Estimation Of Viscosity In Nanofluids Using Neural Network-Genetic

Algorithm Method

Engku Ahmad Fariz Bin Engku Ahmad Khairul Anuar,

Department of Chemical and Process Engineering,
Faculty of Engineering & Built Environment, Universiti Kebangsaan Malaysia, Bangi, Malaysia

ABSTRACT

An accurate and efficient artificial neural network based on a genetic algorithm (GA) is developed for predicting
nanofluid viscosity. The genetic algorithm (GA) is used to optimize the neural network parameters. The experimental
viscosity in eight nanofluids in the range 238.15–343.15 K with the nanoparticle volume fraction up to 9.4% was used.
The obtained results show that the GA-NN model has a good agreement with the experimental data with absolute
deviation 2.48% and high correlation coefficient (R C 0.98). The Results also reveals that GA-NN model outperforms
to the conventional neural nets in predicting the viscosity of nanofluids with the overall percentage improvement of
39%. Furthermore, the results have also been compared with Einstein, Batchelor and Masoumi et al. models. The
findings demonstrate that this model is an efficient method and have better accuracy.
Keywords: Viscosity, Nanofluids, Genetic Algorithm, Neural Network, Artificial Intelligence.

INTRODUCTION properties of nanofluids, viscosity is very important

since it describes the resistance of a fluid to flow, and
hence impacts many other phenomena; including heat
Rapid advances in nanotechnology led to the invention
and mass transfer within the fluid. For industrial
of nanofluids in 1995 with a pioneering work by Choi.
applications, the viscosity of nanofluids can
These fluids are mixtures of nanoparticles with
significantly influence pumping power, workability,
average size of 10 – 100 nm that dispersed in base
pressure drop, and convective heat transfer coefficient
fluids such as water, ethylene glycol, propylene
(Kole & Dey 2010; Mahbubul et al. 2014). Therefore,
glycol, engine oil as well as conventional cooling
in spite of the complexities in hydrodynamic and
fluids (Choi & Eastman 1995; Kamiński & Ossowski
particle-particle interactions of dispersed
2014). The nanoparticles can be metals (Ni, Cu, Al,
nanoparticles, precise prediction of the nanofluid
Ag, Au, etc.), metal oxide (NiO2, CuO, Al2O3, TiO2 ,
viscosity is of great importance (Mehrabi et al. 2013).
Fe3O4, Fe2 O3 , etc.), nonmetals (Si, Graphene, single
and multiwall carbon nanotube [SWCNT and PROBLEM STATEMENT
MWCNT], etc.), and other particles (CaCo 3, SiC,
SiO2 , AlN, Al2Cu, Ag2 Al, etc.). Due to their unique
Over the past two decades, many investigations have
characteristic, nanofluids find a wide application in
been conducted to evaluate the viscosity of different
enhancing heat transfer, even for microscale heat
types of nanofluids over a wide range of experimental
transfer (Lee & Choi 1996) as they display remarkable
conditions; including temperature, the concentration
properties such as enhanced thermal conductivity,
of nanoparticles, shear rate, nanoparticle size and
long-term stability, minimal clogging in flow
shape, pH, and electrical conductivity. Literature
passages, and homogeneity (Chandrasekar et al.
sources agree that the addition of nanoparticles into
2010). It is expected that the nanofluid will become a
conventional base fluid increases the viscosity of the
new type of heat transfer fluid for thermal engineering.
fluid with a monotonic increase in the viscosity with
Thus, the use of nanofluids, for example in heat
increasing particle concentration (Adio et al. 2016;
exchangers, may result in energy and cost savings and
Duangthongsuk & Wongwises 2009; Nguyen et al.
should facilitate the trend of device miniaturization.
2007; Prasher et al. 2006; Sundar, Hortiguela, et al.
To select a nanofluid for any particular application,
2016; Syam Sundar et al. 2013). Moreover, the
it is essential to determine its thermophysical
viscosity of nanofluids increases with decreasing
properties including viscosity, thermal conductivity,
temperature (Kole & Dey 2010; Pastoriza-Gallego et
density, and specific heat. Among the different
al. 2014; Sundar, Singh, et al. 2016; Turgut et al. network through adequate training from the
2009), although some studies report that the relative experimental data (Jain & Fanelli 2000).
viscosity of nanofluids (viscosity of nanofluid over the These methods provide a parameterized, non-linear
viscosity of its base fluid) is independent of mapping between inputs and outputs and are
temperature (Chen, Ding & Tan 2007; Chen, Ding, extremely useful in recognizing patterns in complex
He, et al. 2007). Similarly, most literature reported an data. The resulting quantitative models are
increase in nanofluid viscosity with decreasing transparent; they can be interrogated to reveal the
particle size (Anoop et al. 2009; Lu & Fan 2008; patterns and the model parameters and can be studied
Pastoriza-Gallego et al. 2011) or a decrease in to clarify the significance of particular variables.
nanofluid viscosity with an increase in particle size Several authors have discussed the history,
(He et al. 2007; Nguyen et al. 2007, 2008) with only a architecture, and operation of ANN.
few exceptions showing particle size-independent Hornik et al. 1989 have proved that a three-layered
viscosity (Prasher et al. 2006). feedforward neural network with a backpropagation
Due to above mention problems, the experimental algorithm can map any non-linear relationship with a
data for the effective viscosity of nanofluids are desired degree of accuracy. In this study, a three-layer
limited to certain nanofluids. Therefore, the ranges of back propagation network (Fig. 1) is presented to
the investigated variables such as the particle volume predict viscosity, where the transfer functions in the
concentration, particle size and temperature are also hidden and output layer are sigmoid and linear,
limited. respectively. The five variables including temperature,
Most of the suggested correlations reported for the particle volume concentration, particle size, fluid base
determination of nanofluids viscosity are based on a viscosity and density ratio of fluid base to particle are
simple Einstein model and experimental determination chosen as input to the BPNN, the nanofluid viscosity
of nanofluid viscosity at different temperatures, is the output of the network.
particle sizes and concentrations is time-consuming
and expensive (Hemmat Esfe et al. 2015; Meybodi et
al. 2015). Therefore, the main objective of this study
is to develop a rigorous estimation method for the
viscosity of nanofluids with ability to predict the
viscosity of different nanofluids accurately and
robustly with respect to temperature and nanoparticle
and base fluid properties.
METHODOLOGY

