Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

International Communications in Heat and Mass Transfer 119 (2020) 104930

Contents lists available at ScienceDirect

International Communications in Heat and Mass Transfer


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

A neural network-based predictive model for the thermal conductivity of T


hybrid nanofluids
Humphrey Aduna, Ifeoluwa Wole-Oshoa, Eric C. Okonkwob, , Olusola Bamisilec,

Mustafa Dagbasia, Serkan Abbasoglua


a
Department of Energy Systems Engineering, Cyprus International University, North Cyprus, via Mersin-10, Turkey
b
Division of Sustainable Development, College of Science and Engineering, Hamad Bin Khalifa University, Qatar Foundation, Education City, Doha, Qatar
c
School of Mechanical and Electrical Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan 611731, China

ARTICLE INFO ABSTRACT

Keywords: Nanofluids are known to have immense potential for heat transfer applications because of their unique ther­
Nanofluids mophysical properties when compared to the conventional heat transfer fluid. Predicting the thermophysical
Thermal conductivity features like thermal conductivity has posed a challenge to their application. This article addresses some of the
Artificial neural networks challenges posed in their prediction by using data sets from several experimental research on various hybrid
Prediction
nanofluids to train an intelligent neural network. The thermal conductivity of hybrid nanofluids is predicted
Nanoparticle
using seven different input variables namely, volume concentration, temperature, the acentric factor of the base
fluid, nanoparticle bulk density, mixture ratio of particles, the thermal conductivity, and size of nanoparticles.
715 experimental data points from studies using different hybrid nanoparticles are used in developing a multi-
layer perceptron artificial neural network (ANN) and support vector regression (SVR) models. The performance
validation of the models is computed using the mean square error (MSE) and the coefficient of determination
(R2). The performance result showed an R2 value of 0.99997 and 0.99788 in the validation phase of the ANN and
SVR model, respectively. This indicates that the models are capable of accurately predicting the thermal con­
ductivity of hybrid nanofluids over a wide range of hybrid nanoparticle combinations. Finally, a universal
formula using MLP-ANN for predicting the thermal conductivity of hybrid nanofluids is presented.

1. Introduction conventional nanofluids like Al2O3 water-based nanofluid and show a


similar trend line when compared to experimental results; significant
Nanofluids have the potential to revolutionize the fields of en­ deviations occur when Maxwell model is compared to experimental
gineering and material science. The dispersions of nanoparticles into thermal conductivity values for hybrid nanofluids. There have been
fluids have shown to possess distinctive properties which can improve other models to predict this thermophysical property. Some models like
the efficiencies of current engineering systems. However, predicting the Hamilton-Crosse Model [2], Wasp Model [3], and Yu and Choi [4]
thermodynamic properties of nanofluids has been a bit of a quagmire models modified the Maxwell model to improve its predictive accuracy.
and has limited their applications. However, these models still deviate significantly when predicting the
Since the very end of the 19th century, there have been formulas to thermal conductivity of conventional and hybrid nanofluid.
predict the thermal conductivity of spherical dispersions in a fluid. These deviations are likely because most of the equations proposed
Maxwell predicted that the thermal conductivity of a fluid that contains do not account for all the variables that affect the thermal conductivity
solid conducting particles will depend on the thermal conductivity of of nanofluids. These variables include base fluid thermal conductivity,
the base fluid, the thermal conductivity of the solid conducting dis­ nanoparticle thermal conductivity, the shape of the constituent nano­
persion, and the volume concentration of the dispersion in the fluid [1]. particle, particle aggregation over time, volume concentration of the
While the values predicted by Maxwell are relatively accurate for nanoparticle, interfacial layer within the nanofluid, particle mixture

Abbreviations: ANFIS, Adaptive neuro-fuzzy inference system; ANN, Artificial neural networks; BR, Bayesian regularization; LM, Levenberg-Marquard back pro­
pagation; LSSVM, Least square support vector machine; MLP, Multilayer perceptron neural network; RBF, Radial basis function neural networks; RMS, Root mean
square; RP, Resilient propagation; R2, Coefficient of determination; SCG, Scaled conjugate gradient; SSE, Sum of squared errors; SVR, Support vector regression

Corresponding author.
E-mail address: eokonkwo@hbku.edu.qa (E.C. Okonkwo).

https://doi.org/10.1016/j.icheatmasstransfer.2020.104930

Available online 13 October 2020


0735-1933/ © 2020 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
H. Adun, et al. International Communications in Heat and Mass Transfer 119 (2020) 104930

ratio for hybrid nanofluids, the temperature of the nanofluids, etc. Each While there exists a comprehensive ANN model for predicting the
of these variables has a unique relationship with the thermal con­ thermal conductivity of nanofluids [24], the thermal conductivity of
ductivity of the nanofluid and with each other. For instance, the higher metals, and metal oxides based nanofluids [21,25] and thermal con­
the thermal conductivity of the base fluid and constituent nanoparticle ductivity of ethylene glycol-based hybrid nanofluids [26], no study is
the higher the overall thermal conductivity of the fluid. Also, the yet to propose a comprehensive model for predicting the thermal con­
smaller the particle sizes of nanoparticles in the fluid the higher the ductivity of hybrid nanofluids using various base fluids. The study
thermal conductivity of the fluid, and the higher the rate of particle presents a novel comprehensive artificial neural network-based model
aggregation over time. It can also be observed that an increase in vo­ for predicting the thermal conductivity of different hybrid nanofluids.
lume concentration and temperature of nanofluids tends to improve the The network uses seven different input variables namely, volume con­
overall thermal conductivity of the fluid. While many of these re­ centration, temperature, the acentric factor of the base fluid, nano­
lationships can be obtained from experiments on different nanofluids particle bulk density, mixture ratio of particles, the thermal con­
the exact extents to these relationships have not been fully explored. ductivity, and size of nanoparticles. 715 experimental data points from
In recent years, many researchers have used regression correlation various studies using different hybrid nanoparticles are used in devel­
equations to reflect the relationship between the thermal conductivity oping a Multi-Layer Perceptron artificial neural network (MLP-ANN)
of nanofluids and these variables. Regression correlation equations and Support Vector Regression (SVR) models. The performance vali­
have been proposed for several nanofluids; these nanofluids include dation of the models is carried out using the mean square error (MSE)
Al2O3 water nanofluid, TiO2 water nanofluid, Al2O3 ethylene glycol and the coefficient of determination (R2). Finally, a universal formula
(EG) nanofluid, Al2O3 water(40%)–EG(60%) nanofluid [5], Multi­ using MLP-ANN for predicting the thermal conductivity of hybrid na­
walled carbon nanotube (MWCNT) water nanofluid, graphene oxide- nofluids is presented.
Al2O3 water nanofluid [6], SiC-CuO/C EG nanofluid [7], CNT-Fe3O4
water nanofluid [8], Al2O3- MWCNTs EG nanofluid [9], Al2O3-Fe water 2. Modeling methods
nanofluid [10], Al2O3 10w40-engine oil nanofluid [11], Al2O3-TiO2
water nanofluid [12], TiO2-ZnO EG nanofluid [13], MgO-TiO2 water 2.1. Support vector regression
nanofluid [14] and many others [15,16]. These regression correlation
equations mainly reflect the relationship between the thermal con­ Support vector regression (SVR) and support vector classification
ductivity of nanofluid, volume concentration of nanoparticles, and (SVC) are both subsets of support vector machines [27]. SVR was first
temperature. While these regression equations show significantly less proposed by Vapnik, and as implied in the name, SVR is used for re­
deviation compared to classical models they are significantly limited in gression problems and has been applied to many engineering tasks. The
range. The regression equations are accurate within the range in which concept of the SVR model is to map the non-linear relationship between
the thermal conductivity measurements were taken. input and output variables. Hence, the goal of the SVR model is to
More recently many researchers have turned to artificial in­ predict the function from an observed function with minimal error
telligence (AI) and machine learning models to provide a more accurate [27].
predictive model for the thermal conductivity of nanofluids; these In theory, the SVR model fits a regression model y = f(x) to the
models include Multilayer perceptron neural network (MLP), Radial ε − SVR model, to make an accurate prediction of target {yi} which has
basis function neural networks (RBF) [17], Least square support vector corresponding input variables of {xi}. With the training data of
machine (LSSVM), Adaptive-Network Based Fuzzy Inference Systems {(xi, yi, i = 1, 2, 3, …)}, where x ∈ RN and y ∈ RN are the vector of scalar
(ANFIS) [17] genetic algorithm-polynomial neural network (GA-ANN) input and corresponding scalar output [28]. For mapping, the non-
[18], FCM-based neuro-fuzzy inference system (FCM-ANFIS), Con­ linear relationship between input variables to the corresponding output
jugate hybrid-PSO ANFIS (CHPSO ANFIS), ANFIS-PSO [19] and ANFIS- set, a model is constructed in high dimensional feature space, by using a
GA [19]. non-linear mapping function of (hi(x)) as shown in Eq. (1) [29].
ANFIS, MLP-ANN, RBF-ANN, and LSSVM techniques were used to
n
predict the thermal conductivity of TiO2 nanofluid. The input para­ f (x , w ) = wi hi (x ) + b
meters were taken from several experimental measurements obtained i (1)
from different published articles [20]. These input parameters include
temperature, particle volume concentration, and average particle size. where w and b are the weight and bias respectively. The coefficients are
The temperature range considered was between 5 °C to 90 °C, particle determined through the minimization of the risk function. This is given
volume concentration range was between 0 and 4% and the range for as [29];
TiO2 particle size was between 0 and 220 nm. Of all the considered l
1 1
techniques the LSSVM showed the highest ability to predict the thermal R (C ) = C Lc yi, f (x i , w ) + w 2
l 2 (2)
conductivity of nanofluids containing TiO2 nanoparticles. Adaptive i=1

