17)30834-3
DOI: doi:10.1016/j.powtec.2017.10.038
Reference: PTEC 12896

To appear in: Powder Technology

Received date: 16 June 2017

Revised date: 8 October 2017
Accepted date: 19 October 2017

Please cite this article as: Ehsan Gholami, Behzad Vaferi, Mohammad Amin Ariana,
Prediction of viscosity of several alumina-based nanofluids using various artificial intelli-
gence paradigms - Comparison with experimental data and empirical correlations, Powder
Technology (2017), doi:10.1016/j.powtec.2017.10.038

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

Prediction of viscosity of several alumina-based nanofluids using various

artificial intelligence paradigms - comparison with experimental data

PT
and empirical correlations

RI
Ehsan Gholami a, Behzad Vaferi b‫٭‬, Mohammad Amin Ariana c

SC
a
Department of Mechanical Engineering, Gachsaran Branch, Islamic Azad University, Gachsaran, Iran
b
Young Researchers and Elite Club, Shiraz Branch, Islamic Azad University, Shiraz, Iran

NU
c
Department of Petroleum Engineering, Gachsaran Branch, Islamic Azad University, Gachsaran, Iran
MA
ABSTRACT

Nanofluids have recently been considered as one of the most popular working fluid in heat transfer
D
TE

and fluid mechanics. Viscosity is one of the most important thermo-physical properties that

influence both momentum and heat transported by the nanofluids. Accurate estimation of this
P
CE

parameter is required for investigation the heat transfer performance of nanofluids. Therefore, in

this study 1- the most influential variables on viscosity of the nanofluids are determined 2- various
AC

artificial intelligence (AI) models are developed for prediction of viscosity of alumina nanoparticle

in various base fluids, 3- by comparing predictive accuracy of the developed models and available

empirical correlations, the best one is selected. Correlation matrix analyses confirmed that the

reduced pressure, invers of reduced temperature, acentric factor of the base fluids, and diameter

and volume concentration of the nano particles in base fluids are the most influential independent

variables on viscosity of nanofluids. Various statistical indices including mean square errors

‫٭‬
Corresponding Author
Email : vaferi@iaushiraz.ac.ir; behzad.vaferi@gmail.com (B. Vaferi)
Tel.: +989388419266; Fax: +987136410059

1
 ACCEPTED MANUSCRIPT

(MSE), root mean square errors (RMSE), average absolute relative deviation percent (AARD%),

and regression coefficient (R2) are used for numerical evaluation of accuracy of various models.

The results demonstrate that predictive accuracy of the multi-layer perceptron neural network

PT
(MLPNN) outperforms other intelligence/empirical models, and therefore it was considered as the

RI
best approach for the considered task. This model predicted the viscosity of various alumina-based

nanofluids with overall MSE=0.1422, RMSE=0.3797, AARD=4.13%, and R2=0.99947. Based on

SC
our best knowledge, this study is the only work that compared the performance of various

NU
intelligent/empirical paradigms for estimation of viscosity of various alumina-based nanofluids.
MA
Keywords: Nanofluids, viscosity; intelligence approaches; empirical correlations; predictive

accuracy
D
P TE
CE
AC

2
 ACCEPTED MANUSCRIPT

1. Introduction

Demand for increasing the energy efficiency of industrial processes has resulted in high attentions

on the technologies that enhance the performance of heat transfer devices/machines [1]. This

PT
enhancement may be achieved by either modifying the manufacturing techniques of heat devices

RI
or improving the thermo-physical properties of the operating fluids [2-3]. Since the traditional heat

transfer fluids have relatively poor thermo-physical properties, the idea of adding some sort of

SC
solid particles to improve their heat transfer characteristics was evolved [4]. Rapid progress of

NU
nano technology has motivated researchers to disperse various nano-scale particles (1-100 nm) in

the operating fluids to form a new class of heat transfer fluids namely nanofluids [5-11].
MA
Viscosity is important characteristic that has high influences on the both flow behavior and heat

transfer applications of the nanofluids. Accurate determination of viscosity is required for

D
TE

calculation/estimation of convection heat transfer coefficient, Reynolds and Prandtl numbers,

pressure drop and its associated required power for pumping.

P

Large amounts of theoretical/empirical approaches for estimation of viscosity of nanofluids are

CE

recommended up to now [12-21]. Einstein [12], Brinkman [13], Frankel and Acrivos [14], Nguyen
AC

et al. [15], Batchelor [16], Maiga et al. [17], Thomas and Muthukumar [18], Rea et al. [19],

Chandrasekar et al. [20], and Heyhat et al. [21] are some of the most well-known models of this

type. Many researchers have investigated the effect of operating temperature [22-25], size of nano-

particle [26-29], and volume concentration of nano-particles [30-34] on the viscosity of nanofluids.

1.1. Our study, novelty and its outline

Although, various AI-based models such as support vector machines [35-37], artificial neural

networks [38-40], genetic algorithm optimization methods [41], and neuro fuzzy models [42-43]

3
 ACCEPTED MANUSCRIPT

have been developed for prediction of thermo-physical properties of nanofluids, there are limited

works on prediction of viscosity of nanofluids using computer-aided approaches [35-36, 41].

Moreover, these researches focused on estimation of viscosity of alumina nanoparticle in specific

PT
base fluid, and proposed different models for different alumina-based nanofluids. But we proposed

RI
a single model for estimation of viscosity of alumina nanoparticle in various base fluids including

water, ethylene glycol, propylene glycol, different mixtures of water in ethylene glycol. Moreover,

SC
there exists no study that compared the performance of various intelligent/empirical paradigms for

NU
estimation of viscosity of various alumina-based nanofluids.

In the following sections, at first the most well-known empirical correlations for prediction of
MA
viscosity of suspensions/nanofluids are reviewed. Thereafter, short explanations about the

employed artificial intelligence models i.e., various types of artificial neural networks (ANN), and
D
TE

least squares support vector machines (LS-SVM) are given. Finally by comparing the predictive

accuracy of various intelligent/empirical approaches, the best model is found and its results are
P

reported.
CE
AC

2. Available correlations for viscosity of nanofluids

In this section a detailed review of the available empirical/theoretical formulas for prediction of

viscosity of suspensions as well as nanofluids is given, their strengths, limitations, and the ranges

of their applications are highlighted.

2.1. Model by Einstein

4
 ACCEPTED MANUSCRIPT

Using amount of required work for transportation of spherical solid particle through linearly

viscous fluids, Einstein proposed Eq. (1) for calculation of viscosity of liquid-solid homogeneous

mixtures [12].

PT
d  bf 1 2.5  (1)
where d and bf are viscosity of suspension and base fluid, respectively and  denotes volume

RI
SC
fraction of the solid particles in suspension. This equation can be used for estimation of viscosity

of dilute nanofluids containing low volume percent of spherical nano-particles (   0.02 % ).

NU
Since this model is assumed very slow flow, and not consider the inertia effect inside the fluids,

some researchers tried to modify this equation for improving its predictive accuracy.
MA
2.2. Model by Brinkman
D
TE

By considering the effect of both size and structure of particles on the viscosity of suspensions,

Brinkman has modified the Einstein’s model, and extended its ranges of application [13]. The
P

obtained model by Brinkman can be defined by Eq. (2).

CE

d  bf 12.5 (2)

AC

This model can be used for estimation of viscosity of suspension with moderate solid-particle

volume fraction (   0.04 ).

2.3. Model by Frankel and Acrivos

Considering the effect of maximum volume concentration of nano-particles in the base fluids,

Frankel and Acrivos proposed the following equation for calculation of viscosity of nanofluids

[14].

5
 ACCEPTED MANUSCRIPT

  / max  0.33 
d  1.125 bf  0.33 
1  / max  
(3)

here max indicates value of the maximum volume fraction of nano-particles in base fluids. This

PT
model is valid for calculation of viscosity of nanofluids containing broad range of concentration

of spherical nano-particles by considering 0.5236max  0.7405 .

RI
SC
2.4. Model by Nguyen et al.

NU
Nguyen et al. by comprehensive investigation of the effect of size and volume fractions of alumina

nanoparticle, proposed Eq. (4) for dynamic viscosity of alumina-water nanofluids [15].
MA
nf  bf 1  0.025  0.015 2  (4)

here nf is the dynamic viscosity of nanofluids. It has been claimed that this empirical correlation
D
TE

can be used in the range of 0.15 to 12 volume percent of nanoparticle in base fluid.
P

2.5. Model by Batchelor

CE

Batchelor studied the effect of Brownian motion of particles on shear stress of homogeneous
AC

suspensions containing spherical particles [16]. By this way Eq. (5) derived for calculation of

viscosity of nanofluids.

d  bf 1  2.5  6.5 2  (5)

In the case of very low particle volume concentration, this model simplifies to the Einstein’s

formula. The Batchelor’s model can calculate the viscosity of isotropic nanofluids containing

spherical nano-particles.

2.6. Model by Maiga et al.

6
 ACCEPTED MANUSCRIPT

By employing the least squared approach, Maiga et al. fitted Eqs. (6) and (7) on some experimental

data [17]. The proposed models by Eq. (6) and Eq. (7) are applicable for calculation of viscosity

of water-alumina, ethylene glycol - alumina nanofluids, respectively.

PT
nf  bf 1  7.3  123  2  (6)

nf  bf 1  0.19  306 2 

RI
(7)

SC
2.7. Model by Thomas and Muthukumar

NU
Thomas and Muthukumar developed Eq. (8) for calculation of viscosity of suspension containing

hard sphere solid particles [18]. The effect of solid hydrodynamics is considered during deriving
MA
this correlation.

d  bf 1  2.5  4.83 2  6.4 3 

D

(8)
TE

2.8. Model by Rea et al.

P

Rea et al. by fitting models on experimental data, developed some correlations for the calculation
CE

of the viscosity of alumina water-based nanofluid [19]. The proposed correlation i.e. Eq. (9) is
AC

claimed to be applicable for the up to 6% volumetric loadings of alumina naoparticles.

nf  bf exp4.91 0.2092   (9)

2.9. Model by Chandrasekar et al.

By focusing on mean free path between alumina nanoparticles, Chandrasekar et al. derived Eq.

