Powder Technology 257 (2014) 11–19

Contents lists available at ScienceDirect

Powder Technology
journal homepage: www.elsevier.com/locate/powtec

Development of empirical models with high accuracy for estimation of drag

coefficient of flow around a smooth sphere: An evolutionary approach
Reza Barati a, Seyed Ali Akbar Salehi Neyshabouri b,⁎, Goodarz Ahmadi c
a
Faculty of Civil and Environmental Engineering, Tarbiat Modares University, Tehran, Iran
b
Water Engineering Research Center, Department of Civil Engineering, Tarbiat Modares University, Tehran, Iran
c
Department of Mechanical and Aeronautical Engineering, Clarkson University, Potsdam, NY, USA

a r t i c l e i n f o a b s t r a c t

Article history: An accurate correlation for the smooth sphere drag coefficient with wide range of applicability is a useful tool in
Received 23 October 2013 the field of particle technology. The present study focuses on the development of high accurate drag coefficient
Received in revised form 12 February 2014 correlations from low to very high Reynolds numbers (up to 106) using a multi-gene Genetic Programming
Accepted 14 February 2014
(GP) procedure. A clear superiority of GP over other methods is that GP is able to determine the structure and
Available online 22 February 2014
parameters of the model, simultaneously, while the structure of the model is imposed by the user in traditional
Keywords:
regression analysis, and only the parameters of the model are assigned. In other words, in addition to the
Particle motion parameters of the model, the structure of it can be optimized using GP approach. Among two new and high
Sphere drag accurate models of the present study, one of them is acceptable for the region before drag dip, and the other is
Reynolds number applicable for the whole range of Reynolds numbers up to 106 including the transient region from laminar to
Multi-gene Genetic Programming turbulent. The performances of the developed models are examined and compared with other reported models.
The results indicate that these models respectively give 16.2% and 69.4% better results than the best existing
correlations in terms of the sum of squared of logarithmic deviations (SSLD). On the other hand, the proposed
models are validated with experimental data. The validation results show that all of the estimated drag
coefficients are within the bounds of ±7% of experimental values.
© 2014 Elsevier B.V. All rights reserved.

1. Introduction on experimental data when inertial effects are significant (i.e. higher
Reynolds numbers).
The motion of particles in fluids is a key subject in many problems The drag coefficient of a smooth sphere in incompressible flow is a
in the fields of chemical and metallurgical engineering as well as function of Re based on both theoretical investigations and numerous
mechanical and environmental engineering. The solution of these experimental data (Kreith [2]). The main classes of the dependence of
problems generally involves determining the local behavior of flow drag coefficient on Reynolds number are (1) very low Reynolds number
and the interaction between solid and liquid phases through the flow (i.e. creeping flow), (2) moderate Reynolds number flow (i.e.
knowledge of hydrodynamic forces such as drag. The drag force is laminar boundary layer), and (3) very large Reynolds number flow
the combination of the normal (i.e. pressure) and tangential (i.e. (i.e. turbulent boundary layer) (Munson et al. [3]). In the first class
wall shear stress) forces on the body in the flow direction. However, (Re b 1), the flows reflect entirely the viscous effect of flow with no
the distributions of the pressure and wall shear stress are often very separation results. By increasing Reynolds numbers (i.e. increasing the
difficult to achieve, so the magnitude of the drag force can be particle size or flow velocity for a given Kinematic viscosity), the
determined only through the knowledge of drag coefficient. separation region can be observed at Re ≈ 10, and the region increases
Analytical determination of the drag coefficient such as Stokes' law until Re ≈ 1000, where most of the drag is due to pressure drag rather
is only valid for Reynolds number, Re, less than 0.1 (Flemmer and than frictional drag. Parenthetically, it should be noted that the value of
Banks [1], Kreith [2]), although the drag coefficient can be the drag coefficient decreases, as wake area becomes larger. At a
ascertained using empirical and semi-empirical correlations based sufficiently high Reynolds number (103 b Re b 105), the drag coefficient
is relatively constant (Munson et al. [3]). When transition from laminar
to turbulent flow occurs, a dramatic dip (up to almost 80%) in the drag
coefficient appears at critical value Re ≈ 2 × 105 since the turbulent
⁎ Corresponding author at: Tarbiat Modares University, Al-Ahmad Ave., Tehran I.R. Iran
P.O. Box 14115-143. Tel.: +982182883316; fax:+982182883381.
boundary layer travels further along the surface into the adverse
E-mail addresses: [email protected], [email protected] (R. Barati), pressure gradient on the rear portion of the sphere before the
[email protected] (S.A.A.S. Neyshabouri), [email protected] (G. Ahmadi). separation, so the wake is smaller, causing less pressure drag. After