neuro-fuzzy inference system (ANFIS) and least square support vector where R(C) is the risk function, which is dependent on the empirical
machine (LSSVM) were applied in predicting the thermal conductivity 1 l
form (C l i = 1 Lc (yi, f (x i , w )) ) and the smoothness of the
of conventional nanofluids using temperature, the volume fraction of
the particle, the thermal conductivity of the base fluid, the thermal ( 1
)
function 2 w 2 . Factor C describes the trade-off between the
conductivity of the nanoparticles and nanoparticle size as the input smoothness of the model and the empirical form.
variables. In the study 1109 data points were collected from 29 con­ Also, in the empirical form, the ε parameter accounts for the errors.
ventional nanofluids [21]. In measuring the “upper” and “lower” values of the ε tube, the slack
An MLP-ANN was applied in predicting the thermal conductivity of variables of ξiand ξi∗ are introduced. The slack variables convert the
Cu-TiO2 water/EG hybrid nanofluid. The input parameters considered SVR model to an optimization problem whose objective function is to
were particle volume concentration and temperature. Results show that be minimized. The standard form is shown as [28];
the best performing MLP-ANN configuration had an R2 value of 0.999 n
1
and a mean square error (MSE) of 0.00002 [22]. An MLP-ANN was also minization w 2 +C (i+ i )
used to predict the thermal conductivity of SWCNT–Al2O3 EG nano­ 2 i=1
fluid, 40 data points were applied to the ANN model and the results subject to yi (wh (x i ) + b) i + i
observed that when predicted values are compared to experimentally where (x i ) + b yi + ; i, 0
i i i (3)
obtained values a maximum deviation of 1.94% was observed [23].

2
H. Adun, et al. International Communications in Heat and Mass Transfer 119 (2020) 104930

In overcoming the complexity in computation and high dimensional major issue in ANN modeling is over-fitting [42]. This is the inability of
featured space, kernel functions are defined in SVR models. Kernel a model to predict new datasets, though it fits known datasets in the
functions are either symmetric, positive, or semi-definite [30]. Different training process. Several methods have been suggested to avoid this
kernel functions models have been suggested in literature, although the problem, such as model selection, data augmentation, use of regular­
most utilized are the linear, polynomial, sigmoid, and radial basis ization techniques, and early stopping method [43].
functions [31]. The functions of the linear kernel, polynomial kernel, In this study, the early stopping technique is used to improve the
and radial basis function kernel are(x × y),(⌈(x × xi) + 1⌉d), exp generalization of the model and to eliminate the problem of overfitting.
{−y|x − xi|2} respectively [31]. This technique involves splitting the dataset into subsets of training,
In this study, the cross-validation method is used, and a polynomial testing, and validation. The test set is used in the verification of the
kernel function is adopted to define the non-linear mapping from the generalization and accuracy of the trained model [44]. This is vital
input space to some high dimensional feature space. because at the later stage of the prediction, accuracy begins to worsen
and at this point, the model is stopped and then retrained, as this is
2.2. Artificial neural networks likely to cause overfitting.
The training algorithms used in this study are the scaled conjugate
Traditional modeling is limited in finding relationships between gradient (SCG), Levenberg-Marquardt (LM), Bayesian regularization
input and output data, as there are scenarios where the relationships (BR), and Resilient backpropagation (RP). The retrieved dataset is split
are either complex or non-linear. Artificial neural networks are known into training and validation dataset. The validation dataset was not
to be an efficient algorithm for mapping input-output data. The neural used in the training and testing process.
network reproduces the functionality of the human brain, in other to The performance of the model is tested with two different ANN
learn the relationship that exists among the corresponding variables architecture, these are the ANN models with one layer and with two
[32,33]. The application of artificial neural networks (ANN) spans di­ hidden layers. In both structures, the number of neurons is gradually
verse fields [34–36]. Some of the applications of ANN include pattern increased to find the optimum.
recognition, classification, estimation, and so on [33]. The original idea Table 1 shows the performance parameters used for learning in the
of the ANN is to compute and analyze problems in the same pattern as a ANN model. In this study, the logistic sigmoid is used as the neuron
working human brain. The architecture of ANN is based on the inter­ activation function as shown in Eq. (5) and the hidden and output layer
connection of units termed “artificial neurons” (similar in pattern to the is activated by the Purelin function as shown in Eq. (6). The Purelin
axons and neurons in the brain). The perceptrons, which consist of the function is chosen as the activation function for the output neuron
inputs, outputs, and hidden layers are the interconnected structures because it is ideal for continuous-valued targets [45]. The testing da­
that provide the basis for the algorithm model [37]. The multilayer- taset is used in examining the generalization of the model. Fig. 2 out­
layer perceptron (MLP) network is one of the most used ANN archi­ lines the steps taken to obtain the optimum architecture.
tecture [38]. It is designed to function similarly to the learning process
of the human brain. The different learning algorithms of ANN models (
f (x ) = 1 1 + e cx
) (5)
include Levenberg-Marquard backpropagation (LM), scaled conjugate
gradient (SCG), Bayesian regularization (BR), and resilient propagation
f (x ) = x (6)
(RP). Primarily, the ANN architecture consists of input layers, hidden
layers, and output layers. The neurons in one layer do not have a re­ where mu is the momentum update, mu-dec is the mu decrease, mu-
lationship with each other [39]. Signals from one layer are transferred inc is the mu increase and max-fail is the maximum validation fails.
to the next layer through a non-linear transfer function. A challenge in
utilizing the MLP algorithm is getting an optimal model. This can
however be resolved by using different choices of the MLP algorithm to 3. Computational methodology
arrive at an efficient model [33]. In the MLP algorithm, the weighted
sum of the input is linked to the hidden neurons which is then trans­ The MATLAB 2018a environment is used in carrying out the com­
formed by the activation function. The input to the output neuron is the putation in this study. To model the relative thermal conductivity of
output of the hidden neurons, where further transformation takes place. hybrid nanofluids the following variables were considered, tempera­
Eq. (4) gives the output layer of the MLP neural network [32]; ture, volume fraction, mixture ratio, the acentric factor of the base
fluid, nanoparticle's thermal conductivity, nanoparticle's density, and
q
yij = f j wji Xi + bij nanoparticle's size.
i=1 (4) For faster training and to avoid getting stuck at a local optimum,
preprocessing of the dataset is important. Before training and testing,
where yij represents the i neurons output of the j layer, the bij is the bias all data are normalized between [0–1] and are de-normalized after­
of the i neuron in the j layer. The wji is the linked weight of each ward. Eqs. (7) and (8) determine the normalization and denormaliza­
variable in the i neuron in the j layer, which is chosen at random at the tion functions, respectively.
start of the training process. The Xi is the input vector and the fj is the
non-linear activation function which may be binary sigmoid, bipolar ((X actual Current_min)(Newmax Newmin ))
Xn = + Newmin
sigmoid, Gaussian, identity function, and linear function [40]. (Current max Current min ) (7)
The schematic representation of the MLP neural network is shown
in Fig. 1. The most utilized transfer function for the output layer is the ((Xn Newmin )(Current max Current min ))
purelin, while the tangent sigmoid (tansig) and logarithmic sigmoid Xactual = + Current min
(Newmax Newmin )
(logsig) hyperbolic functions are the most used for the hidden layers
(8)
[41]. In the training process, to minimize errors, there is iterative
random initialization of the weight values and bias values. The back­ where Xn represent the normalized dataset and Xactual is the denorma­
propagation method is usually employed to minimize the errors at lized dataset.
every iteration by back-propagating the errors to reduce the values of The performance criteria used in assessing the accuracy of the dif­
the weight and bias errors [41]. ferent model architecture are the mean square error (MSE) and the
One of the most important considerations in machine learning re­ coefficient of determination (R2) shown in Eqs. (9) and (10), respec­
gression modeling is how a model generalizes to unseen data, hence a tively [46].