(10) for estimation of dynamic viscosity of water-alumina nanofluid with the volume fraction of

0.0033 to 0.05 [20].

7
 ACCEPTED MANUSCRIPT

    
200

nf  bf 1  5  (10)
  1    


PT
2.10. Model by Heyhat et al.

RI
Heyhat et al. obtained an exponential correlation between the viscosity of alumina-water nanofluid

and its associated volume fraction [21]. This model that is expressed by Eq. (11) shows good

SC
accuracy for volume fractions and temperatures in the ranges of 0.001 to 0.02 and 20 to 60 °C,

NU
respectively.

 5.989 
MA
nf  bf exp  (11)
 0.278  
However, as it was previously explained, large parts of these empirical correlations are initially
D

developed for the suspension of microparticles, and very few of them have specially been
TE

developed based on the unique characteristics of the nano-particles and base fluids [44-45].
P

Moreover, the application of these empirical models is often limited to specific types of nanofluids,
CE

covered small ranges of nanoparticle sizes, and low amount of particle concentration [46-47].

Therefore, it can be said that large parts of these correlations are not enable to accurately predict
AC

the viscosity of nanofluids in wide ranges of operating conditions. A summary of the most well-

known correlations for estimation of dynamic viscosity of both suspensions and nanofluids is listed

in Table 1.

Table 1

3. Methods

8
 ACCEPTED MANUSCRIPT

In this section, short explanations for the considered AI paradigms are provided, their

mathematical background is reviewed, and their strength and weakness are highlighted from

practical point of view.

PT
RI
3.1. Artificial intelligence approaches

Artificial intelligence is a branch of science and technology which was stablished by mathematical

SC
simulation of procedure of thinking, learning, decision making, and working of human beings to

NU
solve problems [48-51]. Indeed these approaches try to extract even the most complicated

relationships among various variables and present them as block box intelligent entities [52-58].
MA
AI includes variety of subfields such as artificial neural networks, support vector machines,

adaptive neuro fuzzy interface, genetic algorithm and so on [59]. They results in significant and
D
TE

impressive progress in different branch of science and technology [60].

P

3.1.1. Artificial neural networks

CE

Simulation of working procedure of biological nervous system of the human’s brain is the basic
AC

idea for designing of artificial neural networks [61]. The smallest and the most common processing

element among various types of these networks is an artificial neuron or perceptron. This node has

two adjustable parameters i.e. weight ( w ) and bias ( b ). Artificial neuron receives information

from either external source or other nodes, and generates an output according to Eq. (12).

k 
out f  wr xr  b (12)
 r1 
where out denotes the perceptron’s output, and f is the activation function.

ANN models often require different numbers of this processing unit in their layers for solving

specific problems. Artificial neural networks can extract a function g

 : RInd RDep by training
9
 ACCEPTED MANUSCRIPT

on a dataset, where Ind and Dep indicates the number of dimensions for independent and

dependent variables, respectively. Given a set of independent variables X  Ind1 Ind2    Indn 
T

and their associated target variable(s), the parameters of ANN can be adjusted by the

PT
backpropagation training algorithm to solve even the most nonlinear problems [62-63]. ANN has

RI
generated a huge progress for solving of specific application including parameter estimation,

SC
pattern recognition, data classification, fault detection, speech recognition, computer vision and

text processing and so on [64].

NU
MA
3.1.1.1. Multilayer perceptron neural networks

Multilayer perceptron neural network is a feedforward supervised learning approach with one or
D

more hidden layers between independent and dependent variables. Feedforward means that the
TE

entry signals flow inside the neural network from input layer passing hidden layers to output layer.

Indeed MLP has multiple layers of nodes i.e. an input layer, one or more hidden layer(s), and an
P
CE

output layer. This type of ANN methodology is well-known tool for both classification and

regression problems [61-63].

AC

3.1.1.2. Cascade feedforward neural networks

Fahlman and Lebiere developed cascade feedforward neural network (CFNN) by modifying the

structure of the MLP network [65]. Instead of just synaptic connections between neurons of

adjacent layers, CFNN provides with synaptic connections between neuron of each layer with

neurons of all successive layers [66]. Learning process of the cascade feedforward neural networks

is claimed to be quicker than other topologies [67].

10
 ACCEPTED MANUSCRIPT

3.1.1.3. Radial basis neural networks

Radial basis function (RBF) neural network may be imagined as a two-layer feedforward network

with Gaussian and linear transfer function in hidden and output layer, respectively [68]. This type

PT
of ANN models which was developed by Broomhead and Lowe in 1988 are powerful techniques

RI
for interpolation in multidimensional space [69-70].

SC
3.1.2. Least square support vector machines

NU
Support vector machines (SVM) are supervised learning approaches that uses a convex

optimization problems, typically quadratic programming for solving classification as well as

MA
regression problems [71]. Originally, it has been proposed by Vapnik in the field of statistical

learning theory and minimization of structural risk [72]. Support Vector Machines is a powerful
D
TE

tool for solving nonlinear classification, function approximation as well as density estimation [73].

Least squares support vector machine (LS-SVM) which is a reformulations of SVM, solves a
P

system of linear equations for both regression and classification analyses [74].
CE

Similar to ANN model, the LS-SVM methodology has also some adjusting parameter namely
AC

regularization parameter (  ) and kernel parameter (  2 ). The first one determines the tradeoff

between the fitting error minimization and the smoothness of the estimated function and the second

parameter plays a crucial role in establishing an efficient LS-SVM model. Optimum values of

these parameters are adjusted during training stage.

3.2. Accuracy assessment of AI models

Several accuracy indices including mean square errors, root mean square errors, average absolute

relative deviation percent, and regression coefficient have been applied to compare an accuracy of

11
 ACCEPTED MANUSCRIPT

various models for prediction of viscosity of alumina-based nanofluids and finding the most

accurate one. Values of MSE, RMSE, AARD%, and R2 are mathematically calculated by Eq. (13)

to Eq. (16), respectively.

PT
MSE 
 i   i
N exp. cal. 2
(13)
i 1 N

RI

 N  exp. i   cal. i 2  
0.5

RMSE  

SC
(14)
 i1 N 

100 N   exp. i   cal. i 

N 

NU
AARD%    exp. i  (15)
i1  

 i  2  exp. i   cal. i2

N N
exp.
MA
R2  i 1 i 1
(16)
 i  
N
exp. 2

i 1
D

where  and N represent average value of viscosity of nanofluids and number of experimental
TE

data of viscosity of nanofluids, respectively.

P

It should be mentioned that a model that provides the smallest MSE, AARD%, RMSE, and R2
CE

close to one is considered as the most accurate model. In addition to these indices, the size of
AC

network is sometimes considered during finding the best ANN type.

4. Results and discussion

In this section, summary of experimental databank collected from nine different literatures is

presented, procedure of selection of the most influential independent variables using Pearson

correlation matrix is explained, the best structure of various AI model is determined, comparisons

among accuracies of various model are done, and finally the best model is selected and its results

are analyzed.

12
 ACCEPTED MANUSCRIPT

4.1. Experimental database

Both empirical correlations [12-21] and published literature data [22-30, 75-77] were confirmed

that the dynamic viscosity of nanofluids mainly depends on type of base fluid, operating

PT
conditions, and characteristics of nanoparticle.

RI
According to three-parameter corresponding state theory, it was common to define base fluids

based on their acentric factor, critical pressure, and critical temperature simultaneously [78]. Table

SC
2 reports the value of these parameters for different base fluids. It should be mentioned that the

NU
properties of water-ethylene glycol mixtures are calculated using the Kay’s method [79].

Table 2
MA
Temperature is the main operating condition that has high influence on the viscosity of both base

fluids and nanofluids. For nanoparticles, effects of diameter as well as its volume percent (Vp) in
D
TE

base fluid are also considered as independent variables.

Table 3 shows summary of the experimental databank including independent and dependent
P

variable(s), their ranges, and number of collected data points from different literatures [22-30]. It
CE

should be mentioned that, in almost all cases, preparation method is the same and no surfactant or
AC

dispersant were added. Therefore, the selection of data is based on the fact that the conditions are

almost identical in the tests and the similar parameters are evaluated.

Table 3

It can be seen from Table 3 that our databank constitutes of 674 experimental data points for

viscosity of different base fluids - alumina nanoparticle. It includes seven different base fluids in

temperature range of 273-345 K, covers up to 9.4 volume percent of alumina nanoparticle in base

fluids, alumina diameter of 8 to 50 nm. For these conditions dynamic viscosity was reported to be

in the range of 0.43 to 81.51 m Pa. The AI-based models could provide good prediction for

13
 ACCEPTED MANUSCRIPT

interpolation between these ranges of experimental data that were previously used for their

development. Similar to other regression-based models, AI-based models provides poor accuracy

for extrapolation.

PT
It should be mentioned that, since the alumina nanoparticle is a safe product, can be simply

RI
synthesized in large scale, and its suspensions in base fluids often show an excellent stability,

Al2O3-based nanofluids have considered as one of the most popular working fluids in various

SC
thermal processes.

NU
4.1.1. Scaling the data during preprocessing stage
MA
It should be mentioned that all of the experimental data are mapped to interval of [0.01 0.99]. This

is done for increasing the convergence rate during training stage, and avoiding of saturation of
D
TE

parameters for considered ANNs. Normalization is done using Eq. (17).

V Vmin
0.99  0.01
P

Vnormal 0.01 (17)

Vmax Vmin
CE

here, V can be any independent or dependent variable, Vmax and Vmin represent the maximum and
AC

minimum values of each variable, and Vnormal is associated normal value.