http://dx.doi.org/10.1016/j.powtec.2014.02.045
0032-5910/© 2014 Elsevier B.V. All rights reserved.
12 R. Barati et al. / Powder Technology 257 (2014) 11–19

Fig. 1. Illustration of the variations of the drag coefficient with Reynolds numbers using reliable data points of Stokes regime and available experiments in the literature [5,8,11].

Table 1
Summary of some empirical relationships for Re b 2 × 105.

Investigator Model and Reynolds number range Equation no.

Rouse [14] ^ D ¼ 24 þ 30:5 þ 0:34 for Reb2  105

C (1)
Re Re

Engelund and Hansen [13] ^ D ¼ 24 þ 1:5 for Reb2  105

C (2)
Re


Clift and Gauvin [17]a ^ D ¼ 24 1 þ 0:152Re0:677 þ
C 0:417
for Reb2  105 (3)
Re 1þ5070Re−0:94

8 24
>
> for Reb0:1;
> Re
>
>
>
22:7300
þ 0:0903 þ 3:6900 for 0:1bReb1;
>
>
Re Re2
>
> 29:1667
− 3:8889
þ 1:2220 for 1bReb10;
>
> Re Re2
>
> 46:5000 − 116:6700
< Re 2 þ 0:6167 for 10bReb100;
Morsi and Alexander [11] ^ D ¼ 98:3300 2778
C
Re
(4)
> − Re2 þ 0:3644 for 100bReb1000;
>
>
Re
>
> − 4:7510
4
þ 0:3570
> 148:6200
> for 1000bReb5000;
>
>
Re Re2
>
> − Re þ
490:5460 57:8710 4
þ 0:4600 for 5000bReb10; 000;
>
> 2
> Re
: 1662:5000 5:4167106
− Re þ Re2
þ 0:5191 for 10; 000bReb50; 000

Graf [36] ^ D ¼ 24 þ 7:30:5 þ 0:25 for Reb2  105

C (5)
Re 1þRe

Flemmer and Banks [1]a ^ D ¼ 24 10E where E ¼ 0:383Re0:356 −0:207Re0:396 −

C 0:143
for Reb2  105 (6)
Re 1þð logReÞ2

3:18
Khan and Richardson [20]a ^ D ¼ 2:49Re−0:328 þ 0:34Re0:067
C for Reb2  105 (7)

( 
−0:25 )0:25
h    i2:5 2
Swamee and Ojha [21] ^ D ¼ 0:5 16 24 1:6 þ 130 0:72
C þ 40;000 þ1 for Reb1:5  105 (8)
Re Re Re

pffiffiffiffiffiffi 
Yen [37] ^ D ¼ 24 1 þ 0:15 Re þ 0:017Re − 0:208
C for Reb2  105 (9)
Re 1þ104 Re−0:5


Haider and Levenspiel [19]a ^ D ¼ 24 1 þ 0:150Re0:681 þ 0:407 −1 for Reb2  105
C (10)
Re 1þ8710Re

h
i
Cheng [6] ^ D ¼ 24 ð1 þ 0:27ReÞ0:43 þ 0:47 1− exp −0:04Re0:38
C for Reb2  105 (11)
Re

Terfous et al. [22] ^ D ¼ 2:689 þ 21:683 þ 0:131

C − 10:616 þ 12:216 for 0:1bReb5  104 (12)
Re Re2 Re0:1 Re0:2

^ D ¼ 3808½ð1;617;933=2030Þþð178;861=1063ÞReþð1219=10842ÞRe  for Reb118; 300

2
Mikhailov and Freire [7] C 681Re½ð77;531=422Þþð13;529=976ÞRe−ð1=71;154ÞRe 
(13)

a
These models were improved by Brown and Lawler [5].
 R. Barati et al. / Powder Technology 257 (2014) 11–19 13

Table 2
Empirical relationships for Reynolds numbers up to 106.