3
H. Adun, et al. International Communications in Heat and Mass Transfer 119 (2020) 104930

Fig. 1. Schematic diagram of the MLP neural network.

Table 1 1
N

Optimized parameter selection for the development of the MLP model. MSE = (yi yi ) 2
N I=1 (9)
Parameters Values
1 N
(yi yi )2
Maximum iteration 1000 R2 = 1 N I=1
Minimum gradient 1e-07 1 N
N I=1
(yi yi ) 2 (10)
mu, mu-inc, mu-dec, max-fail 0.001, 10, 0.1, 6
Training dataset 80%
whereyi, yi , yi , and n are experimental data, predicted data, the
Validation dataset 20%
Learning rate 0.1 average value of experimental data, and the number of data points,
Number of hidden layers 1 layer, 2 layers respectively.
Node in hidden layers 1 Hidden layer-10, 15, 20, 25
2 Hidden layer −10/15, 15/15, 20/15, 25/15
Learning algorithm LM, SCG, RP, GDX 3.1. Dataset
Threshold function Logistic sigmoid – neuron activation function
Purelin – activation function for the output The studies that have investigated the relative thermal conductivity
neuron
of hybrid nanofluids were reviewed and 715 experimental data were
Performance metric R2, MSE
obtained from 20 different literatures [6,23,47–64] and used for de­
signing and validating the performance of the predictive model. The

4
H. Adun, et al. International Communications in Heat and Mass Transfer 119 (2020) 104930

Fig. 2. Algorithm to obtain MLP-ANN optimized model.

5
H. Adun, et al. International Communications in Heat and Mass Transfer 119 (2020) 104930

Table 2 Table 3
Data set used for model development. Data set used for model development.
Parameters Range No Nanoparticle Base fluid No of data References
points
Temperature (°C) 20–70
Volume concentration 0–3.5 1 Al2O3/MWCNT Thermal Oil 36 [47]
Particle size (nm) 1.5–70 nm 2 Al2O3/Cu Ethylene Glycol 24 [64]
Mixture ratio 0.15–0.85 3 Al2O3-CNT Water 28 [51]
Acentric factor of base fluid 0.343–0.659 4 Al2O3-SWCNT Ethylene Glycol 40 [23]
Thermal conductivity (W/m.k) 0.161299125–1.440480492 5 Al2O3-Graphene Water 30 [6]
Nanoparticle density (kg/m3) 1000–10,500 oxide
Nanoparticle thermal conductivity (W/m.k) 1.36–3007.4 6 Al2O3/SiO2 Water 20 [48]
7 Al2O3-Fe Water 60 [57]
8 TiO2/ZnO Ethylene Glycol 54 [49]
9 TiO2/SiO2 Water/ Ethylene Glycol (60:40) 21 [63]
results of these studies indicate the factors that primarily affect the
10 TiO2/SiO2 Water/ Ethylene Glycol (60:40) 30 [53]
thermal conductivity of the nanofluids. These factors are temperature, 11 TiO2/CuO Ethylene Glycol 32 [52]
particle size, surfactants, volume concentration, nanoparticle's shape, 12 TiO2-MWCNT Ethylene Glycol /water 36 [55]
size, interfacial layers, mixture ratio, inter-particle spacing, and nano­ 13 TiO2-MWCNT Ethylene Glycol /water 49 [54]
particle's thermal conductivity. Also, the thermal conductivity of hybrid 14 ZnO-Ag Water 36 [56]
15 ZnO-MWCNT Ethylene Glycol /water 35 [59]
nanofluid is influenced by the nanoparticle mixture ratio and the type
16 ZnO-SWCNT Ethylene Glycol /water 42 [58]
of base fluid. To make a comprehensive model covering different hybrid 17 MgO/FMWCNT Ethylene Glycol 24 [62]
nanofluids, the most significant factors that influence thermal con­ 18 Fe3O4/FMWCNT Ethylene Glycol 48 [61]
ductivity must be recognized. In the present study, the temperature, the 19 SiO2/MWCNT Ethylene Glycol 35 [60]
20 SiO2/MWCNT Ethylene Glycol 35 [50]
volume concentration, the particle size, the mixture ratio, the thermal
conductivity of nanoparticle, the nanoparticle density and the acentric
factor of base fluid were selected as input parameters. To improve the
accuracy of the computing model the acentric factor of the base fluid
was used to identify the different base fluids within the study. For the individually. Also, Table 3 shows the number of data points selected
critical properties of mixtures, the acentric factor is determined using from each study. The selected data points are within the temperature
Kay's equation as used in the study conducted by Khalif and Vaferi [24]. and volume concentration ranges considered in this study. The re­
Table 2 shows the general specification of the data set used in the ANN trieved dataset is split into two, 80% for training, and 20% for vali­
and support vector model development. dation. The 20% validation dataset, representing 143 data points ran­
Fig. 3 shows the retrieved hybrid nanofluid from reviewed studies domly selected from the literature [59,61,63,64] [55], was not used in
used to predict the thermal conductivity of the different nanofluids the training process but was used as an external validator.

Fig. 3. Hybrid nanofluid used in this study.