The soft computing techniques employed in this study

include the artificial neural networks and genetic
algorithms. Implementation algorithm of the network Figure 1 Architecture of three layers back
propagation neural network for nanofluids
training is programmed within the Matlab viscosity prediction
environment using GA and Neural Networks Tool
Boxes. A brief overview of these techniques is
Genetic Algorithm (GA)
presented here.
Artificial Neural Network (ANN)
GA has been proved to be capable of finding global
optima in complex problems by exploring virtually all
Artificial neural networks have been introduced as regions of the state space and exploiting promising
flexible mathematical function, which can imitate areas through mutation, crossover and selection
complex and non-linear relationships through the operations applied to individuals in the populations. It
application of many non-linear processing units called applies selection, crossover and mutation operators to
neurons. The relationship can be learned by a neural construct fitter solutions. A genetic algorithm
processes the populations of chromosomes by The selection operation is to choose the individual
replacing unsuitable candidates according to the who has the large fitness from the population. It has
fitness function. It can solve how to optimize the chance to propagate offspring. A roulette wheel
connectivity weights and thresholds of BPNN. It is selection method is used to choose new individual in
employed to optimize the weights and thresholds of this study
the network in this study. The operation is divided into
five steps as follows: 4. Crossover operation

The crossover operation for GA creates variation in

1. Initialization Of Population the population by producing new offspring that
consists of the parts taking from each parent.
After initializing of weights wij and thresholds bi of
Subsequently some genes of two chromosomes are
the network, chromosome is generated at random
changed to produce new individual. In this study, two
based on these weights and thresholds and makes up
parental chromosomes and bunch’s crossing position
an initial population size as [W, b]mxn where m is
are determined a randomly.
number of weights and thresholds and n is the number
of generation. We start iterative search using the initial
5. Mutation operation
population as a start point. Finally, population size,
selection probability, crossover probability and The mutation operation introduces random changes in
mutation probability are determined by experiments. the structure of the population. It is used to change
some values of chromosomes of weight and threshold
2. Fitness function with little probability. It cannot only avoid some
information to be lost perpetually resulting from
Fitness function is an important principle on selection and crossover operations, but also ensure
evaluating an individual. In this study, the fitness is validity of genetic arithmetic. By using the above
function of the average deviation between genetic algorithm operation, appropriate network
experimental and predicted values of viscosity (l). To weight and threshold are obtained.
do this, firstly the objective function according to
mean square error of neural network is defined in Eq.
1. Then the fitness value of a chromosome is Back Propagation Neural Network (BPNN) Learning Based
On GA Optimization
calculated using Eq. 2.

According to the GA method the optimization is to

maximize the fitness values (Ffitness) which would
(1) lead to the minimization of the total mean squared
(2) error from Eq. 1. This makes optimization of the
weights and thresholds of the BPNN and successively
where N stands for the number of the training data, µ ical BPNN be is improved during achieved in training. As
is the expected output, and µ iexp is the experimental seen in Eq. 1, the minimizing process of d value is the
output. Thus, the smaller the network’s total mean adjustment and optimization process of weights and
squared error, the closer a fitness value to 1 thresholds of the BPNN. Therefore, the GA is used to
(maximum). Once fitness values of all chromosomes optimize the weights and thresholds of the BPNN. It is
are evaluated, a population of chromosomes is updated the weights and thresholds optimization that is
using three genetic operators: selection, crossover and addressed in the current work. The BPNN learning
mutation. algorithm composed of two stages: firstly employing