4.2. Independent variable selection using correlation matrix analysis

In this study it has been tried to rearrange the independent variables in such a way that they present

the highest relationship with the viscosity of nanofluids. To doing so the correlation matrix

analyses using the Pearson approach are applied for measuring the strength of relation between

transformed variables with the dependent variable [8, 38]. Pearson approach provides a value

between -1 and +1 for indicating both strength and kind of relation between each pair of variables.

14
 ACCEPTED MANUSCRIPT

The value of -1 indicates the highest inverse relationship, +1 indicates the most strength direct

relationship, while the zero value referred as no relationship between the considered variables.

Values of the correlation coefficients for various independent-dependent variables and their

PT
absolute average value are presented in Table 4A and Table 4B. Since the correlation coefficients

RI
for all of the independent variables have non-zero values, their selection is likely verified. Since it

is interested to use the most powerful relationship, it’s better to focus on their absolute average

SC
values. It can be simply seen that the rearranged independent variables in Table 4B show the

NU
highest absolute average value. Therefore, these transformed variables which provide more strong

relationship with the viscosity of nanofluid are selected as the independent variables.
MA
Table 4A

Table 4B
D
TE

Figure 1 simply illustrates the procedure of our study to find the best model for estimation of

viscosity of alumina-based nanofluids. It contain three main parts i.e. transformed independent
P

variables, the considered models, and finding the best model.

CE

Fig. 1
AC

4.3. Selection of configuration of the AI approaches

The parameter of LS-SVM model, number of hidden layers of ANN models and most importantly

number of neurons in each hidden layer should be determined by a systematic procedure [55, 56,

61, 63, 74, 80]. Literature has demonstrated that 95% of multi-variable system can be accurately

modeled by a single hidden layer neural network [81]. Accordingly, it was supposed that ANN

models having one hidden layer are able to estimate viscosity of nanofluids with acceptable

accuracy. On the other hand, the optimum numbers of hidden neurons of different types of ANN

15
 ACCEPTED MANUSCRIPT

model, and the parameters of LS-SVM model have tried to be determined by minimizing of

AARD%, MSE, and RMSE for both training ang testing datasets.

Our AI approaches were trained by 80% of experimental data (i.e. 540 data points), hereafter they

PT
called training dataset. The remaining 134 data points were used for evaluation of their

RI
performance by the data that were not seen by the models previously, i.e. testing dataset.

SC
4.3.1. MLPNN model

NU
The results of sensitivity accuracy analyses for finding the best number of hidden neurons for the

MLP network was summarized in Table 5. Various MLP networks with one to fifteen hidden
MA
neurons were designed, trained, tested, and their quantitative accuracy were calculated. The

optimum number of hidden neurons is often selected by finding the smallest MLP network that
D
TE

provides an acceptable accuracy.

It’s obvious that the efficiency of the MLP network improves by increasing the number of hidden
P

neuron up to fourteen, and thereafter worst result is observed for testing dataset even by developing
CE

a larger MLP approach and spending higher computational effort. Therefore, a single hidden layer
AC

feed-forward network containing 14 hidden neurons (the bold rows) was found to be the optimum

topology for prediction of viscosity of various alumina-based naofluids. It can be clearly seen that

the selected MLP network provides overall AARD%=4.13, MSE=0.1442, RMSE=0.3797, and

R2=0.99947.

Table 5

4.3.2. CFNN model

16
 ACCEPTED MANUSCRIPT

The optimum number of hidden neurons for the cascade feedforward neural network was also

found by trial and error procedure. Number of hidden neurons was increased up to fifteen, the

performances of various topologies were calculated, and the best one was selected. The results of

PT
sensitivity error analyses for the CFNN based on MSE, AARD%, RMSE, and R2 were reported in

RI
Table 6.

Table 6

SC
It is obvious that the CFNN with one hidden layer containing 14 hidden neurons (the bold rows)

NU
which shows an overall AARD%=4.22, MSE= 0.118, R2=0.9959, and RMSE=0.3343 is the best

topology.
MA
4.3.3. RBFNN model
D
TE

The best number of hidden neurons for radial basis neural network is determined by the same

procedure that was previously used for the MLP and CFNN models. Table 7 summarizes the results
P

of trial and error procedure for finding the most accurate topology of the RBFNN model. It can be
CE

simply observed that the 15 hidden neurons RBFNN model that provides AARD%=56.81,
AC

MSE=6.922, R2=0.9745, and RMSE=2.631 is the best topology amongst the considered structures

(the bold rows).

Table 7

It’s worthy to note that to eliminate the effect of initial guess of weights and biases on the

performance of the developed ANN networks, each topology in Table 5 to Table 7 was trained 20

times and only the best results were reported.

4.3.4. LS-SVM model

17
 ACCEPTED MANUSCRIPT

As mentioned earlier the LS-SVM model has two adjustable parameters. Therefore, for finding the

best values for these parameters the LS-SVM were trained and tested several times. The results of

various LS-SVM models differing from adjustable parameters were presented in Table 8.

PT
Table 8

It is clear that the LS-SVM model with  =7.183 and  2 =2.24×1010 has the best structure. This

RI
SC
developed LS-SVM model predicted the overall dataset with AARD%=6.92, MSE=0.4492,

R2=0.99834, and RMSE=0.6702.

4.4. Finding the best model

NU
MA
Since the aim of this study was to find the best AI approach for estimation of viscosity of various

alumina-based nanofluids, it’s common to compare the predictive accuracies of various AI

D
TE

approaches to find the best model. Table 9 reports results of the best topology of the MLP, CFNN,

RBFNN, and LS-SVM models for training, testing, and overall datasets in term of AARD%, MSE,
P

R2, and RMSE.

CE

Table 9
AC

It can be simply concluded that the MLP model which provides the best predictive accuracy for

estimation of viscosity of alumina-based nanofluids, is the best model amongst the considered AI

approaches (it’s highlighted by the bold font). On the other hands, the RBFNN shows the worst

predictions for the considered task.

4.5. Accuracy analysis of the best model

4.5.1. Prediction of experimental data

18
 ACCEPTED MANUSCRIPT

Figure 2 illustrates comparison between the experimental datasets for dynamic viscosity of

nanofluids and their corresponding estimated values by the optimum MLP network for both

training as well testing. Concentration the symbols for both training and testing groups around the

PT
45˚line confirms that the optimum MLP network predicted the experimental values of dynamic

RI
viscosity of alumina-based nanofluids with an acceptable accuracy.

Fig. 2

SC
NU
4.5.2. Parametric study

In this section effect of various independent variables on dynamic viscosity of alumina-based

MA
nanofluids are investigated both experimentally and theoretically. Moreover, some lines of

reasoning were provided for explanation the obtained results.

D
TE

4.5.2.1. Effect of temperature

P

Variation of viscosity of EG-alumina, water-alumina, as well as PG-alumina nanofluids as

CE

function of temperature for different volume percent of nanoparticles in base fluids are shown in
AC

Figs. 3A to 3C both experimentally and theoretically. It is clear that the viscosity of nanofluids

decreases by increasing temperature.

Heating and increasing the temperature of nanofluid generally supplies the molecules of fluid with

higher energy and increases their random motion. These are responsible for weakening of

intermolecular forces holding the fluid molecules and reducing resistance of the fluid to shearing

flow. Overall effect of these factors results in, reduction of nanofluids viscosity by increasing

temperature.

Fig. 3A

19
 ACCEPTED MANUSCRIPT

Fig. 3B

Fig. 3C

PT
4.5.2.2. Effect of volume fraction of nanoparticle

RI
Both experimental and calculated results of variation of viscosity of EG-alumina, water-alumina,

as well as PG-alumina nanofluids as function of volume percent of nanoparticles at different

SC
temperatures are depicted in Figs. 4A through 4C, respectively. These figures clearly demonstrate

NU
that the viscosity of various alumina-based nanofluids increase by increasing volume concentration

of nanoparticle. By increasing the concentration of nanoparticles in base fluid, the Van der Waals
MA
force of attraction significantly increases and the chance of agglomeration/aggregation of

nanoparticles increases [82-83]. Moreover, increasing of volume percent of nanoparticles in base

D
TE

fluid increases the drag force on both individual particles as well as the whole medium. Other

forces such as solvation, hydration, hydrophilic and hydrophobic forces become very important
P

and influence the rheology of nanofluids when the interparticle distance is reduced due to an
CE

increase in volume fraction [84].

AC

It seems that these are the main reasons for increasing viscosity of nanofluids by increasing the

concentration of nanoparticles.

Fig. 4A

Fig. 4B

Fig. 4C

4.5.2.3. Effect of nanoparticle diameter

20
 ACCEPTED MANUSCRIPT

Effect of nanoparticle diameter on the viscosity of water-alumina and ethylene glycol-alumina

nanofluid at 313 K and 1% nanoparticle loading is shown in Fig. 5. This figure shows that,

increasing the size of nanoparticle caused decreasing viscosity of nanofluid. Particle-fluid

PT
interaction and particle-particle interaction are two major interactions in fluid medium containing

RI
suspended nanoparticles. These interactions have been termed as a first and second electroviscous

effects, respectively [85-86]. Van der Waals force of attraction between particles is another

SC
prevalent interaction in the fluid. The electroviscous effect present in the nanofluid medium,

NU
determines the agglomeration and hence the degree of Brownian motion effects of the particles.

For constant nanoparticle concentration, smaller nanoparticles provide higher overall surface area
MA
between both solid-liquid and solid-solid systems. It can increase the electroviscous effects present

in the nanofluids, which, in turn, gives rise to an increase in viscosity. On the other hand, by
D
TE

reducing the overall surface area for solid-liquid and solid-solid, the electroviscous effects and

viscosity of nanofluid reduces.