Investigator Model and Reynolds number range Equation no.

8 24 3
Clift et al. [16] >
> þ 16
Re
 for Reb0:01 (14)
>
>
>
> 0:82−0:05½ logRe
Re 1 þ 0:1315Re 0:01bRe ≤ 20
24
> for
>
>

>
>
> 24 1 þ 0:1935Re0:6305
> for 20bRe ≤ 260
>
> Re
> 1:6435−1:1242½ logReþ0:1558½ logRe2
<
^ D ¼ 10
C 2 3
for 260bRe ≤1500
>
> 10−2:4571þ2:5558½ logRe−0:9295½ logRe2 þ0:1049½ logRe for 1500bRe≤1:2  104
>
>
>
> 10 −1:9181þ0:6370½ logRe−0:0636½ logRe
for 1:2  104 bRe≤ 4:4  104
>
>
> −4:3390þ1:5809½ logRe−0:1546½ logRe2
>
>
> 10 for 4:4  104 bRe≤ 3:38  105
>
>
>
> 29:78−5:3½ logRe for 3:38  105 bRe ≤4  105
:
0:1½ logRe−0:49 for 4  105 bRe≤ 106


^ D ¼ 1−0:5 expð0:182Þ þ 10:11Re−2=3 exp 0:952Re−1=4
C


−0:03859Re−4=3 exp 1:30Re−1=2
Ceylan et al. [9]
 (15)
þ 0:037  10−4 Re exp −0:125  10−4 Re


−0:116  10−10 Re2 exp −0:444  10−5 Re for 0:1bReb106

h i0:1
^D ¼
C þ φ4
1
ðφ1 þφ2 Þ−1 þðφ3 Þ−1

10
10
10
where φ1 ¼ 24Re−1 þ 21Re−0:67 þ 4Re−0:33 þ ð0:4Þ10 ;
Almedeij [10] 
−10 −1 (16)
φ2 ¼ 0:148Re0:11 þ ð0:5Þ−10 ;

10 
−10 −1
φ3 ¼ 1:57  108 Re−1:625 ; φ4 ¼ 6  10−17 Re2:63 þ ð0:2Þ−10 for Reb106

2:6ðRe
−7:94

5Þ 0:411ð263000Þ
Re
Morrison [23] ^ D ¼ 24 þ
C 1:52 þ −8:00 þ Re0:8
for Reb106 (17)
Re 1þðRe 461000
5Þ 1þð263000Þ
Re