3. Selection operation
GA to search for optimal or approximate optimal completed by applying a BP algorithm on the GA
connection weights and thresholds for the network, established initial connection weights and thresholds.
then uses the back-propagation learning rule and If the BP network’s total mean squared error is larger
training algorithm to adjust the final weights and than the expected error, the weights and thresholds
thresholds. The operations are as follows: The BPNN will be updated; otherwise, they are saved as initial
weights Wij and thresholds bi are initialized as genes value of BP network training. After that, they are
of chromosome, and then the global optimum is further adjusted under BP learning rule to the best
searched through selection, crossover and mutation results, by which the viscosity can be accurately
operators of genetic algorithm. This procedure is
predicted. The overall GA-NN training algorithm is GA-NN algorithm was performed using initial
represented in Fig. 2. population and based on results, population size,
selection probability, crossover probability and
mutation probability are determined. The values were
Training And Testing Database
found to be 40, 0.8, 0.90 and 0.001, respectively. All
optimal calculated weights and thresholds of net are
Experimental data presented in literature for viscosity presented in Table 2.
in eight nanofluids Al 2O3 in water, Al2O3 in propylene Based on testing data, the optimized neural
glycol (PG), CuO in water, CuO in 40% water and network model was used to predict the nanofluid
60% ethylene glycol, TiO2 in water TiO2 in ethylene viscosity; the comparison between predictive values
glycol (EG), SiO2 in water and SiO2 in ethanol and experimental values is curried out and shown in
(EtOH), were used in this study. Dataset was Fig. 3. The results of Fig. 3 show good agreement
characterized by five input variables and an output between the predicted and the experimental values of
variable as follows: temperature, volume fraction, viscosity with relatively low mean absolute error,
nanoparticle size, base fluid viscosity and density ratio MARE = 0.0248 and high correlation coefficients, R
of base fluid to nanoparticle are input variables and = 0.999.
nanofluid viscosity is the output variable. The ranges The computational results were compared to
of input–output variables for each nanofluid are given obtain ones by a conventional back-propagation neural
in Table 1. network (BPNN). For this propose, the respective
deviation for each model is presented in Fig. 4. Based
RESULT AND DISCUSSION on deviation criteria we utilized the advantage term
(Huang & Moraga 2004) of GA-NN relative to BPNN
as follow.
The GA-NN is trained and tested based on the
database. Randomly 70% (267 data points) of the data
set were used for training and the rest 30% (114 data
points) was used for testing. Because of deference in
dimension and magnitude of original experimental where MARE is mean absolute relative error as
data, the training and testing data should be follow:
normalized before fed to GA-NN. A problem in BPNN
is the possibility of over-training. This problem can be
overcome by proper selection of the number of
neurons in hidden layer for generalization and avoid
over-training. The numbers of hidden neurons were As it can seen form Fig. 4 the analysis of results
determined by trial and error. At first, analysis was showed that the MARE corresponding to GA-NN and
started with 2 neurons in the hidden layer and conventional BPNN for all nanofluids, are 0.0248 and
gradually the number neurons increased. The 0.0409 respectively. Based on Eq. 3, Ad = 39%, it
performance of network in training phase must means GANN model can improve the accuracy by
increase with increasing the number of neuron, while 39%, that show suitability and robustness of the