P

Fig. 5
CE
AC

4.5.2.4. Effect of types of base fluid

Effect of base fluid on the viscosity of alumina-based nanofluid at (Dp= 36 nm, Vp=1.5%) is

shown in Fig. 6. This figure clearly shows that, viscosity of nanofluids will change with changing

the type of basefluid. It can be simply seen that pure propylene glycol has the highest dynamic

viscosity, and viscosity of pure water is minimum. It can also be seen that the dynamic viscosity

of various mixtures of water-EG continuously decreases by increasing the water content of the

mixture. Some properties of base fluid such as density and viscosity are definitely the factors that

affect the thermophysical properties of nanofluids such as viscosity and thermal conductivity.

21
 ACCEPTED MANUSCRIPT

Fig. 6

4.6. Comparison with various empirical correlations

PT
For justification the priority of using the AI approaches for estimation of viscosity of nanofluis

over the available empirical correlations, it’s better to compare the predictive accuracy of these

RI
two different methodologies. Therefore, the viscosity of aluimna-based nanofluids was predicted

SC
by ten well-known empirical correlations [12-21]. Table 10 reports the numerical values of

NU
provided AARD% by these empirical correlations for prediction of experimental data points of

various researchers. The developed model by Maiga et al. [17] is the best empirical correlation,
MA
while the model by Frankel and Acrivos [14] provides the worst results. The first one presented an

overall AARD%= 16.76, while the later one has shown AARD%=180.25.
D
TE

Table 10

Table 11 summarizes the provided AARD% by various AI approaches as well as the best results
P

obtained by the considered empirical correlations. It is clear that the best results of ten empirical
CE

correlations are only better than the RBFNN and predictive accuracy of other AI approaches
AC

superior than the empirical correlations.

Table 11

5. Conclusion

Nanofluids have recently found high popularity in various fields of engineering. Viscosity is one

of the most important thermophysical properties that directly has influence on both pressure drop

and required power of pumps. In this study, accuracy of various AI-based models including MLP,

RBFNN, CFNN, LS-SVM and ten empirical correlations for estimating the viscosity of alumina-

22
 ACCEPTED MANUSCRIPT

based nanofluids, were compared. The comparison is made using 674 experimental data points

collected from nine different literatures. The Pearson correlation matrix analysis approved that

working with transformed independent variables is better than rely on raw variables. Moreover,

PT
sensitivity accuracy analyses confirmed that the MLP paradigm including 14 hidden neurons is

RI
the best model for the considered task. This model provides an excellent agreement with

experimental data and predicts both individual experimental data and the trend of variation of

SC
viscosity with various independent variables. Overall MSE=0.1422, RMSE=0.3797,

NU
AARD=4.13%, and R2=0.99947 was observed between actual data and calculated results by the

MLP model. This developed model can simply incorporate in the mathematical model of fluid
MA
dynamics for prediction of both pressure drop and required power for pumping scenario.
D
P TE
CE
AC

23
 ACCEPTED MANUSCRIPT

Nomenclatures

b bias
D Diameter (nm)
h Inter-particle spacing (nm)

PT
N number of experimental data of viscosity of Nanofluids
P pressure (MPa)

RI
T temperature (K)
Vp volume percent of particle in base fluid (%)

SC
w Weight

NU
Abbreviations

AAC Average of absolute of coefficients

MA
AARD% average absolute relative deviation percent
AI artificial intelligence
ANN artificial neural networks
D

CFNN cascade feedforward neural networks

TE

EDL electric double layer

EG Ethylene Glycol
LS-SVM least squares support vector machines
P

MLP multi-layer perceptron

CE

MLPNN multi-layer perceptron neural networks

MSE mean squared errors
PG Propylene Glycol
AC

RBF radial basis neural networks

RMSE root mean square errors
R2-value regression coefficient
SVM support vector machines

Greek symbols

 viscosity (mPa.s)
 volume percent of solid particles in base fluids (%)

 regularization parameter
2 kernel parameter

24
 ACCEPTED MANUSCRIPT

Subscripts/superscripts

bf base fluid
c critical
cal. calculated variable

PT
d suspension
exp. experimental variable

RI
max maximum
nf Nanofluid

SC
normal normalized values
p particle

NU
MA
D
P TE
CE
AC

25
 ACCEPTED MANUSCRIPT

References

[1] M. Khoshvaght-Aliabadi, F. Hormozi, A. Zamzamian, Role of channel shape on performance

of plate-fin heat exchangers: Experimental assessment, Int. J. Therm. Sci. 79 (2014) 183-193.

PT
[2] C. Marques, K.W. Kelly, Fabrication and performance of a pin fin micro heat exchanger, J.

RI
Heat. Trans-T. ASME, 126 (2004) 434-444.

[3] G. Huminic, A. Huminic, Application of nanofluids in heat exchangers: a review, Renew. Sust.

SC
Energ. Rev. 16 (2012) 5625-5638.

NU
[4] H. Salimi-Yasar, S.Z. Heris, M. Shanbedi, A. Amiri, A. Kameli, Experimental investigation of

thermal properties of cutting fluid using soluble oil-based TiO2 nanofluid, Powder. Technol. 310
MA
(2017) 213-220.

[5] S.U.S. Choi, Enhancing thermal conductivity of fluids with nanoparticles, in D.A. Siginer, H.P.
D
TE

Wang (Eds.), Developments and applications of non-Newtonian flows, ASME, New York, 66

(1995) 99-105.
P

[6] L.S. Sundar, K.V. Sharma, M.T. Naik, M.K. Singh, Empirical and theoretical correlations on
CE

viscosity of nanofluids: A review, Renew. Sust. Energ. Rev. 25 (2013) 670-686.

AC

[7] H.W. Chiam, W.H. Azmi, N.A. Usri, R. Mamat, N.M. Adam, Thermal conductivity

and viscosity of Al2O3 nanofluids for different based ratio of water and ethylene

glycol mixture, Exp. Therm. Fluid. Sci. 81 (2017) 420-429.

[8] M.A. Ariana, B.Vaferi, G. Karimi, Prediction of thermal conductivity of alumina water-based

nanofluid by artificial neural networks, Powder. Technol. 278 (2015) 1-10.

[9] L. Yang, K. Du, X. Zhang, A theoretical investigation of thermal conductivity of nanofluids

with particles in cylindrical shape by anisotropy analysis, Powder. Technol. 314 (2017) 328-338.

26
 ACCEPTED MANUSCRIPT

[10] R. Mondragón, C. Segarra, R. Martínez-Cuenca, J.E. Juliá, J.C. Jarque, Experimental

characterization and modeling of thermophysical properties of nanofluids at high temperature

conditions for heat transfer applications, Powder. Technol. 249 (2013) 516-529.

PT
[11] L. Yang, J. Xu, K. Du, X. Zhang, Recent developments on viscosity and thermal conductivity

RI
of nanofluids, Powder. Technol. 317 (2017) 348-369.

[12] A. Einstein, Eine neue bestimmung der moleküldimensionen, Ann. Phys-Berlin. 324 (1906)

SC
289-306.

NU
[13] H.C. Brinkman, The viscosity of concentrated suspensions and solution, J. Chem. Phys. 20

(1952) 571-581.
MA
[14] N.A. Frankel, A. Acrivos, On the viscosity of a concentrate suspension of solid spheres,

Chem. Eng. Sci. 22 (1967) 847-853.

D
TE

[15] C.T. Nguyen, F. Desgranges, N. Galanis, G. Roy, T. Maré, S. Boucher, H. Angue Mintsa, T.

Maré, H.A. Mintsa, Viscosity data for Al2O3 - water nanofluid - hysteresis: is heat transfer
P

enhancement using nanofluids reliable?, Int. J. Therm. Sci. 47 (2008) 103-111.

CE

[16] G.K. Batchelor, The effect of Brownian motion on the bulk stress in a suspension of spherical
AC

particles, J. Fluid. Mech. 83 (1977) 97-117.

[17] S.E.B. Maiga, C.T. Nguyen, N. Galanis, G. Roy, Heat transfer behaviours of nanofluids in a

uniformly heated tube, Superlattice. Microst. 35 (2004) 543-557.

[18] C.U. Thomas, M. Muthukumar, Three-body hydrodynamic effects on viscosity of suspensions

of spheres, J. Chem. Phys. 94 (1991) 5180-5189.

[19] U. Rea, T. McKrell, L.W. Hu, J. Buongiorno, Laminar convective heat transfer and viscous

pressure loss of alumina-water and zirconia-water nanofluids, Int. J. Heat. Mass. Tran 52 (2009)

2042-2048.

27
 ACCEPTED MANUSCRIPT

[20] M. Chandrasekar, S. Suresh, A.C. Bose, Experimental investigations and theoretical

determination of thermal conductivity and viscosity of Al2O3/water nanofluid, Exp. Therm. Fluid.

Sci. 34 (2010) 210-216.

PT
[21] M.M. Heyhat, F., Kowsary, A.M. Rashidi, M.H. Memenpour, A. Amrollahi, M.H.

RI
Momenpour, Experimental investigation of laminar convective heat transfer and pressure drop of

water-based Al2O3 nanofluids in fully developed flow regime, Exp. Therm. Fluid. Sci. 44 (2013)

SC
483-489.

NU
[22] M.J. Pastoriza-Gallego, L. Lugo, J.L. Legido, M.M. Piñeiro, Thermal conductivity and

viscosity measurements of ethylene glycol-based Al2O3 nanofluids, Nanoscale. Res. Lett. 6 (2011)
MA
1-11.

[23] M.A. Hachey, C.T. Nguyen, N. Galanis, Catalin V. Popa, Experimental investigation of Al2O3
D
TE

nanofluids thermal properties and rheology - Effects of transient and steady-state heat exposure,

Int. J. Therm. Sci. 76 (2014) 155-167.