this abrupt descent, the value of the drag coefficient increases by in- to develop high accurate models for the estimation of the drag coeffi-
creasing Reynolds numbers. Finally, for Re N 106 a constant value of cient of the free falling smooth sphere. Unlike traditional regression
the drag coefficient (≈0.2) is acceptable (Potter et al. [4]). analysis in which the structure of the model must be specified, GP auto-
Most of the information pertaining to drag force on the sphere arises matically evolves both the structure and parameters of the drag coeffi-
from numerous experiments with wind tunnels, water tunnels, towing cient estimation model. Therefore, both parameters and structure of
tanks, and other ingenious devices (Munson et al. [3]). Experimental the model will be optimized. The experimental data points of Voloshuk
data of the drag coefficient of spherical particles have been presented and Sedunow [8], and Brown and Lawler [5] will be used to develop
in the literature having a wide range of Re. However, some of the avail- empirical models. Seventeen popular correlations will be reviewed
able experimental data are not accurate, adequately. Brown and Lawler for comparison purposes, and experimental data points of Morsi and
[5] reviewed the experimental studies of sphere drag coefficient for Re b Alexander [11] along with analytical solution of Stokes regime will be
2 × 105. They assembled 606 data points which were originally present- considered for the validation of the developed models.
ed in tabular form. By excluding some experimental data for various
reasons, Brown and Lawler [5] presented 480 very high quality data 2. Literature review
points by considering wall effects. This data set seems acceptable
among other researchers for developing correlations (Cheng [6], Many empirical or semi-empirical correlations that vary somewhat
Mikhailov and Freire [7]). On the other hand, Voloshuk and Sedunow in form have been developed to estimate the standard drag curve of
[8] presented the experimental data for higher Reynolds numbers smooth sphere using regression techniques. Seventeen of them which
with good quality. This data set was also used in several studies such as are allocated in two groups based on range of applicability are presented
Ceylan et al. [9] and Almedeij [10]. Fig. 1 illustrates all of the mentioned in Tables 1 and 2. The first group covers Reynolds numbers up to 2 × 105
data. The variations of the drag coefficient with Reynolds numbers can while the second covers Reynolds numbers up to 106. A critical discus-
follow as explained in the previous paragraph by considering Fig. 1. sion about these models will be presented in the next paragraph.
In the previous studies, the regression analyses were applied to ob- Rubey [12] suggested a simple equation for estimation of the drag
tain a correlation for the estimation of the drag coefficient of spherical coefficient of sphere based on combining Stokes' law with the impact
particles. Several forms and procedures such as multi-segment poly- formula. Engelund and Hansen [13] revised Rubey's equation to im-
nomial, exponential function, piecewise matched, power function and prove it for natural sands and gravels. Rouse [14] proposed a relation
rational fraction were used in these studies. These forms of the correla- that has been widely used in the environmental field. Morsi and Alexan-
tions were developed by imposing general arithmetic operations (i.e. der [11] developed a model based on the multi-segment polynomial
plus, minus, multiplication and division), and/or some function set equations that are valid for the range of Re b 5 × 104 (Nasrollahi et al.
(e.g. logarithm, and exponential functions) without considering other [15]). Clift et al. [16] analyzed the available experimental data by consid-
popular functions such as sin, cos, tan, tanh, and natural logarithm. ering wall effects, compressibility effects, non-continuum effects and
Therefore, the performance of existing correlations is less than perfect. support interference. They established another multi-segment model
In the present study, multi-gene Genetic Programming (GP) is adopted using several polynomial equations. Two recent models have good
14 R. Barati et al. / Powder Technology 257 (2014) 11–19

Fig. 2. A typical multi-gene tree-based model.