the performance of network in testing data phase lead proposed model. Among of nanofluids, for instance
to optimum value at an optimal number of hidden the performance of trained GANN for Al2O3–H2O,
neurons. In this way, the trained network is CuO–H2O and TiO2–EG are presented in Figs. 5 and
generalized and has best prediction on testing data. 6. Among of nanofluids that have been studied the
Thus use of more or less than the optimal number of nanofluids CuO–H2O, and TiO2–EG, have the
hidden neurons was discouraged. For this study, the minimum advantage (Ad = 9.8%) and maximum
mean square error (MSE) was chosen as a measure of advantage (Ad = 57%) respectively.
the performance of the nets. The network model with Finally, it can be that concluded the proposed
one hidden layer and nine neurons in hidden layer with model has more accurate prediction ability for all
MSE = 0.0074 lead to the best prediction. nanofluids, while due to a few available experimental
data slightly less accuracy can be seen in the case of between other models prediction and experimental
Al2O3–PG with MARE = 4.1%. data especially in Fig. 7b, c and d. For example in Fig.
GA-NN was compared to other analytical and 7b, c and d the mean absolute relative error for
empirical models and results shown in Fig. 7. This Batchelor and Einstein models are 50% up to 75%.
figure compares the temperature variation of the Also in Fig. 7c for Masoumi et al. model the mean
predicted results of Al 2 O3– H2O, CuO–H2O and TiO2– absolute relative error is 23%.
EG nanofluids with the experimental results and with The Einstein and Batchelor models are a simple
other models (Batchelor 1977; Masoumi et al. 2009). correlation for fluid with a low concentration of
As is shown the GA-NN model could well predict the spherical nanoparticles. Furthermore in these models
effect of the temperature variations on the nanofluid the effect of nanoparticle size is not considered in
viscosity. The trends of this model in outside of effective viscosity. Therefore as shown in Fig. 7b, c
experimental data are determined using unseen data and d, the accuracy of those models reduce seriously
and shown in Fig. 7. There are a great deal of deviation in large volume fraction of nanoparticle in suspension.
In Masoumi et al. 2009 the model parameters are relatively low MARE (2.48%) in prediction, the high
adjusted for limited nanoparticle size (i.e. nanofluid correlation coefficient between predicted and
Al2O3– H2O; with nanoparticles 28 and 13 nm), experimental value is a powerful tool for accurate
consequently this model not usable for wide range of determination of viscosity of nanofluids. GA-NN
type and size of nanoparticles. model results were compared with conventional back
propagation neural network (BPNN), for this purpose
CONCLUSION an advantage factor (Eq. 3) base on GA-NN model was
defined. According to this equation, Ad = 39%, it
means GA-NN model can reduce the error by 39%.
In this study we successfully employed the hybrid This is another proof that the proposed GA-NN can
neural network and genetic algorithm for the build a useful and robust model. Compared with the
prediction of the effective viscosity of nanofluids. The other theoretical and empirical models that are
model could calculate the effective viscosity as a available in literature, the presented model in general
function of the temperature, the mean particle size, the has higher accuracy and precision.
nanoparticle volume fraction and the base fluid
viscosity. The GA-NN due to advantages including
 REFERENCES

Adio, S. A., Mehrabi, M., Sharifpur, M. & Meyer, J. P. 2016. Experimental investigation and
model development for effective viscosity of MgO–ethylene glycol nanofluids by using
dimensional analysis, FCM-ANFIS and GA-PNN techniques. International Communications
in Heat and Mass Transfer 72: 71–83. doi:10.1016/j.icheatmasstransfer.2016.01.005