P

[24] C.T. Nguyen, F. Desgranges, G. Roy, N. Galanis, T. Mare, S. Boucher, H. Angue Mintsa,
CE

Temperature and particle-size dependent viscosity data for water-based nanofluids – Hysteresis
AC

phenomenon, Int. J. Heat. Fluid. Fl. 28 (2007) 1492-1506.

[25] B.C. Pak, Y.I. Cho, hydrodynamic and heat transfer study of dispersed fluids with submicron

metallic oxide particles, Exp. Heat. Transfer. 11 (1998) 151-170.

[26] J.H. Lee, Effective viscosities and thermal conductivities of aqueous nanofluids containing

low volume concentrations of Al2O3 nanoparticles, Int. J. Heat. Mass. Tran 51 (2008) 2651-2656.

[27] C.J. Ho, L.C. Wei, Z.W. Li, An experimental investigation of forced convective cooling

performance of a microchannel heat sink with Al2O3/water nanofluid, Appl. Therm. Eng. 30 (2010)

96-103.

28
 ACCEPTED MANUSCRIPT

[28] L. Syam Sundar, E. Venkata Ramana, Manoj K. Singh, Antonio C.M. Sousa, Thermal

conductivity and viscosity of stabilized ethylene glycol and water mixture Al2O3 nanofluids for

heat transfer applications: An experimental study, Int. Commun. Heat. Mass 56(2014) 86-95.

PT
[29] W. Yu, H. Xie, Y. Li, L. Chen, Q. Wang, Experimental investigation on the heat transfer

RI
properties of Al2O3 nanofluids using the mixture of ethylene glycol and water as base fluid,

Powder. Technol. 230 (2012) 14-9.

SC
[30] R. Prasher, D. Song, J. Wang, P. Phelan, Measurements of nanofluid viscosity and its

NU
implications for thermal applications, Appl. Phys. Lett. 89 (2006) 133108-1.

[31] W. Yu, S. Choi, The role of interfacial layers in the enhanced thermal conductivity of
MA
nanofluids: a renovated Hamilton-Crosser model, J. Nanopart. Res. 6 (2004) 355-361.

[32] N, Ahammed, L.G. Asirvatham, S. Wongwises, Effect of volume concentration and

D
TE

temperature on viscosity and surface tension of graphene-water nanofluid for heat transfer

applications, J. Therm. Anal. Calorim. 123 (2016) 1399-1409.

P

[33] X. Wang, X. Xu, S.U.S. Choi, Thermal conductivity of nanoparticle-fluid mixture, J.

CE

Thermophys. Heat. Tr. 13 (1999) 474-480.

AC

[34] M.H.U. Bhuiyan, R. Saidur, A. Amalina, R.M. Mostafizur, A. Islam, effect of nanoparticles

concentration and their sizes on surface tension of nanofluids, Procedia. Eng. 105 (2015) 431-437.

[35] M. Mehrabi, M. Sharifpur, J.P. Meyer, Viscosity of nanofluids based on an artificial

intelligence model, Int. Commun. Heat. Mass. 43 (2013) 16-21.

[36] M. Kalantari Meybodi, S. Naseri, A. Shokrollahi, A. Daryasafar, Prediction of viscosity of

water-based Al2O3, TiO2, SiO2, and CuO nanofluids using a reliable approach, Chemometr. Intell.

Lab. 149 (2015) 60-69.

29
 ACCEPTED MANUSCRIPT

[37] M. Sahooli, S. Sabbaghi, R. Maleki, M. M. Nematollahi, Thermal conductivity of Water-

based nanofluids: prediction and comparison of models using machine learning, Int. J. Nano.

Dimens. 5(2014) 47-55.

PT
[38] B. Vaferi, F. Samimi, E. Pakgohar, D. Mowla, Artificial neural network approach for

RI
prediction of thermal behavior of nanofluids flowing through circular tubes, Powder. Technol. 267

(2014) 1-10.

SC
[39] F. Yousefi, Z. Amoozandeh, A new model to predict the densities of nanofluids using

NU
statistical mechanics and artificial intelligent plus principal component analysis, Chinese. J. Chem.

Eng. 25 (2017) 1273-1281.

MA
[40] M. Hassanpour, B. Vaferi, M.E. Masoumi, Estimation of pool boiling heat transfer coefficient

of alumina water-based nanofluids by various artificial intelligence (AI) approaches, Appl. Therm.
D
TE

Eng. Accepted Manuscript.

[41] H. Karimi, F. Yousefi, M. Rahimi, Correlation of viscosity in nanofluids using genetic

P

algorithm-neural network (GA-NN), Heat. Mass. Transfer. 47 (2011) 1417-1425.

CE

[42] M. Mehrabi, M. Sharifpur, J.P. Meyer, Application of the FCM-based neuro-fuzzy inference
AC

system and genetic algorithm-polynomial neural network approaches to modelling the thermal

conductivity of alumina-water nanofluids, International Communn. Heat. Mass. 39 (2012) 971-

977.

[43] H. Salehi, S. Zeinali-Heris, M. Esfandyari, M. Koolivand, Nero-fuzzy modeling of the

convection heat transfer coefficient for the nanofluid, Heat. Mass. Transfer. 49 (2013) 575-583.

[44] N. Masoumi, N. Sohrabi, A. Behzadmehr, A new model for calculating the effective viscosity

of nanofluids, J. Phys. D. Appl. Phys. 42 (2009) 0555011 - 0555016.

30
 ACCEPTED MANUSCRIPT

[45] S. Hosseini, A. Moghadassi, D.E. Henneke, A new dimensionless group model for

determining the viscosity of nanofluids, J. Therm. Anal. Calorim. 100 (2010) 873-877.

[46] P.K. Namburu, D.P. Kulkarni, D. Misra, D.K. Das, Viscosity of copper oxide nanoparticles

PT
dispersed in ethylene glycol and water mixture, Exp. Therm. Fluid. Sci. 32 (2007) 397-402.

RI
[47] M.N. Rashin, J. Hemalatha, Viscosity studies on novel copper oxide - coconut oil nanofluid,

Exp. Therm. Fluid. Sci. 48 (2013) 67-72.

SC
[48] A. Canakci, S. Ozsahin, T. Varol, Modeling the influence of a process control agent on the

NU
properties of metal matrix composite powders using artificial neural networks, Powder. Technol.

228 (2012) 26-35.

MA
[49] S. Russell, N. Peter, Artificial Intelligence, a modern approach, Prentice-Hall, Englewood

Cliffs, New Jersey, 1995.

D
TE

[50] T.J. Bench-Capon, P.E. Dunne, Argumentation in artificial intelligence, Artif. intel. 171

(2007) 619-641.
P

[51] J. Kokina, T.H. Davenport, the emergence of artificial intelligence: how automation is
CE

changing auditing, J. Emerg. Technol. Account. 14 (2017) 115-122.

AC

[52] G. Weiss, Multiagent systems: a modern approach to distributed artificial intelligence, MIT

press, Massachusetts, 1999.

[53] D. Hassabis, Artificial Intelligence: Chess match of the century, Nature. 544 (2017) 413-414.

[54] D.T. Bui, Q.T. Bui, Q.P. Nguyen, B. Pradhan, H. Nampak, P.T. Trinh, A hybrid artificial

intelligence approach using GIS-based neural-fuzzy inference system and particle swarm

optimization for forest fire susceptibility modeling at a tropical area, Agr. Forest. Meteorol. 233

(2017) 32-44.

31
 ACCEPTED MANUSCRIPT

[55] B. Vaferi, R. Eslamloueyan, S. Ayatollahi, Application of recurrent networks to classification

of oil reservoir models in well-testing analysis, Energ. Source. Part. A. 37(2015 174-180.

[56] B. Vaferi, R. Eslamloueyan, S. Ayatollahi, Automatic recognition of oil reservoir models

PT
from well testing data by using multi-layer perceptron networks, J. Petrol. Sci. Eng. 77 (2011) 254-

RI
262.

[57] M. Sheikholeslami, F.B. Sheykholeslami, S. Khoshhal, H. Mola-Abasia, D.D. Ganji, H.B.

SC
Rokni, Effect of magnetic field on Cu-water nanofluid heat transfer using GMDH-type neural

NU
network, Neural. Comput. Appl. 25 (2014) 171-178.

[58] B. Vaferi, M. Karimi, M. Azizi, H. Esmaeili, Comparison between the artificial neural
MA
network, SAFT and PRSV approach in obtaining the solubility of solid aromatic compounds in

supercritical carbon dioxide, J. Supercrit. Fluid. 77 (2013) 44-51.

D
TE

[59] D.S.H.Y. Gan, P.G.M.M. Gromiha, Advanced intelligent computing theories and

applications, 7th International Conference, ICIC 2011, Zhengzhou, China, 2007.

P

[60] E. Mjolsness, D. DeCoste, Machine learning for science: state of the art and future prospects,
CE

Science. 293 (2001) 2051-2055.

AC

[61] B. Vaferi, M. Bahmani, P. Keshavarz, D. Mowla, Experimental and theoretical analysis of

the UV/H2O2 advanced oxidation processes treating aromatic hydrocarbons and MTBE from

contaminated synthetic wastewaters, J. Environ. Chem. Eng. 2 (2014) 1252-1260.

[62] N. Ghaffarian, R. Eslamloueyan, B. Vaferi, Model identification for gas condensate reservoirs

by using ANN method based on well test data, J. Petrol. Sci. Eng. 123 (2014) 20-29.

[63] B. Vaferi, R. Eslamloueyan, N. Ghaffarian, Hydrocarbon reservoir model detection from

pressure transient data using coupled artificial neural network-Wavelet transform approach Appl.

Soft. Comput. 47 (2016) 63-75.

32
 ACCEPTED MANUSCRIPT

[64] H.B. Demuth, M.H. Beale, O. De Jess, M.T. Hagan, Neural network design, Martin Hagan,

Oklahoma, 2014.