6
H. Adun, et al. International Communications in Heat and Mass Transfer 119 (2020) 104930

4. Results and discussions Table 5


Optimum statistical results for SVR modeling.
In this section, the results from SVR, ANN-BR, ANN-LM, ANN-SCG, Parameters Training Validation dataset
and ANN-RP were discussed and analyzed. Also, this section compares
the results from all the applied prediction techniques and compares
−4
MSE 1.3142e 1.2769e−4
R2 0.99883 0.997883
them with each other to determine the techniques with the best per­
formance. Ultimately the relative error in prediction is determined for
each technique.
validation data, respectively. The result shows that the error was neatly
spread around the zero line, and this points to the excellence of SVR as a
4.1. SVR configurations predictive model for the thermal conductivity of nanofluids.

As earlier stated, the SVR model is constructed using the input


4.2. ANN configuration
combinations of temperature, volume concentration, nanoparticle size,
mixture ratio, the acentric factor of the base fluid, nanoparticle density,
Studies have shown that a small number of neurons in the hidden
and nanoparticle thermal conductivity. High prediction and general­
layers can make the performance of the ANN model unsatisfactory,
izability are only ensured by searching for the optimal parameters of
likewise too large number of neurons can also cause overfitting in the
the SVR model [25]. The kernel function used in this study is the
model [32]. In finding an optimum configuration, there is no general
polynomial function as it gave the best performance. The type of
rule, as the model configuration would depend on the complexity of the
polynomial function used in this study is the cubic polynomial. The
system. The different MLP architecture utilized in this study was ex­
kernel contains a bias term b, accommodated within the kernel func­
amined using trials and errors to obtain the best configuration [66].
tion. The accuracy of the SVR model is defined here in terms of the
generalization ability. In practice, this is inferred as cross-validation
(CV) over the training set of known data. The most common CV is the k- 4.2.1. Evaluation of proposed ANN configuration
fold CV which is usually set between 5 and 10 [65]. The k-fold cross Table 6 shows the statistical results retrieved for the training and
validation used in this study is set to 5. The error between the predicted validation phase of the model. An ANN with one (1) hidden layer and
and real values in a high-dimensional space is the epsilon (ε). The box an ANN with two (2) hidden layers were selected. A range of neurons
constraint is described as a parameter penalizing vectors laying outside was modeled to determine the optimum performing configuration.
the epsilon-margin. Varying the box constraint values, increases, or The optimization of the training algorithm is first determined. For
decreases the support vectors. An optimum box constraint value allows optimal training process results, the training process for each archi­
for more regularization, as more points lie within the margin and this tecture is carried out three times, and the average value is computed.
reduces the risk of overfitting. The points within the margin are support Also, based on literature [32], the trial and error method is used in the
vectors. The optimum parameters for the SVR models are shown in training algorithm, as regards the number of neurons in the hidden
Table 4. layers, in determining the best network architecture. In this study, the
number of hidden layers and the number of neurons in the hidden
4.1.1. Evaluation of the proposed SVR model layers are varied from one to two hidden layers. As observed from
The accuracy of the SVR model was assessed using the MSE and the Table 6 the optimum MLP configuration was achieved with two hidden
coefficient of determination (R2) values between the predicted and the layers having 10 neurons in the first hidden layer and 15 neurons in the
experimental relative thermal conductivity results. From Eqs. (9) and second hidden layer. The Bayesian regularization (BR) method was
(10), the values for both the MSE and R2 are obtained for assessing the used in the training process of the relative thermal conductivity pre­
accuracy of the model in predicting the relative thermal conductivity of diction model. The choice of BR was made by investigating previous
the hybrid nanofluid, the statistical and graphical result is shown in literature [67–71] which shows that the Bayesian method can solve the
Table 5 and Figs. 4 and 5. Table 5 shows the statistical results retrieved challenge posed by overfitting effectively.
for the training and validation phase of the model. The result shows a While BR has some advantages, according to Okut et al. [67], the
high correlation coefficient and low MSE values. This is an indication of performance of BR algorithm may not be significantly better than the
the high computational performance of the developed model for re­ other training algorithms, hence this study also utilized the RP, LM, and
lative thermal conductivity prediction. SCG training algorithms on the determined optimum MLP configuration
From Table 5, the R2 value of 0.99883 in the training dataset shows to validate the best prediction accuracy of the model. The result in
a high correlation coefficient between the experimental and predicted Table 7 shows the statistical performance of different training techni­
values during the training process. However, the R2 value of the vali­ ques in predicting the relative thermal conductivity of hybrid nano­
dation data set was calculated to be 0.99783. While this is 0.001 lower fluids. Based on the recorded results, it is obtained that the BR training
than the value obtained from the training set, it still reflects an ac­ algorithm is the best predictive training model; having both the highest
ceptable performance of the prediction model. The validation process R2 value and least MSE values compared to RP, SCG, and LM. Of the
confirms the accuracy of the prediction model. Fig. 4b gives a graphical considered training algorithms, the SCG performed the worst. This can
illustration of the validation model. be attributed to the model challenges with initial minimization.
Fig. 5a and Fig. 5b show the relative error of the training and Fig. 6a shows the agreement between the predicted and experi­
mental data used in the training process. The R2 value of 0.99989 in the
Table 4 training process shows a high correlation coefficient. Also, as can be
Optimum parameters for SVR modeling. observed from Fig. 6a, all the data points are aligned along the diagonal
line of the graph, which shows a high-performance prediction model.
Model parameters Optimal values
Fig. 6b shows the R2 value of the validation dataset of 0.99993 also
Function Polynomial confirming a high prediction accuracy. Furthermore, Fig. 7 shows the
Kernel scale 1.3372 residual analysis of the training and validation model. The concentra­
Bias 0.8038 tion of the errors around the zero line also proves that the high per­
K folds 5
formance of the models.
Epsilon (ε) 0.0741
Box constraint 0.741 The universal formula for the relative thermal conductivity of hy­
brid nanofluid using the optimum MLP model is derived in this study.

7
H. Adun, et al. International Communications in Heat and Mass Transfer 119 (2020) 104930

Fig. 4. Cross plots of experimental and predicted values of relative thermal conductivity for (a) training (b) validation for the SVR model.

Eq. (11) presents this universal ANN formula. 4.3. Comparative evaluation of the intelligent models to classical predictive
models
15
wi31 (1 1 + exp 10

i=1
w 2 1 1 + exp
j = 1 ji (( 11 1
))
1
k = 1 wkj uk + j + i2 + 13
From the previous sections, MLP and SVR models were developed
for the prediction of the relative thermal conductivity of hybrid nano­
(11)
fluids. Table 8 shows that the MLP has a higher prediction accuracy as
k k
where the bias weights are given as βj and wji is the link weights for compared to the SVR model in this study. The Maxwell model for hy­
the neuron j in layer k. uk is the input variable matrix for the seven brid nanocomposite [72] given in Eqs. (12) and (13) is used in com­
model input parameters. An excel file is attached that contains the puting the relative thermal conductivity of the validation dataset of the
weights and bias of the proposed ANN model. The weight and bias hybrid nanofluid and compared with the machine learning models.
values of the optimum MLP model can be found in the supplementary
excel file.

8
H. Adun, et al. International Communications in Heat and Mass Transfer 119 (2020) 104930

Fig. 5. Relative error distribution over the experimental relative thermal conductivity (a) for training data (b) Validation data.