accuracy, but these complicated regression equations suffer from and compared it with seven previously-developed formulas. This corre-
gradient discontinuities between subintervals. Clift and Gauvin [17] lation consists of two terms (1) an extended Stokes' law applicable ap-
proposed a simple five-constant correlation to estimate the drag coeffi- proximately for Re b 100, and (2) an exponential function accounting
cient. After them, this correlation has been improved by Turton and for slight deviations from the Newton's law for high Re. Cheng [6]
Levenspiel [18], Haider and Levenspiel [19], and Brown and Lawler [5] claimed that the proposed correlation gives the best representation of
based on regression analysis by minimizing the sum of squared of loga- the historical data for the region before drag dip. Terfous et al. [22] de-
rithmic deviations (SSLD). Flemmer and Banks [1] used a regression veloped an equation for Re b5 × 104 by using the simple series function
technique to generate another expression of an exponential form for and applying the least-squares method. However, this model is not ap-
the range of Re b 3 × 105. Brown and Lawler [5] improved this expres- plicable for Re b 0.1 because they used the term 21.683/Re instead of the
sion both in range and accuracy using local minimizations. Khan and traditional term 24/Re in the equation. Morrison [23] used a piecewise
Richardson [20] compiled experimental results of various researchers. matched procedure to develop a model that covers the transition from
They proposed a drag equation of a power form using nonlinear regres- laminar to turbulent flow region. Recently, Mikhailov and Freire [7]
sion on 300 data points. This equation was also improved by Brown and presented an expression for the estimation of the drag coefficient of a
Lawler [20]. Swamee and Ojha [21] presented another model by com- falling sphere for the ranges of Re b 118,300 based on Shanks transfor-
bining four expressions that were obtained for different ranges of Reyn- mation which is a non-linear rational fractional transformation of the
olds numbers (i.e. a piecewise matched procedure). The authors series of Goldstein to Oseen's equation.
claimed that the resulting equation faithfully described the drag Briefly, it can be said that the above equations of the drag coefficient
coefficient-versus-Reynolds number variation of a spherical particle estimation suffer from the complicated correlation, bounds of applica-
for R b 1.5 × 105. Ceylan et al. [9] developed a model in the form of bility, and/or low accuracy. Therefore, the development of a simple
the series by analyzing several transfer phenomena for the ranges of and accurate model with a high range of applicability is vital.
0.1 b Re b 106. The effects of the form drag and skin friction drag were
separately considered using two series in the model. However, the accu-
racy of this model is low, as will be shown. Almedeij [10] presented a 3. Genetic Programming
model, which is valid for the ranges of Re b 106, using the concept of
the matching asymptotically, and employing regression techniques. Genetic Programming (GP) is a random-based procedure for auto-
He developed the model by using four functions that can be combined matically learning the most “fit” computer programs by means of artifi-
into an overall relationship. However, this model suffers from the com- cial evolution (Johari et al. [24]). Recently, GP has been successfully
plicated form. Cheng [6] recommended a five-parameter correlation, applied in many applications such as the prediction of the soil–water
characteristic of soils (Johari et al. [24]), the estimation of the bridge
pier scour (Azamathulla et al. [25]), and the prediction of the outflow
hydrograph from earthen dam breach (Hakimzadeh et al. [26]).
Table 3 GP, which is a branch of the conventional genetic algorithm (GA;
Range of parameters of multi-gene GP. Holland [27]), was initially developed by Koza [28]. The significant
Parameter Range superiority of GP over GA is that in GP there is no need to define the
structure of the model a priori. In other words, GP can determine not
Population size 200–10,000
Number of generations 400–10,000 only the coefficients and parameters of the model, but also and more
Maximum number of genes 1–10 importantly, the form of the model itself (Johari et al. [24]). GP random-
Maximum number of nodes per tree 1–15 ly generates a population of equal or unequal length computer pro-
Maximum depth of trees 3–15
grams (or symbolic expression) with a high level of diversity. The
Probability of multi-gene GP tree mutationa 0.1–0.2
Probability of multi-gene GP tree cross overa 0.75–0.85
programs are represented by tree-based structures using variables (ter-
Probability of multi-gene GP tree direct copya 0.05 minal) and several mathematical operators (function) that can be se-
Size of the tournament 2–3 lected from sets of terminals and functions, respectively. The terminal
Arithmetic operations and function set (+, −, ×, ÷, square, tanh, set contains the numerical constants and external inputs of the program
sin, cos, tan, log, ln, exp)
while the function set contains the basic arithmetic operators such as +,
−, ∗, and / and function calls such as ex, sin, cos, tan, tanh, log, ln, sqrt
a
Sum must be equal to 1.
 R. Barati et al. / Powder Technology 257 (2014) 11–19 15

Fig. 3. Developed drag coefficient model for the region before drag dip along with experimental data of Brown and Lawler [5].