[65] E.S. Fahlman, C. Lebiere, The cascade-correlation learning architecture, Advances in neural

PT
information processing systems, 2 D. S. Touretzky, Ed. San Francisco, 1990.

RI
[66] M. Lashkarbolooki, B. Vaferi, A. Shariati, A.Z. Hezave, Investigating vapor-liquid equilibria

of binary mixtures containing supercritical or near-critical carbon dioxide and a cyclic compound

SC
using cascade neural network, Fluid .Phase. Equilibri. 343 (2013) 24-29.

NU
[67] G. Cheng, J. Zhou, X.J. Zhang, Z.J. Zhang, A daily load forecasting method based on cascaded

back propagation and radial basis function neural networks. Power. Syst. Technol. 33 (2009) 101-
MA
105.

[68] S. Chen, C.F.N. Cowan, P.M. Grant, Orthogonal least squares learning algorithm for radial
D
TE

basis function networks, IEEE. T. Neural. Networ. 2 (1991) 302-309.

[69] D.S. Broomhead, D. Lowe, Multivariable functional interpolation and adaptive networks,
P

Comp. Sys. 2 (1988) 321-355.

CE

[70] N. Zhao, X. Wen, J. Yang, S. Li, Z. Wang, Modeling and prediction of viscosity of water-
AC

based nanofluids by radial basis function neural networks, Powder. Technol. 281 (2015) 173-183.

[71] M.A. Hearst, S.T. Dumais, E. Osuna, J. Platt, B. Scholkopf, Support vector machines, IEEE.

Intell. Syst. 13 (1998) 18-28.

[72] V. Vapnik, the nature of statistical learning theory, Springer-Verlag, New York, 1995.

[73] P. Mantero, G. Moser, S.B. Serpico, Partially supervised classification of remote sensing

images through SVM-based probability density estimation, IEEE. T. Geosci. Remote. 43 (2005)

559-570.

33
 ACCEPTED MANUSCRIPT

[74] K. Pelckmans, J.A. Suykens, T. Van Gestel, J. De Brabanter, L. Lukas, B. Hamers, B. De

Moor, J. Vandewalle, LS-SVMlab: a matlab/c toolbox for least squares support vector machines,

Tutorial. KULeuven-ESAT. Leuven, Belgium, 2002.

PT
[75] L. Yang, H. Yuhan, Toward TiO2 Nanofluids-Part 1: Preparation and Properties, Nanoscale

RI
Res. Lett. 12 (2017) 417.

SC
[76] L. Yang, H. Yuhan, Toward TiO2 Nanofluids—Part 2: Applications and Challenges,

Nanoscale Res. Lett. 12 (2017) 446.

NU
[77] L. Yang, W. Jiang, X. Chen, K. Du, Dynamic characteristics of an environment-friendly
MA
refrigerant: ammonia-water based TiO2 nanofluids, International Journal of Refrigeration.

Accepted Manuscript.
D
TE

[78] G.F. Carruth, R. Kobayashi, Extension to low reduced temperatures of three-parameter

corresponding states: vapor pressures, enthalpies and entropies of vaporization, and liquid fugacity
P

coefficients. Ind. Eng. Chem. Fundam. 11 (1972) 509-517.

CE

[79] B.E. Poling, J.M. Prausnitz, J.P. O'connell, The properties of gases and liquids, Mcgraw-Hill,
AC

New York, 2001.

[80] H. Adib, A. Sabet, A. Naderifar, M. Adib, M. Ebrahimzadeh, Evolving a prediction model

based on machine learning approach for hydrogen sulfide removal from sour condensate of south

pars natural gas processing plant, J. Nat. Gas. Sci. Eng. 27 (2015) 74-81.

[81] G. Cybenko, Approximation by superpositions of a sigmoidal function, Math. Control. Signal.

2 (1989) 303-314.

34
 ACCEPTED MANUSCRIPT

[82] K.B. Anoop, S. Kabelac, T. Sundararajan, S.K. Das, Rheological and flow characteristics of

nanofluids: Influence of electroviscous effects and particle agglomeration, J. Appl. Phys. 106

(2009) 034909.

PT
[83] H. Chen, S. Witharana, Y. Jin, C. Kim, Y. Ding, Predicting thermal conductivity of liquid

RI
suspensions of nanoparticles (nanofluids) based on rheology, Particuology. 7 (2009) 151-157.

[84] R.G. Larson, The structure and rheology of complex fluids, Oxford University Press, Inc.,

SC
New York, 1999.

NU
[85] S.M.S. Murshed, K.C. Leong, C. Yang, Determination of the effective thermal diffusivity of

nanofluids by the double hot-wire technique, J. Phys. D. Appl. Phys. 39 (2006) 5316-5322.
MA
[86] J.P. Meyer, S.A. Adio, M. Sharifpur, The Viscosity of Nanofluids: A Review of the

Theoretical, Empirical, and Numerical Models, Heat. Transfer. Eng. 37 (2016) 387-421.
D
P TE
CE
AC

35
 ACCEPTED MANUSCRIPT

Figures caption:

Fig. 1. Flowchart of finding the best model for prediction of viscosity of alumina-based nanofluids

Fig. 2. Performance of the best AI model for estimation of viscosity of alumina-based nanofluids

PT
Fig. 3A. Influence of variation of temperature on the viscosity of EG-alumina nanofluid (Dp= 43

RI
nm)

Fig. 3B. Effect of variation of temperature on the viscosity of water-alumina nanofluid (Dp= 47

SC
nm)

NU
Fig. 3C. Effect of variation of temperature on the viscosity of PG-alumina nanofluid (Dp= 50 nm)

Fig. 4A. Influence of variation of volume percent of nanoparticle on the viscosity of EG-alumina
MA
nanofluid (Dp= 43 nm)

Fig. 4B. Influence of variation of volume percent of nanoparticle on the viscosity of water-alumina
D
TE

nanofluid (Dp= 30 nm)

Fig. 4C. Effect of variation of volume percent of nanoparticle on the viscosity of PG-alumina
P

nanofluid (Dp= 27 nm)

CE

Fig. 5. Effect of nanoparticle diameter on the viscosity of alumina-based nanofluid (T= 313 K,
AC

Vp=1%)

Fig. 6. Effect of temperature on the viscosity of base fluids (without nanoparticle) and alumina-

based nanofluid (Dp= 36 nm, Vp=1.5%)

36
 ACCEPTED MANUSCRIPT

PT
RI
SC
NU
MA
D
TE
P
CE
AC

Fig. 1

37
 ACCEPTED MANUSCRIPT

PT
RI
80 Training dataset

SC
Testing dataset
Perfect fit
70
Predicted viscosity of nanofluid (mPa.s)

NU
60
MA
50

40
D
TE

30
P

20
CE

10
AC

10 20 30 40 50 60 70 80
Experimental viscosity of nanofluid (mPa.s)

Fig. 2

38
 ACCEPTED MANUSCRIPT

PT
RI
90
Exp [22]. Vp=0.5%
Calc. Vp=0.5%

SC
80 Exp [22]. Vp=1%
Calc. Vp=1%

NU
70 Exp [22]. Vp=1.5%
Viscosity of nanofluid (mPa.s)

Calc. Vp=1.5%
60 Exp [22]. Vp=2.1%
MA
Calc. Vp=2.1%
Exp [22]. Vp=3.1%
50
Calc. Vp=3.1%
D

Exp [22]. Vp=4.8%

40 Calc. Vp=4.8%
TE

Exp [22]. Vp=6.6%

30 Calc. Vp=6.6%
P
CE

20

10
AC

0
280 285 290 295 300 305 310 315 320 325
Temperature (K)

Fig. 3A

39
 ACCEPTED MANUSCRIPT

PT
RI
5
Exp [24]. Vp=1%

SC
4.5 Calc. Vp=1%
Exp [24]. Vp=4%
4 Calc. Vp=4%

NU
Exp [24]. Vp=7%
Viscosity of nanofluid (mPa.s)

3.5 Calc. Vp=7%

Exp [24]. Vp=9.4%
MA
3 Calc Vp=9.4%

2.5
D
TE

1.5
P
CE

0.5
AC

0
290 300 310 320 330 340 350
Temperature (K)

Fig. 3B

40
 ACCEPTED MANUSCRIPT

PT
40
Exp [30]. Vp=0.5%

RI
Calc. Vp=0.5%
35 Exp [30]. Vp=2%

SC
Calc. Vp=2%
Exp [30]. Vp=3%
Viscosity of nanofluid (mPa.s)

30

NU
Calc. Vp=3%
MA
25

20
D
TE

15
P
CE

10
AC

5
300 305 310 315 320 325 330 335
Temperature (K)

Fig. 3C

41
 ACCEPTED MANUSCRIPT

PT
RI
40
Exp [22]. 298 K

SC
Calc. 298 K
35 Exp [22]. 303 K
Calc. 303 K

NU
Viscosity of nanofluid (mPa.s)

Exp [22]. 313 K

30
Calc. 313 K MA
Exp [22]. 323 K
25 Calc. 323 K
D

20
TE

15
P
CE

10
AC

5
0 1 2 3 4 5 6 7
Nanoparticle volume percent (%)

Fig. 4A

42
 ACCEPTED MANUSCRIPT

PT
1.05
Exp [26]. 295 K

RI
Calc. 295 K
1

SC
Exp [26]. 300 K
Calc. 300 K
0.95 Exp [26]. 310 K
Viscosity of nanofluid (mPa.s)

NU
Calc. 300 K
Exp [26]. 312 K
0.9
MA
Calc. 312 K

0.85
D

0.8
TE

0.75
P
CE

0.7
AC

0.65
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Nanoparticle volume percent (%)

Fig. 4B

43
 ACCEPTED MANUSCRIPT

PT
40
Exp [30]. 303 K

RI
Calc. 303 K
35 Exp [30]. 313 K

SC
Calc. 313 K
Exp [30]. 323 K
Viscosity of nanofluid (mPa.s)