Table 6
Statistical results of ANN configurations for training and validation.
Hidden layers No of neurons MSE (training) MSE (validation) SSE (validation) R2 (training) R2 (validation)

1 10 5.36879e–5 1.0609e−4 0.0148 0.99914 0.99825


1 15 1.56079e−5 3.249 e−5 0.0047 0.99976 0.99947
1 20 3.31880e−5 4.761 e−5 0.0067 0.99956 0.99928
1 25 1.24703e−5 2.209 e−5 0.0031 0.99986 0.99966
2 5–15 1.1720e−04 1.6641 e−4 0.0235 0.99858 0.99725
2 10–15 8.6675e−06 4.000 e−5 5.585e−4 0.99989 0.99993
2 15–15 9.0519e−06 6.25 e−6 8.5456e−4 0.99987 0.99992
2 20–15 2.1438e−04 1.156 e−5 0.0016 0.99738 0.99983
2 25–15 7.1306e−06 7.29 e−5 9.9542e−4 0.99955 0.99945

The boldened text in Table 6 shows the optimum MLP configuration used in the study from all the possible configurations tested.

9
H. Adun, et al. International Communications in Heat and Mass Transfer 119 (2020) 104930

Table 7
Evaluation of optimum MLP architecture using different training algorithms. Khnf
=
( ( p1 kp1 + p2 kp2 )
tot
+ 2kbf + 2( p1 kp1 + p2 kp2 ) 2 tot kbf )
Kbf ( p1 kp1 + p2 kp2 )
Training MSE (training) MSE R2 (training) R2 (validation) + 2kbf ( p1 k p1 + p2 k p2 ) + tot kbf (12)
tot
algorithms (validation)
tot = p1 + p2 (13)
BR 8.6675e−06 1.0609e−4 0.99989 0.99993
RP 0.0014 1.0240e−3 0.98247 0.98265
where kbf, kp1, kp2, ∅p1, ∅p2, ∅tot is the thermal conductivity of the
SCG 0.0078 5.806e−3 0.90306 0.89226
LM 7.1930e−04 4.4521e−4 0.99086 0.99246 base fluid, the thermal conductivity of particle 1, the thermal con­
ductivity of particle 2, the volume fraction of particle 1, volume fraction
The boldened text in Table 7 indicates the training algorithm with the best of particle 2, and total volume fraction respectively.
result. Table 9 shows model accuracy using the experimental results

Fig. 6. Experimental and predicted values of relative thermal conductivity for the (a) training (b) validation model in MLP.

10
H. Adun, et al. International Communications in Heat and Mass Transfer 119 (2020) 104930

Fig. 7. Experimental and predicted values of relative thermal conductivity for the (a) training (b) validation model in MLP.

Table 8 volume fraction, mixture ratio, and size of the nanoparticles.


Performance of proposed machine learning algorithms. The dataset separated for validation was used in analyzing the effect
Machine learning algorithms ANN SVR
of some of the considered parameters on the relative thermal con­
ductivity of the fluids. TiO2-SiO2 EG/water (60:40) nanofluid as used in
MSE 1.0609e−4 1.2769e−4 Hamid et al. [53] experiment was mainly used to validate and compare
R2 0.99993 0.99783 with the proposed intelligent models.
Fig. 8 shows the effect of volume fraction on the effective thermal
The boldened text in Table 8 shows that the statistical results of the ANN model
performed better than the SVR. conductivity of the hybrid nanofluid. The result shows that by only
considering the effect of volume concentration the intelligent predictive
obtained from Al2O3/Cu water hybrid nanofluid at a 50%–50% mixture models to a degree of accuracy match the experimental data. The MLP
ratio. From Table 9, it can be observed that the Maxwell model has a with BR learning had the best agreement with the experimental data.
higher error and deviation from the actual values, as volume fraction However, SVR also has a near-perfect prediction as well, and only
increases. This is in agreement with comparative studies done between slightly overestimates thermal conductivity values as the volume con­
experimental values of hybrid nanofluid relative thermal conductivity centration of the nanofluid increases. The Maxwell model, on the other
compared with results obtained by Maxwell's equation [10,73]. Also, hand, shows a significant deviation from the experimental data. This
the results presented validate the ANN–BR as the most accurate thermal proves the superiority of machine learning algorithms over Maxwell's
conductivity model presented. model.
Table 10 shows how the mixture ratio of the nanoparticles within
the hybrid nanofluid influences the effective thermal conductivity of
4.4. Relationship between different parameters with the relative thermal the nanofluid. The measured values of the thermal conductivity of
conductivity of hybrid nanofluids hybrid nanofluids are compared to the forecasted values for thermal
conductivity as the mixture ratio of Al2O3 in the Al2O3-ZnO hybrid
In this section, the effect of the different parameters on the thermal nanofluid is varied. Both the MLP model and the SVR model predict the
conductivity of nanofluids is analyzed. These parameters include curve-like relationship between the mixture ratio and thermal

11
H. Adun, et al. International Communications in Heat and Mass Transfer 119 (2020) 104930

Table 9
Model validations of Al2O3/Cu hybrid nanofluid (50%–50% mixture ratio).
Volume fraction Experimental values SVR MLP Maxwell model SVR-deviation MLP deviation Maxwell model deviation

25 °C
0.125 1.0118 1.0230 1.0109 1.0037 1.1035 −0.0861 −0.7972
0.25 1.0417 1.0402 1.0415 1.0075 −0.1368 −0.0148 −3.2798
0.5 1.0767 1.0720 1.0749 1.0150 −0.4324 −0.1692 −5.7280
1 1.1344 1.1279 1.1295 1.0030 −0.5746 −0.4339 −11.5877
1.5 1.1895 1.1794 1.1881 1.0045 −0.8449 −0.1114 −15.5506
2 1.2438 1.2336 1.2413 1.0060 −0.8216 −0.2038 −19.1185

35 °C
0.125 1.0191 1.0262 1.0215 1.0037 0.6958 0.2343 −1.5064
0.25 1.0490 1.0447 1.0519 1.0075 −0.4160 0.2738 −3.9601
0.5 1.0843 1.0789 1.0859 1.0150 −0.4919 0.1491 −6.3858
1 1.1424 1.1397 1.1411 1.0030 −0.2380 −0.1109 −12.2022
1.5 1.1978 1.1959 1.1994 1.0045 −0.1527 0.1332 −16.1368
2 1.2525 1.2548 1.2523 1.0060 0.1890 −0.0168 −19.6796

45 °C
0.125 1.0290 1.0312 1.0301 1.0037 0.2185 0.1102 −2.4515
0.25 1.0590 1.0506 1.0608 1.0075 −0.7942 0.1647 −4.8670
0.5 1.0945 1.0867 1.0959 1.0150 −0.7125 0.1304 −7.2628
1 1.1531 1.1511 1.1525 1.0030 −0.1789 −0.0565 −13.0214
1.5 1.2090 1.2109 1.2108 1.0045 0.1563 0.1427 −16.9183
2 1.2642 1.2733 1.2636 1.0060 0.7192 −0.0531 −20.4277

50 °C
0.125 1.0349 1.0348 1.0337 1.0037 −0.0071 −0.1224 −3.0124
0.25 1.0651 1.0546 1.0646 1.0075 −0.9823 −0.0439 −5.4053
0.5 1.1007 1.0914 1.1006 1.0150 −0.8463 −0.0108 −7.7835
1 1.1596 1.1571 1.1581 1.0030 −0.2183 −0.1334 −13.5079
1.5 1.2158 1.2183 1.2166 1.0045 0.2008 0.0592 −17.3824
2 1.2713 1.2820 1.2694 1.0060 0.8366 −0.1522 −20.8718

conductivity. However, the Maxwell model inaccurately shows a linear Table 10


relationship between the mixture ratio and thermal conductivity. The Effect of mixture ratio on relative thermal conductivity.
MLP is again more accurate than the SVR models in predicting this Mixture ratio Experimental data Machine learning Maxwell model
relationship. algorithms
The result obtained in this study corroborates with those in litera­
ture. The study by Kavitha and Kumar [74] compares the performance R1 R2 MLP SVR

of the SVR and MLP models in the prediction of the thermal properties 0.33 0.67 0.782 0.787621 0.785223 1.02935793
of nanofluids. Their study showed that the MLP model performed better 0.5 0.5 0.775 0.777154 0.77265 1.029408144
than the SVR model, which had a cubic polynomial kernel as the best 0.67 0.33 0.805 0.806927 0.781868 1.029453291
kernel function. However, the R2 value obtained in this study showed a

Fig. 8. Effect of volume concentration on relative thermal conductivity.