and power. After choosing the best symbolic expressions through defin- where y is the output variable, a0 is the bias term and ai is weight of the
ing a fitness function, an improved population will be created using ith gene.
crossover and mutation operators. The iterations proceed until they An example of multi-gene tree-based model is presented in Fig. 2,
meet stopping criteria (e.g. reaching the specified number of genera- where an output variable y is predicted using an input variable x. As
tions). It is informative to express that GP is often known as symbolic re- can be seen in Fig. 2, multi-gene GP is a linear combination of nonlinear
gression, when an empirical model is developed using collected data terms, and this feature makes it possible to recognize the pattern of
from a process or system. More details of GP procedure can be found engineering problems in a highly precise manner (Hinchliffe et al. [29]).
in Koza [28]. In the present study, GPTIPS, which is a new “Genetic Programming
In order to improve fitness for non-linear behavior procedures, the and Symbolic Regression” code (Searson [31]), was adapted to perform
multi-gene GP which is a branch of standard GP can be used (Hinchliffe a multi-gene GP for the development of the sphere drag coefficient
et al. [29], and Hiden [30]). In the traditional GP, the model expresses as models. GPTIPS which uses lexicographic tournament selection as an
a single tree while, in the multi-gene GP, several trees may define the effective technique for controlling the bloating of the model has the
model through a weighted linear combination of each gene plus a bias capability of setting some restrictions on initial parameters such as the
term as maximum number of genes, maximum depth of trees and genes and
maximum number of nodes per tree to avoid the most recent problems
of bloat in GP. There are several successful applications of GPTIPS in the
y ¼ a0 þ a1  gene1 þ a2  gene2 þ … þ an  genen ð18Þ different fields such as predicting the liquefaction resistance of sand–silt

Fig. 4. Developed drag coefficient model for the wide range of particle Reynolds numbers along with experimental data of Brown and Lawler [5], and Voloshuk and Sedunow [8].
16 R. Barati et al. / Powder Technology 257 (2014) 11–19

Table 4 It is notable that SSLD and RMSLD were calculated based on logarith-
Comparison of the performance evaluation criteria of the relationships for Re b 2 × 105. mic values of the drag coefficients while SRE can be used to evaluate the
Order of accuracy Reference SSLD RMSLD SRE original value of the drag coefficients.
In order to select appropriate models, GPTIPS is run over 100 times,
1 Present study Eq. (22) 0.0881 0.0136 10.7171
2 Cheng [6] 0.1051 0.0148 11.8567 because multi-gene GP is a stochastic procedure. As will be shown, the
3 Morsi and Alexander [11] 0.1314 0.0176 13.0465 best representation of available experimental data of the smooth sphere
4 Clift and Gauvin [17]a 0.1502 0.0177 15.4402 drag coefficient in the literature for Reynolds numbers up to 2 × 105 and
5 Haider and Levenspiel [19]a 0.1510 0.0177 15.5260
up to 106 can be estimated by Eqs. (22) and (23), respectively.
6 Flemmer and Banks [1]a 0.1978 0.0203 17.4739
7 Khan and Richardson [20]a 0.2091 0.0209 17.4464

8 Swamee and Ojha [21] 0.2778 0.0242 20.4107 ^ ¼ 5:4856  109 tanh 4:3774  10−9 =Re þ 0:0709 tanhð700:6574=ReÞ
CD
9 Terfous et al. [22] 0.4670 0.0322 23.1950 þ0:3894 tanhð74:1539=ReÞ−0:1198 tanhð7429:0843=ReÞ
10 Mikhailov and Freire [7] 0.4703 0.0319 25.9766 5
þ1:7174 tanh½9:9851=ðRe þ 2:3384Þ þ 0:4744 for Reb2  10
11 Yen [37] 0.8475 0.0420 38.7227
12 Rouse [14] 2.1313 0.0666 51.3652 ð22Þ
13 Graf [36] 5.2163 0.1042 63.1865
14 Engelund and Hansen [13] 61.5283 0.3580 545.4005
h i
a
These models were improved by Brown and Lawler [5]. ^ ¼ 8  10−6 ðRe=6530Þ2 þ tanhðReÞ−8 ln ðReÞ= ln ð10Þ
C D

mixtures (Baziar et al. [32]), and predicting the ultimate bearing −2:081043 =½ReþRe2 
4
− ½lnðRe2 þ10:7563Þ= ln ð10Þ
2
þ9:9867 =Re
−0:4119e −2:1344e
capacity of shallow foundations on cohesionless soils (Shahnazari and
þ0:1357e ½
Tutunchian [33]). The ranges of the initial parameters of multi-gene − ðRe=1620Þ2 þ10370=Re −3
−8:5  10 f2 ln ½ tanhð tanhðReÞÞ=
GP for runs of this study are summarized in Table 3. More descriptions
6
about the initial parameters and specifications of multi-gene GP can ln ð10Þ−2825:7162g=Re þ 2:4795 for Reb10
be found in Searson [31]. ð23Þ