30

NU
Calc. 323 K
Exp [30]. 333 K
Calc. 333 K
25
MA
20
D
TE

15
P

10
CE
AC

5
0 0.5 1 1.5 2 2.5 3 3.5 4
Nanoparticle volume percent (%)

Fig. 4C

44
 ACCEPTED MANUSCRIPT

PT
RI
14

SC
12

NU
Viscosity of nanofluid (mPa.s)

10 MA
8
D

6
TE

4
P
CE

Exp. EG-alumina [22]

2 Calc. EG-alumina
Exp. Water-alumina [24, 27]
AC

Calc. Water-alumina
0
5 10 15 20 25 30 35 40 45 50
Nanoparticle diameter (nm)

Fig. 5

45
 ACCEPTED MANUSCRIPT

PT
RI
40
Pure PG - Base fluid

SC
Pure EG, Base fluid
35 Exp [28]. (60% EG in water) - Nanofluid
Calc. (60% EG in water) - Nanofluid

NU
30 Exp [28]. (40% EG in water) - Nanofluid
MA Calc. (40% EG in water) -Nanofluid
45% EG in water - Base fluid
Viscosity (mPa.s)

25
Exp [28]. (20% EG in water) - Nanofluid
Calc. (20% EG in water) - Nanofluid
20
D

Pure water - Base fluid

TE

15
P

10
CE

5
AC

0
270 280 290 300 310 320 330 340 350 360 370
Temperature (K)

Fig. 6

46
 ACCEPTED MANUSCRIPT

Tables caption:

Table 1. Summary of the most well-known empirical correlations for viscosity of

suspension/nanofluids

PT
Table 2. Critical conditions of the employed base fluids

RI
Table 3. Summary of experimental databank for viscosity of alumina-based nanofluids

Table 4A. Coefficients of correlation matrix analyses between independent variables and

SC
viscosity of nanofluid

NU
Table 4B. Coefficients of correlation matrix analyses between independent variables and

viscosity of nanofluids
MA
Table 5. Trial and error procedure for finding the best structure for MLPNN model

Table 6. Sensitivity error analyses for finding the best topology of CFNN model
D
TE

Table 7. Finding the best structure of RBFNN model through statistical error analyses

Table 8. Evaluation the best adjusted parameters for LS-SVM model

P

Table 9. Comparison among the capabilities of different AI approach in prediction of viscosity

CE

of nanofluids
AC

Table 10. Provided AARD% for prediction of experimental datasets by different empirical

correlations

Table 11. Provided AARD% for prediction of experimental datasets by different methodologies

47
 ACCEPTED MANUSCRIPT

Table 1

Model Correlation Application Ref.

PT
Einstein d  bf 1 2.5  Viscosity of dilute suspension of rigid solid spheres, applicable for  < 0.02 [12]

RI
Brinkman d  bf 12.5 Viscosity of suspension with moderate concentration, applicable for  < 0.04 [13]

SC
  / max  0.33 
d  1.125 bf  0.33 
Viscosity of suspension with uniform solid sphere, for 0.5236< max <0.7405

NU
1  / max  
Frankel and Acrivos [14]

nf  bf 1  0.025  0.015 2  Viscosity of Al2O3 /water nanofluids, for 0.0015<  <0.12

MA
Nguyen et al. [15]

Batchelor d  bf 1  2.5  6.5 2  Spherical particles and dilute suspension [16]

ED
Maiga et al. nf  bf 1  7.3  123  2  Viscosity of Al2O3 /water nanofluids [17]

Thomas and Muthukumar

PT
d  bf 1  2.5  4.83 2  6.4 3  Viscosity of suspension with solid sphere [18]
CE

Rea et al. nf  bf exp4.91 0.2092   Viscosity of Al2O3 /water nanofluids, for  <0.06 [19]
AC

   
200

nf  bf 1  5   Viscosity of Al2O3 /water nanofluids, for 0.0033<  <0.05
Chandrasekar et al.
  1   [20]
 

 5.989 
Heyhat et al. nf  bf exp  Viscosity of Al2O3 /water nanofluids, for 0.001<  <0.02 [21]
 0.278  

48
 ACCEPTED MANUSCRIPT

PT
RI
SC
NU
Table 2

Base fluid Tc (K) Pc (MPa) Acentric factor

MA
Ethylene glycol (EG) 719.7 7.71 0.487

Water 647.1 22.06 0.343

D

EG/Water (20/80) 652 21.09 0.353

TE

EG/Water (40/60) 658.9 19.73 0.366

EG/Water (45/55) 661 19.31 0.371

P

EG/water (60/40) 669.1 17.71 0.387

CE

Propylene Glycol (PG) 626 6.04 1.102

AC

49
 ACCEPTED MANUSCRIPT

PT
RI
SC
Table 3

Base fluid Temperature (K) Vp (%) Dp (nm)  (mPa.s) No. of data Ref.

NU
Ethylene glycol (EG) 283-323 0.5-6.6 8-43 7.51-81.51 96 [22]

Ethylene glycol (EG) 298-328 1-5 10 8.14-37.28 21 [23]

MA
Water 294-343 1-9.4 47 0.43-4.91 81 [24]

Water 293-345 1.34-2.78 13 0.63-2.49 14 [25]

D

Water 294-312 0.01-0.3 30 0.66-1.00 114 [26]

TE

Water 293-313 1-2 33 0.65-1.09 10 [27]

EG/Water (20/80) 273-333 0-1.5 36 0.59-4.6 89 [28]

P

EG/Water (40/60) 273-333 0-1.5 36 0.96-13.64 90 [28]

CE

EG/Water(45/55) 283-333 0-2 30 1.54-11.08 33 [29]

EG/water(60/40) 273-333 0-1.5 36 1.5-35.35 90 [28]

AC

Propylene Glycol (PG) 303-333 0.5-3 27-50 7.9-38.6 36 [30]

Overall ranges 273-345 0-9.4 8-50 0.43-81.51 674 [22-30]

50
 ACCEPTED MANUSCRIPT

PT
RI
Table 4A

SC
Temperature (K) Tc (K) Pc (MPa) Acentric factor Vp (%) Dp (nm) AAC*

-0.3399 0.6232 -0.7576 0.4333 0.2044 -0.0821 0.4068

NU
*
Average of absolute of coefficients

MA
D
TE

Table 4B
P

Tc /T 1/Pc (MPa) Acentric factor vol.% Dp (nm) AAC

CE

0.6241 0.7135 0.4333 0.2044 -0. 0821 0.4938

AC

51
 ACCEPTED MANUSCRIPT

Table 5

Number of hidden Sensitivity accuracy analyses

Database
neurons
AARD% MSE R2 RMSE

PT
Training 24.62 3.4708 0.98488 1.8630
2 Testing 27.07 6.5978 0.99033 2.5686
Overall 24.99 3.9394 0.98562 1.9848

RI
Training 17.02 1.5977 0.99414 1.2640

SC
4 Testing 17.79 1.5941 0.99402 1.2626
Overall 17.14 1.5972 0.99409 1.2638
Training 12.94 0.5584 0.99795 0.7473

NU
6 Testing 15.09 0.7587 0.99726 0.8710
Overall 13.26
MA 0.5884 0.99783 0.7671
Training 8.80 0.3267 0.99876 0.5715
8 Testing 10.79 1.0471 0.99723 1.0233
Overall 9.10 0.4346 0.99841 0.6593
Training 8.20 0.1343 0.99946 0.3665
D

9 Testing 9.64 0.2289 0.99961 0.4784

TE

Overall 8.41 0.1485 0.99946 0.3853

Training 6.55 0.1183 0.99956 0.3440
P

10 Testing 7.81 0.1864 0.99940 0.4318

Overall 6.74 0.1285 0.99953 0.3585
CE

Training 6.53 0.1263 0.99956 0.3554

11 Testing 8.19 1.7631 0.99352 1.3278
AC

Overall 6.78 0.3716 0.99866 0.6096

Training 5.63 0.0906 0.99968 0.3011
12 Testing 6.04 0.2209 0.99891 0.4700
Overall 5.69 0.1102 0.99959 0.3319
Training 4.38 0.0859 0.99969 0.2931
13 Testing 6.30 0.3259 0.99869 0.5709
Overall 4.67 0.1219 0.99955 0.3491
Training 4.11 0.1025 0.99962 0.3202
14 Testing 4.22 0.3804 0.99867 0.6167
Overall 4.13 0.1442 0.99947 0.3797
Training 4.72 0.0799 0.99971 0.2827
15 Testing 5.91 0.3091 0.99861 0.5560
Overall 4.90 0.1143 0.99958 0.3381

52
 ACCEPTED MANUSCRIPT

Table 6

Number of hidden Sensitivity accuracy analyses

Database
neurons
AARD% MSE R2 RMSE

PT
Training 25.33 3.4364 0.98767 1.8537
2 Testing 27.96 4.4143 0.97988 2.1010
Overall 25.72 3.5829 0.98670 1.8929

RI
Training 22.12 1.9388 0.99322 1.3924

SC
4 Testing 28.41 3.2538 0.98281 1.8038
Overall 23.06 2.1359 0.99209 1.4615
Training 12.37 0.5823 0.99778 0.7631

NU
6 Testing 14.42 1.4567 0.99595 1.2070
Overall 12.68
MA 0.7133 0.99737 0.8446
Training 6.21 0.1587 0.99939 0.3983
8 Testing 8.08 1.5538 0.99645 1.2465
Overall 6.49 0.3677 0.99868 0.6064
Training 5.37 0.2134 0.99924 0.4619
D