12
H. Adun, et al. International Communications in Heat and Mass Transfer 119 (2020) 104930

better prediction performance when compared with that obtained in Acknowledgements


their study. A similar study by Jamei et al. [26] used three models;
genetic algorithm, model tree, and multiple linear regression for pre­ Open Access funding provided by the Qatar National Library.
diction of thermal conductivity of hybrid nanofluid. Their best per­
forming model was the genetic algorithm with an R2 value of 0.9028 References
which is less accurate when compared to the results obtained from this
study. A study by Wang et al. [66] also compared the prediction per­ [1] J.C. Maxwell, A Treatise on Electricity and Magnetism, 2 Clarendon Press, London,
formance of hybrid nanofluids using different machine learning algo­ 1881.
[2] R.L. Hamilton, O.K. Crosser, Thermal conductivity of heterogeneous two-compo­
rithms. Their result showed that the best performance was obtained nent systems, Ind. Eng. Chem. Fundam. 1 (3) (1962) 187–191.
with the mind evolutionary algorithm back propagation neural network [3] L.E. Nielsen, Generalized equation for the elastic moduli of composite materials, J.
(MEA-BPNN), which gave an R2 value of 0.9997. This also shows that Appl. Phys. 41 (1970) 4626.
[4] W. Yu, S.U.S. Choi, The role of interfacial layers in the enhanced thermal con­
the prediction in this study has better accuracy. ductivity of nanofluids: a renovated Hamilton-Crosser model, J. Nanopart. Res. 6
(2004) 355–361.
5. Conclusion [5] M. Hemmat Esfe, W.M. Yan, M. Afrand, M. Sarraf, D. Toghraie, M. Dahari,
Estimation of thermal conductivity of Al2O3/water (40%)-ethylene glycol (60%) by
artificial neural network and correlation using experimental data, Int. Commun.
To aid in enhancing the thermal application of nanofluids in en­ Heat Mass Transf. 74 (2016) (2016) 125–128.
gineering systems, this study seeks to address the challenges faced with [6] R. Taherialekouhi, S. Rasouli, A. Khosravi, An experimental study on stability and
thermal conductivity of water-graphene oxide/aluminum oxide nanoparticles as a
predicting the thermal conductivity of hybrid nanofluids. The study
cooling hybrid nanofluid, Int. J. Heat Mass Transf. 145 (2019) (2019) 118751.
presents an intelligent based model for predicting the thermal con­ [7] S. Akilu, A.T. Baheta, S. Chowdhury, E. Padmanabhan, K.V. Sharma,
ductivity of various hybrid nanofluids. The network uses seven different Thermophysical profile of SiC–CuO/C nanocomposite in base liquid ethylene glycol,
input variables namely, volume concentration, temperature, the acen­ Powder Technol. 354 (2019) (2019) 540–551.
[8] A. Shahsavar, A. Godini, P.T. Sardari, D. Toghraie, H. Salehipour, Impact of variable
tric factor of the base fluid, nanoparticle density, mixture ratio of fluid properties on forced convection of Fe3O4/CNT/water hybrid nanofluid in a
particles, the thermal conductivity, and size of nanoparticles. 715 data double-pipe mini-channel heat exchanger, J. Therm. Anal. Calorim. 137 (3) (2019)
points from several published articles are used to validate and train the 1031–1043.
[9] A.A.A. Arani, F. Pourmoghadam, Experimental investigation of thermal con­
neural network model. Upon the statistical validation of the models ductivity behavior of MWCNTS-Al2O3/ethylene glycol hybrid Nanofluid: providing
using their R2 and MSE values, the following conclusions can be new thermal conductivity correlation, Heat Mass Transf. Stoffuebertragung 55 (8)
reached: (2019) 2329–2339.
[10] E.C. Okonkwo, I. Wole-Osho, D. Kavaz, M. Abid, Comparison of experimental and

• When BR, RP, LM, and SCG training algorithms are used to train the
theoretical methods of obtaining the thermal properties of alumina/iron mono and
hybrid nanofluids, J. Mol. Liq. 292 (2019) (2019) 111377.
MLP, the BR training produced the most accurate hybrid thermal [11] M. Tahmasebi Sulgani, A. Karimipour, Improve the thermal conductivity of 10w40-
engine oil at various temperature by addition of Al 2 O 3 /Fe 2 O 3 nanoparticles, J.
conductivity prediction results when validated.

Mol. Liq. 283 (2019) (2019) 660–666.
Using the SVR model to predict the thermal conductivity of nano­ [12] G.M. Moldoveanu, A.A. Minea, G. Huminic, A. Huminic, Al2O3/TiO2 hybrid na­
fluids shows fairly accurate results. However, the SVR model does nofluids thermal conductivity, J. Therm. Anal. Calorim. 137 (2) (2019) 583–592.
[13] B. Jacob Rubasingh, P. Selvakumar, R.S. Sanjay Raja, Predicting thermal con­
not perform as well as the MLP.

ductivity behaviour of ZnO, TiO2 and ball milled TiO2/ZnO based nanofluids with
The Maxwell model is not an accurate predictor of the thermal ethylene glycol as base fluid, Mater. Res. Express 6 (9) (2019) 095702.
conductivity for hybrid nanofluid. It fails to predict the trend lines [14] S.M. Mousavi, F. Esmaeilzadeh, X.P. Wang, A detailed investigation on the thermo-
especially when the mixture ratio of the particles is considered. physical and rheological behavior of MgO/TiO 2 aqueous dual hybrid nanofluid, J.