Anoop, K. B., Sundararajan, T. & Das, S. K. 2009. Effect of particle size on the convective heat
transfer in nanofluid in the developing region. International Journal of Heat and Mass
Transfer 52(9–10): 2189–2195. doi:10.1016/j.ijheatmasstransfer.2007.11.063

Batchelor, G. K. 1977. The effect of Brownian motion on the bulk stress in a suspension of
spherical particles. Journal of Fluid Mechanics 83(1): 97–117.
doi:10.1017/S0022112077001062

Chandrasekar, M., Suresh, S. & Chandra Bose, A. 2010. Experimental investigations and
theoretical determination of thermal conductivity and viscosity of Al2O3/water nanofluid.
Experimental Thermal and Fluid Science 34(2): 210–216.
doi:10.1016/j.expthermflusci.2009.10.022

Chen, H., Ding, Y., He, Y. & Tan, C. 2007. Rheological behaviour of ethylene glycol based titania
nanofluids. Chemical Physics Letters 444(4–6): 333–337. doi:10.1016/j.cplett.2007.07.046

Chen, H., Ding, Y. & Tan, C. 2007. Rheological behaviour of nanofluids. New Journal of Physics
9(10): 367–367. doi:10.1088/1367-2630/9/10/367

Choi, S. U. S. & Eastman, J. A. 1995. Enhancing thermal conductivity of fluids with nanoparticles.
International mechanical engineering congress and exhibition. United States.

Duangthongsuk, W. & Wongwises, S. 2009. Measurement of temperature-dependent thermal

conductivity and viscosity of TiO2-water nanofluids. Experimental Thermal and Fluid
Science 33(4): 706–714. doi:10.1016/j.expthermflusci.2009.01.005

He, Y., Jin, Y., Chen, H., Ding, Y., Cang, D. & Lu, H. 2007. Heat transfer and flow behaviour of
aqueous suspensions of TiO2 nanoparticles (nanofluids) flowing upward through a vertical
pipe. International Journal of Heat and Mass Transfer 50(11–12): 2272–2281.
doi:10.1016/j.ijheatmasstransfer.2006.10.024

Hemmat Esfe, M., Saedodin, S., Sina, N., Afrand, M. & Rostami, S. 2015. Designing an artificial
neural network to predict thermal conductivity and dynamic viscosity of ferromagnetic
nanofluid. International Communications in Heat and Mass Transfer 68: 50–57.
doi:10.1016/j.icheatmasstransfer.2015.06.013

Hornik, K., Stinchcombe, M. & White, H. 1989. Multilayer feedforward networks are universal
approximators. Neural Networks 2(5): 359–366. doi:10.1016/0893-6080(89)90020-8
Huang, C. & Moraga, C. 2004. A diffusion-neural-network for learning from small samples.
International Journal of Approximate Reasoning 35(2): 137–161.
doi:10.1016/j.ijar.2003.06.001

Jain, L. & Fanelli, A. M. (Eds.). 2000. Recent Advances in Artificial Neural Networks: Design
and Applications. USA: CRC Press, Inc.

Kamiński, M. & Ossowski, R. L. 2014. Prediction of the effective parameters of the nanofluids
using the generalized stochastic perturbation method. Physica A: Statistical Mechanics and
its Applications 393: 10–22. doi:10.1016/j.physa.2013.09.015

Kole, M. & Dey, T. K. 2010. Viscosity of alumina nanoparticles dispersed in car engine coolant.
Experimental Thermal and Fluid Science 34(6): 677–683.
doi:10.1016/j.expthermflusci.2009.12.009

Lee, S. & Choi, S. U. S. 1996. Application of metallic nanoparticle suspensions in advanced

cooling systems. international mechanical engineering congress and exhibition.