9 Testing 6.01 0.2912 0.99869 0.5396

TE

Overall 5.47 0.2250 0.99917 0.4744

Training 8.68 0.1885 0.99933 0.4342
P

10 Testing 9.73 0.3327 0.99843 0.5768

Overall 8.84 0.2101 0.99923 0.4584
CE

Training 5.22 0.1209 0.99958 0.3477

11 Testing 5.91 0.1400 0.99919 0.3742
AC

Overall 5.33 0.1237 0.99954 0.3518

Training 4.75 0.1044 0.99962 0.3231
12 Testing 5.77 0.4001 0.99837 0.6325
Overall 4.90 0.1487 0.99945 0.3856
Training 4.27 0.1018 0.99962 0.3190
13 Testing 5.92 0.4281 0.99894 0.6543
Overall 4.52 0.1506 0.99945 0.3881
Training 4.13 0.0989 0.99965 0.3144
14 Testing 4.74 0.1850 0.99911 0.4302
Overall 4.22 0.1118 0.99959 0.3343
Training 4.46 0.0739 0.99971 0.2718
15 Testing 6.12 0.9032 0.99803 0.9504
Overall 4.71 0.1982 0.99929 0.4452

53
 ACCEPTED MANUSCRIPT

Table 7

Number of hidden Sensitivity accuracy analyses

Database
neurons
AARD% MSE R2 RMSE

PT
Training 140.33 58.132 0.7713 7.624
2 Testing 124.66 46.062 0.7466 6.787
Overall 137.98 56.323 0.7651 7.505

RI
Training 102.80 27.990 0.8849 5.291

SC
4 Testing 112.65 31.793 0.9123 5.639
Overall 104.28 28.560 0.8886 5.344
Training 68.57 20.001 0.9206 4.472

NU
6 Testing 64.37 19.009 0.9390 4.360
Overall 67.94
MA 19.853 0.9238 4.456
Training 90.17 15.808 0.9409 3.976
8 Testing 77.84 14.494 0.9500 3.807
Overall 88.32 15.612 0.9409 3.951
Training 84.86 15.280 0.9435 3.909
D

9 Testing 91.69 15.837 0.9463 3.980

TE

Overall 85.89 15.363 0.9420 3.920

Training 86.56 12.561 0.9546 3.544
P

10 Testing 90.20 13.369 0.9315 3.656

Overall 87.10 12.682 0.9521 3.561
CE

Training 82.65 12.615 0.9533 3.552

11 Testing 78.22 13.767 0.9421 3.710
AC

Overall 81.98 12.787 0.9517 3.576

Training 84.85 12.043 0.9544 3.470
12 Testing 87.58 13.740 0.9488 3.707
Overall 85.26 12.297 0.9535 3.507
Training 63.39 7.256 0.9739 2.694
13 Testing 80.29 7.986 0.9612 2.826
Overall 65.92 7.366 0.9724 2.714
Training 70.20 10.119 0.9569 3.181
14 Testing 60.34 20.727 0.9583 4.553
Overall 68.72 11.709 0.9562 3.422
Training 57.91 5.599 0.9759 2.366
15 Testing 50.58 14.425 0.9753 3.798
Overall 56.81 6.922 0.9745 2.631

54
 ACCEPTED MANUSCRIPT

Table 8

Adjusted parameters Sensitivity accuracy analyses

Database
2  AARD% MSE R2 RMSE

PT
Training 14.54 0.7776 0.99711 0.8818
22.943 4.54×109 Testing 22.86 1.0734 0.99676 1.0361
Overall 15.79 0.8219 0.99697 0.9066

RI
Training 8.30 0.2198 0.99925 0.4688

SC
11.845 2.28×1010 Testing 17.00 2.3862 0.98455 1.5447
Overall 9.60 0.5444 0.99800 0.7379
Training 16.03 1.0528 0.99613 1.0261

NU
4
2.026 2.11×10 Testing 20.57 1.4596 0.99452 1.2081
Overall
MA 16.71 1.1138 0.99589 1.0554
Training 12.00 0.5227 0.99801 0.7230
9
10.486 1.72×10 Testing 19.32 1.1138 0.99656 1.0554
Overall 13.10 0.6113 0.99774 0.7819
Training 13.39 0.7163 0.99749 0.8463
D

9.747 4.42×108 Testing 17.58 0.8040 0.99554 0.8967

TE

Overall 14.02 0.7294 0.99731 0.8541

Training 13.30 0.6745 0.99748 0.8213
P

9.079 3.53×108 Testing 14.76 0.8614 0.99698 0.9281

Overall 13.52 0.7025 0.99740 0.8382
CE

Training 15.33 0.8606 0.99679 0.9277

6
6.442 5.56×10 Testing 17.91 1.4717 0.99489 1.2131
AC

Overall 15.72 0.9522 0.99648 0.9758

Training 14.62 0.8625 0.99685 0.9287
5
3.072 2.58×10 Testing 19.52 1.7139 0.99367 1.3092
Overall 15.35 0.9901 0.99635 0.9950
Training 15.52 0.7259 0.99744 0.8520
22.898 2.32×109 Testing 18.73 2.1817 0.98951 1.4770
Overall 16.00 0.9441 0.99652 0.9716
Training 6.33 0.0791 0.99971 0.2812
7.183 2.24×1010 Testing 10.27 2.5489 0.99073 1.5965
Overall 6.92 0.4492 0.99834 0.6702
Training 15.68 0.9051 0.99644 0.9514
9 Testing 19.34 1.1502 0.99689 1.0725
31.500 2.7×10
Overall 16.23 0.9418 0.99652 0.9705

55
 ACCEPTED MANUSCRIPT

PT
RI
SC
NU
Table 9

AI approach Dataset MAAARD% MSE R2 RMSE

Training stage 4.11 0.1025 0.99962 0.3202

MLPNN Testing stage 4.22 0.3804 0.99867 0.6167
Overall data 4.13 0.1442 0.99947 0.3797
Training stage 4.13 0.0989 0.99965 0.3144
D

CFNN Testing stage 4.74 0.1850 0.99911 0.4302

TE

Overall data 4.22 0.1118 0.99959 0.3343

Training stage 6.33 0.0791 0.99971 0.2812
LS-SVM Testing stage 10.27 2.5489 0.99073 1.5965
P

Overall data 6.92 0.4492 0.99834 0.6702

Training stage
CE

57.91 5.599 0.9759 2.366

RBFNN Testing stage 50.58 14.425 0.9753 3.798
Overall data 56.81 6.922 0.9745 2.631
AC

56
 ACCEPTED MANUSCRIPT

PT
Table 10

RI
Frankel and Nguyen et Maiga Thomas and Rea et Chandrasekar Heyhat
Einstein Brinkman Batchelor

SC
Experimental data Acrivos al. et al. Muthukumar al. et al. et al.

[12] [13] [14] [15] [16] [17] [18] [19] [20] [21]

NU
Ethylene glycol [22] 24.20 23.98 94.03 27.84 23.90 11.56 23.96 54.46 27.80 34.40

MA
Ethylene glycol [23] 39.37 39.15 82.83 42.93 39.06 26.59 39.13 38.38 42.90 24.95

Water [24] 33.80 33.17 180.63 39.36 32.97 11.92 33.12 383.73 39.31 168.62

ED
Water [25] 44.63 44.53 165.32 47.17 44.48 37.11 44.52 11.04 47.14 16.15

PT
Water [26] 2.97 2.97 173.55 3.23 2.97 2.45 2.97 1.26 3.23 1.32

Water [27] 7.24 7.18 337.49 8.73 7.16 5.80 7.18 34.77 8.70 29.54
CE
EG/Water (20/80) [28] 10.18 10.16 286.76 11.61 10.15 7.08 10.16 6.66 11.60 5.35
AC

EG/Water (40/60) [28] 27.91 27.91 252.39 28.17 27.91 27.68 27.91 31.24 28.16 30.32

EG/Water(45/55) [29] 35.87 35.84 147.00 36.90 35.83 33.14 35.84 24.17 36.89 25.70

EG/water(60/40) [28] 38.94 38.93 120.06 39.72 38.92 37.10 38.93 31.12 39.72 32.01

Propylene Glycol [30] 15.90 15.74 273.78 19.51 15.68 7.06 15.73 37.10 19.48 29.47

Overall AARD% 24.20 23.02 180.25 25.14 22.97 16.76 23.01 68.40 25.12 38.80

57
 ACCEPTED MANUSCRIPT

PT
RI
Table 11

SC
Experimental data MLPNN CFNN LS-SVM Correlation* RBFNN

NU
Ethylene glycol [22] 1.94 1.67 2.28 11.56 7.56

Ethylene glycol [23] 3.61 MA 3.53 2.72 24.95 12.34

Water [24] 2.82 4.58 4.59 11.92 170.58

Water [25] 7.65 5.93 15.68 11.04 100.90

D

Water [26] 2.59 1.19 8.94 1.26 28.11

TE

Water [27] 29.02 11.26 35.64 5.8 53.48

EG/Water (20/80) [28] 5.96 6.30 11.71 5.35 84.15

P

EG/Water (40/60) [28] 4.53 6.12 8.16 27.68 62.74

CE

EG/Water (45/55) [29] 5.42 5.43 7.10 24.17 35.03

EG/water (60/40) [28] 4.75 6.63 4.29 31.12 38.29

AC

Propylene Glycol [30] 1.56 0.90 0.33 7.06 16.70

Overall AARD% 4.13 4.24 6.92 14.50 56.81

*
The best obtained result among all of the considered empirical correlations

58
 ACCEPTED MANUSCRIPT

Graphical Abstract

PT
RI
SC
NU
MA
D
TE
P
CE
AC

59
 ACCEPTED MANUSCRIPT

Highlights

 Performance of various models for estimation of nanofluids viscosity were

compared

PT
 Various types of ANN, LS-SVM, and 10 empirical correlations were considered

RI
 MLPNN found as the most accurate model for the considered task

SC
MLPNN estimated 674 data with MSE=0.1422, RMSE=0.3797, AARD=4.13%

and R2=0.99947

NU
MA
D
P TE
CE
AC

60