Mol. Liq. 282 (2019) 323–339.
Both MLP and SVR can accurately predict the effects of volume [15] I.W. Almanassra, A.D. Manasrah, U.A. Al-Mubaiyedh, T. Al-Ansari, Z.O. Malaibari,
concentration, mixture ratio, and temperature on the thermal con­ M.A. Atieh, An experimental study on stability and thermal conductivity of water/
ductivity of hybrid nanofluids. CNTs nanofluids using different surfactants: a comparison study, J. Mol. Liq. 304
(2020) 111025.
[16] E.C. Okonkwo, M. Abid, E.A. Essien, D. Kavaz, T.A.H. Ratlamwala, Olive leaf-syn­
This research ultimately shows that intelligent systems can help thesized nanofluids for solar parabolic trough collector — thermal performance
scientists and engineers predictive the thermal conductivity of hybrid evaluation, J. Therm. Sci. Eng. Appl. 11 (4) (2019) 1–13 041009.
[17] N. Zhao, Z. Li, Experiment and artificial neural network prediction of thermal
nanofluids to a remarkably high degree of accuracy across different conductivity and viscosity for alumina-water nanofluids, Materials (Basel) 10
base fluid types, nanoparticle types, temperatures, volume concentra­ (2017) 5.
tions, and mixture ratios within hybrid. [18] H. Karimi, F. Yousefi, M.R. Rahimi, Correlation of viscosity in nanofluids using
genetic algorithm-neural network (GA-NN), Heat Mass Transf. Stoffuebertragung 47
(2011) 1417–1425.
Nomenclature [19] I.M. Alarifi, H.M. Nguyen, A.N. Bakhtiyari, A. Asadi, Feasibility of ANFIS-PSO and
ANFIS-GA models in predicting thermophysical properties of Al2O3-MWCNT/oil
hybrid nanofluid, Materials (Basel) 12 (21) (2019) 3628.
b bias
[20] M.H. Ahmadi, A. Baghban, M. Sadeghzadeh, M. Hadipoor, M. Ghazvini, Evolving
i neurons connectionist approaches to compute thermal conductivity of TiO2/water nano­
j layers fluid, Phys. A Stat. Mech. Appl. 540 (2020) (2020) 122489.
k Thermal conductivity (W/m.K) [21] R. Razavi, A. Sabaghmoghadam, A. Bemani, A. Baghban, K. wing Chau, E. Salwana,
Application of ANFIS and LSSVM strategies for estimating thermal conductivity
R Mixture ratio of Al2O3 enhancement of metal and metal oxide based nanofluids, Eng. Appl. Comput. Fluid
w weights Mech. 13 (1) (2019) 560–578.
y output [22] M. Hemmat Esfe, S. Wongwises, A. Naderi, A. Asadi, M.R. Safaei, H. Rostamian,
M. Dahari, A. Karimipour, Thermal conductivity of cu/TiO2-water/EG hybrid na­
nofluid: experimental data and modeling using artificial neural network and cor­
Greek symbols relation, Int. Commun. Heat Mass Transf. 66 (2015) (2015) 100–104.
[23] M.H. Esfe, M. Rejvani, R. Karimpour, A.A. Abbasian Arani, Estimation of thermal
conductivity of ethylene glycol-based nanofluid with hybrid suspensions of
β Ratio of layer thickness to particle radius SWCNT–Al 2 O 3 nanoparticles by correlation and ANN methods using experi­
γ Ratio of layer thermal conductivity to particle thermal con­ mental data, J. Therm. Anal. Calorim. 128 (3) (2017) 1359–1371.
ductivity [24] A. Khalifeh, B. Vaferi, Intelligent assessment of effect of aggregation on thermal
conductivity of nanofluids—comparison by experimental data and empirical cor­
μ dynamic viscosity (Pa s)
relations, Thermochim. Acta 681 (2019) (2019) 178377.
ρ density (kg/m3) [25] I.O. Alade, T.A. Oyehan, I.K. Popoola, S.O. Olatunji, A. Bagudu, Modeling thermal
φ volumetric fraction of nanoparticles (%) conductivity enhancement of metal and metallic oxide nanofluids using support
vector regression, Adv. Powder Technol. 29 (1) (2018) 157–167.