Lu, W.-Q. & Fan, Q.-M. 2008. Study for the particle’s scale effect on some thermophysical
properties of nanofluids by a simplified molecular dynamics method. Engineering Analysis
with Boundary Elements 32(4): 282–289. doi:10.1016/j.enganabound.2007.10.006

Mahbubul, I. M., Chong, T. H., Khaleduzzaman, S. S., Shahrul, I. M., Saidur, R., Long, B. D. &
Amalina, M. A. 2014. Effect of Ultrasonication Duration on Colloidal Structure and Viscosity
of Alumina–Water Nanofluid. Industrial & Engineering Chemistry Research 53(16): 6677–
6684. doi:10.1021/ie500705j

Masoumi, N., Sohrabi, N. & Behzadmehr, A. 2009. A new model for calculating the effective
viscosity of nanofluids. Journal of Physics D: Applied Physics 42(5): 55501.
doi:10.1088/0022-3727/42/5/055501

Mehrabi, M., Sharifpur, M. & Meyer, J. P. 2013. Viscosity of nanofluids based on an artificial
intelligence model. International Communications in Heat and Mass Transfer 43: 16–21.
doi:10.1016/j.icheatmasstransfer.2013.02.008

Meybodi, M. K., Naseri, S., Shokrollahi, A. & Daryasafar, A. 2015. Prediction of viscosity of
water-based Al2O3, TiO2, SiO2, and CuO nanofluids using a reliable approach.
Chemometrics and Intelligent Laboratory Systems 149: 60–69.
doi:10.1016/j.chemolab.2015.10.001

Nguyen, C. T., Desgranges, F., Galanis, N., Roy, G., Maré, T., Boucher, S. & Angue Mintsa, H.
2008. Viscosity data for Al2O3–water nanofluid—hysteresis: is heat transfer enhancement
using nanofluids reliable? International Journal of Thermal Sciences 47(2): 103–111.
doi:10.1016/j.ijthermalsci.2007.01.033

Nguyen, C. T., Desgranges, F., Roy, G., Galanis, N., Maré, T., Boucher, S. & Angue Mintsa, H.
2007. Temperature and particle-size dependent viscosity data for water-based nanofluids –
 Hysteresis phenomenon. International Journal of Heat and Fluid Flow 28(6): 1492–1506.
doi:10.1016/j.ijheatfluidflow.2007.02.004

Pastoriza-Gallego, M. J., Casanova, C., Legido, J. L. & Piñeiro, M. M. 2011. CuO in water
nanofluid: Influence of particle size and polydispersity on volumetric behaviour and
viscosity. Fluid Phase Equilibria 300(1–2): 188–196. doi:10.1016/j.fluid.2010.10.015

Pastoriza-Gallego, M. J., Lugo, L., Cabaleiro, D., Legido, J. L. & Piñeiro, M. M. 2014.
Thermophysical profile of ethylene glycol-based ZnO nanofluids. The Journal of Chemical
Thermodynamics 73: 23–30. doi:10.1016/j.jct.2013.07.002

Prasher, R., Song, D., Wang, J. & Phelan, P. 2006. Measurements of nanofluid viscosity and its
implications for thermal applications. Applied Physics Letters 89(13): 133108.
doi:10.1063/1.2356113

Sundar, L. S., Hortiguela, M. J., Singh, M. K. & Sousa, A. C. M. 2016. Thermal conductivity and
viscosity of water based nanodiamond (ND) nanofluids: An experimental study. International
Communications in Heat and Mass Transfer 76: 245–255.
doi:10.1016/j.icheatmasstransfer.2016.05.025

Sundar, L. S., Singh, M. K. & Sousa, A. C. M. 2016. Enhanced thermal properties of nanodiamond
nanofluids. Chemical Physics Letters 644: 99–110. doi:10.1016/j.cplett.2015.11.028

Syam Sundar, L., Singh, M. K. & Sousa, A. C. M. 2013. Investigation of thermal conductivity and
viscosity of Fe3O4 nanofluid for heat transfer applications. International Communications in
Heat and Mass Transfer 44: 7–14. doi:10.1016/j.icheatmasstransfer.2013.02.014

Turgut, A., Tavman, I., Chirtoc, M., Schuchmann, H. P., Sauter, C. & Tavman, S. 2009. Thermal
Conductivity and Viscosity Measurements of Water-Based TiO2 Nanofluids. International
Journal of Thermophysics 30(4): 1213–1226. doi:10.1007/s10765-009-0594-2