13
H. Adun, et al. International Communications in Heat and Mass Transfer 119 (2020) 104930

[26] M. Jamei, R. Pourrajab, I. Ahmadianfar, A. Noghrehabadi, Accurate prediction of nanoparticles, J. Therm. Anal. Calorim. 124 (1) (2016) 455–460.
thermal conductivity of ethylene glycol-based hybrid nanofluids using artificial [52] S. Akilu, A.T. Baheta, K.V. Sharma, Experimental measurements of thermal con­
intelligence techniques, Int. Commun. Heat Mass Transf. 116 (2020), https://doi. ductivity and viscosity of ethylene glycol-based hybrid nanofluid with TiO2-CuO/C
org/10.1016/j.icheatmasstransfer.2020.104624 In this issue. inclusions, J. Mol. Liq. 246 (2017) 396–405.
[27] Y. Tao, H. Yan, H. Gao, Y. Sun, G. Li, Application of SVR optimized by Modified [53] K.A. Hamid, W.H. Azmi, M.F. Nabil, R. Mamat, K.V. Sharma, Experimental in­
Simulated Annealing (MSA-SVR) air conditioning load prediction model, J. Ind. Inf. vestigation of thermal conductivity and dynamic viscosity on nanoparticle mixture
Integr. 15 (2019) 247–251. ratios of TiO2-SiO2 nanofluids, Int. J. Heat Mass Transf. 116 (2018) 1143–1152.
[28] Y. Xiang, L. Gou, L. He, S. Xia, W. Wang, A SVR–ANN combined model based on [54] A. Akhgar, D. Toghraie, An experimental study on the stability and thermal con­
ensemble EMD for rainfall prediction, Appl. Soft Comput. J. 73 (2018) 874–883. ductivity of water-ethylene glycol/TiO2-MWCNTs hybrid nanofluid: developing a
[29] M.S. Ahmad, S.M. Adnan, S. Zaidi, P. Bhargava, A novel support vector regression new correlation, Powder Technol. 338 (2018) (2018) 806–818.
(SVR) model for the prediction of splice strength of the unconfined beam specimens, [55] A. Akhgar, D. Toghraie, N. Sina, M. Afrand, Developing dissimilar artificial neural
Constr. Build. Mater. 248 (2020) 118475. networks (ANNs) to prediction the thermal conductivity of MWCNT-TiO2/water-
[30] H. Mashhadi Meighani, C. Ghotbi, T. Jafari Behbahani, K. Sharifi, Evaluation of PC- ethylene glycol hybrid nanofluid, Powder Technol. 355 (2019) 602–610.
SAFT model and Support Vector Regression (SVR) approach in prediction of as­ [56] N.N. Esfahani, D. Toghraie, M. Afrand, A new correlation for predicting the thermal
phaltene precipitation using the titration data, Fluid Phase Equilib. 456 (2018) conductivity of ZnO–Ag (50%–50%)/water hybrid nanofluid: an experimental
171–183. study, Powder Technol. 323 (2018) 367–373.
[31] C.H. Wu, J.M. Ho, D.T. Lee, Travel-time prediction with support vector regression, [57] E.C. Okonkwo, I. Wole-osho, D. Kavaz, M. Abid, T. Al-ansari, Thermodynamic
IEEE Trans. Intell. Transp. Syst. 5 (4) (2004) 276–281. evaluation and optimization of a flat plate collector operating with alumina and
[32] E. Heidari, M.A. Sobati, S. Movahedirad, Accurate prediction of nanofluid viscosity iron mono and hybrid nanofluids, Sustain. Energy Technol. Assess. 37 (2020)
using a multilayer perceptron artificial neural network (MLP-ANN), Chemom. 100636.
Intell. Lab. Syst. 155 (2016) 73–85. [58] M. Hemmat Esfe, A.A. Abbasian Arani, M. Firouzi, Empirical study and model de­
[33] M. Roubehie Fissa, Y. Lahiouel, L. Khaouane, S. Hanini, QSPR estimation models of velopment of thermal conductivity improvement and assessment of cost and sen­
normal boiling point and relative liquid density of pure hydrocarbons using MLR sitivity of EG-water based SWCNT-ZnO (30%:70%) hybrid nanofluid, J. Mol. Liq.
and MLP-ANN methods, J. Mol. Graph. Model. 87 (2019) 109–120. 244 (2017) 252–261.
[34] F. Yang, H. Cho, H. Zhang, J. Zhang, Y. Wu, Artificial neural network (ANN) based [59] M. Hemmat Esfe, S. Esfandeh, S. Saedodin, H. Rostamian, Experimental evaluation,
prediction and optimization of an organic Rankine cycle (ORC) for diesel engine sensitivity analyzation and ANN modeling of thermal conductivity of ZnO-
waste heat recovery, Energy Convers. Manag. 164 (February) (2018) 15–26. MWCNT/EG-water hybrid nanofluid for engineering applications, Appl. Therm.
[35] D.A. Fadare, Modelling of solar energy potential in Nigeria using an artificial neural Eng. 125 (2017) 673–685.
network model, Appl. Energy 86 (9) (2009) 1410–1422. [60] M. Hemmat Esfe, S. Esfandeh, M. Rejvani, Modeling of thermal conductivity of
[36] R.C. Deo, M. Şahin, Application of the Artificial Neural Network model for pre­ MWCNT-SiO2 (30:70%)/EG hybrid nanofluid, sensitivity analyzing and cost per­
diction of monthly Standardized Precipitation and Evapotranspiration Index using formance for industrial applications: an experimental based study, J. Therm. Anal.
hydrometeorological parameters and climate indices in eastern Australia, Atmos. Calorim. 131 (2) (2018) 1437–1447.
Res. 161–162 (2015) 65–81. [61] S. Sarbolookzadeh Harandi, A. Karimipour, M. Afrand, M. Akbari, A. D’Orazio, An
[37] K. Abrougui, K. Gabsi, B. Mercatoris, C. Khemis, R. Amami, S. Chehaibi, Prediction experimental study on thermal conductivity of F-MWCNTs-Fe3O4/EG hybrid na­
of organic potato yield using tillage systems and soil properties by artificial neural nofluid: effects of temperature and concentration, Int. Commun. Heat Mass Transf.
network (ANN) and multiple linear regressions (MLR), Soil Tillage Res. 190 (2019) 76 (2016) 171–177.
202–208 (December 2018. [62] M. Vafaei, M. Afrand, N. Sina, R. Kalbasi, F. Sourani, H. Teimouri, Evaluation of
[38] A. Moghadassi, F. Parvizian, S.M. Hosseini, A new approach based on artificial thermal conductivity of MgO-MWCNTs/EG hybrid nanofluids based on experi­
neural networks for prediction of high pressure vapor-liquid equilibrium, Aust. J. mental data by selecting optimal artificial neural networks, Phys. E Low Dimens.
Basic Appl. Sci. 3 (3) (2009) 1851–1862. Syst. Nanostruct. 85 (2017) 90–96.
[39] K.N. Çerçi, M. Daş, Modeling of heat transfer coefficient in solar greenhouse type [63] K.A. Hamid, W.H. Azmi, M.F. Nabil, R. Mamat, Improved thermal conductivity of
drying systems, Sustainability 11 (18) (2019). TiO2-SiO2 hybrid nanofluid in ethylene glycol and water mixture, IOP Conf. Ser.
[40] L.V. Fausett, Fundamentals of Neural Networks, Prentice-Hall, 1994. Mater. Sci. Eng. 257 (1) (2017) 0–7.
[41] P. Amani, K. Vajravelu, Intelligent modeling of rheological and thermophysical [64] A. Parsian, M. Akbari, New experimental correlation for the thermal conductivity of
properties of green covalently functionalized graphene nanofluids containing na­ ethylene glycol containing Al2O3–Cu hybrid nanoparticles, J. Therm. Anal.
noplatelets, Int. J. Heat Mass Transf. 120 (2018) 95–105. Calorim. 131 (2) (2018) 1605–1613.
[42] G. Panchal, A. Ganatra, P. Shah, D. Panchal, Determination of over-learning and [65] K. Cheng, Z. Lu, Y. Zhou, Y. Shi, Y. Wei, Global sensitivity analysis using support
over-fitting problem in back propagation neural network, Int. J. Soft Comput. 2 (2) vector regression, Appl. Math. Model. 49 (2017) 587–598.
(2011) 40–51. [66] J. Wang, Y. Zhai, P. Yao, M. Ma, H. Wang, Established prediction models of thermal
[43] S. Lawrence, C.L. Giles, Overfitting and neural networks: conjugate gradient and conductivity of hybrid nanofluids based on artificial neural network (ANN) models
backpropagation, Proc. Int. Jt. Conf. Neural Networks 1 (2000) 114–119. in waste heat system, Int. Commun. Heat Mass Transf. 110 (2020) (December 2019,
[44] C.H.I.D. Doan, S. Liong, Generalization for multilayer neural network bayesian 104444).
regularization or early stopping, Network (2004) 119260), 1–8. [67] M. Kayri, Predictive abilities of Bayesian regularization and levenberg-marquardt
[45] S.M. Ghoreishi, E. Heidari, Extraction of Epigallocatechin-3-gallate from green tea algorithms in artificial neural networks: a comparative empirical study on social
via supercritical fluid technology: neural network modeling and response surface data, Math. Comput. Appl. 21 (2) (2016).
optimization, J. Supercrit. Fluids 74 (2013) 128–136. [68] T. Chaipimonplin, The Efficiency of Using Different of Learning Algorithms in
[46] E.C. Okonkwo, H. Adun, A.A. Babatunde, M. Abid, T.A.H. Ratlamwala, Entropy Artificial Neural Network Model for Flood Forecasting At Upper River Ping
generation minimization in a parabolic trough collector operating with SiO2 - water Catchment, November (2016), pp. 4–7.
nanofluids using genetic algorithm and artificial neural network, J. Therm. Sci. Eng. [69] L.M. Saini, Peak load forecasting using Bayesian regularization, resilient and
Appl. 12 (3) (2020) (031007 (11 pages). adaptive backpropagation learning based artificial neural networks, Electr. Power
[47] A. Asadi, M. Asadi, A. Rezaniakolaei, L.A. Rosendahl, M. Afrand, S. Wongwises, Syst. Res. 78 (7) (2008) 1302–1310.
Heat transfer efficiency of Al2O3-MWCNT/thermal oil hybrid nanofluid as a cooling [70] J.L. Ticknor, A Bayesian regularized artificial neural network for stock market
fluid in thermal and energy management applications: an experimental and theo­ forecasting, Expert Syst. Appl. 40 (14) (2013) 5501–5506.
retical investigation, Int. J. Heat Mass Transf. 117 (2018) 474–486. [71] S. Chaki, S. Ghosal, Modelling and Optimisation of Laser Assisted Oxygen (LASOX)
[48] G.M. Moldoveanu, G. Huminic, A.A. Minea, A. Huminic, Experimental study on Cutting: A Soft Computing Based Approach, Springer, 2018.
thermal conductivity of stabilized Al2O3 and SiO2 nanofluids and their hybrid, Int. [72] Ç. Yıldız, M. Arıcı, H. Karabay, Comparison of a theoretical and experimental
J. Heat Mass Transf. 127 (2018) 450–457. thermal conductivity model on the heat transfer performance of Al2O3-SiO2/water
[49] D. Toghraie, V.A. Chaharsoghi, M. Afrand, Measurement of thermal conductivity of hybrid-nanofluid, Int. J. Heat Mass Transf. 140 (2019) 598–605.
ZnO–TiO2/EG hybrid nanofluid: effects of temperature and nanoparticles con­ [73] I. Wole-Osho, E.C. Okonkwo, H. Adun, D. Kavaz, S. Abbasoglu, An intelligent ap­
centration, J. Therm. Anal. Calorim. 125 (1) (2016) 527–535. proach to predicting the effect of nanoparticle mixture ratio, concentration, and
[50] M. Hemmat Esfe, P.M. Behbahani, A.A.A. Arani, M.R. Sarlak, Thermal conductivity temperature on thermal conductivity of hybrid nanofluids, J. Therm. Anal. Calorim.
enhancement of SiO2–MWCNT (85:15%)–EG hybrid nanofluids: ANN designing, (2020), https://doi.org/10.1007/s10973-020-09594-y In press.
experimental investigation, cost performance and sensitivity analysis, J. Therm. [74] R. Kavitha, P.C. Mukesh Kumar, A comparison between MLP and SVR models in
Anal. Calorim. 128 (1) (2017) 249–258. prediction of thermal properties of nano fluids, J. Appl. Fluid Mech. 11
[51] M. Hemmat Esfe, S. Saedodin, W.M. Yan, M. Afrand, N. Sina, Study on thermal (Specialissue) (2018) 7–14.
conductivity of water-based nanofluids with hybrid suspensions of CNTs/Al2O3

14

You might also like