4. Developed models based on multi-gene GP The parameters of the above models were rounded through the
sensitivity analysis. In other words, the decimal point accuracy of the
The sum of the squared deviations is a good objective function to parameters of the models was considered with sufficient digits in
minimize the errors between experimental data and calculated results order to fade the rounding off errors.
by considering the previous studies (Turton and Levenspiel [18]; Haider The standard drag curves represented by Eqs. (22) and (23) in
and Levenspiel [19]; Brown and Lawler [5]; Barati [34], Barati [35]). comparison with the experimental data are illustrated in Figs. 3 and 4,
Therefore, the fitness function of multi-gene GP is to minimize the respectively. First of all, it can be seen that these models truthfully fol-
sum of squared of logarithmic deviations (SSLD) between the estimated low the experimental data, and they can represent the mean values of
drag coefficient and experimental data points the experimental data with high accuracy. Furthermore, as shown in
Figs. 3 and 4, both developed models converge to Stokes solution at
N

X 2 very small Reynolds numbers, and the models can capture a minimum
Minimize SSLD ¼ ^
logCD − logC ð19Þ
D for the drag coefficient in a Reynolds number value around 5000
1
which agree with the experimental data. It should be noted that most
^ is the of existing correlations cannot capture this minimum, accurately. On
where CD is the experimental value of the drag coefficient, C D the other hand, the existing correlations of the second group, which
estimated drag coefficient, and N is the total number of the data.
are listed in Table 2, suffer from less accuracy in either moderate or
For assessing the performance of the proposed models compared to
large Reynolds numbers while Eq. (23) can estimate the experimental
existing correlations, root-mean-square of logarithmic deviation
data in all of the creeping flow, transitional, and turbulent regions,
(RMSLD) and sum of the relative errors (SRE), as well as SSLD are
precisely.
used as
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 5. Test of the developed equations
u N
uX
2
RMSLD ¼ t logCD − logC ^
D =N ð20Þ In order to examine the developed models, the results of the drag
1
coefficient obtained from the proposed models were compared with
those from models developed by other researchers in terms of SSLD,
0 1 RMSLD and SRE.
X
N C D −C^ Performance evaluation criteria for the correlation of the first and
D
SRE ¼ @ A: ð21Þ
CD second groups together with the corresponding values of the developed
1
models are listed in Tables 4 and 5, respectively. It should be stated that
the models are ranked in the increasing order of SSLD in Tables 4 and 5.
Generally, other criteria (i.e. RMSLD and SRE) also increase in Tables 4
Table 5
and 5 by increasing order of accuracy. The ranges of using the models
Comparison of the performance evaluation criteria of the relationships for Reynolds in the calculation of the criteria were its applicability range, which is
numbers up to 106. presented in Tables 1 and 2. Therefore, it should be noted that some of
the correlations such as recently developed models of Terfous et al.
Order Reference SSLD RMSLD SRE
of accuracy [22] and Mikhailov and Freire [7] have large values of SSLD (i.e. low ac-
curacy) although SSLD of them were calculated for 0.1 b Re b 5 × 104
1 Present 0.0982 0.0142 11.2853
study Eq. (23) and Re b 118,300, respectively, instead of whole range of Reynolds
2 Almedeij [10] 0.3210 0.0257 23.2094 numbers up to 2 × 105.
3 Clift et al. [16] 0.4467 0.0303 17.4004 As shown in Tables 1 and 2, both proposed models have the lowest
4 Morrison [23] 0.5672 0.0342 29.1768 values of SSLD, RMSLD and SRE (higher accuracy). Eqs. (22) and (23)
5 Ceylan et al. [9] 0.9186 0.0449 33.2925
improved 16.2% and 69.4% to match with the experimental data points
 R. Barati et al. / Powder Technology 257 (2014) 11–19 17

Fig. 5. Comparison between the validation data of the drag coefficient and the associated calculation results.

than the best existing correlations of the first and second groups in region before drag coefficient dip and with Reynolds numbers greater
terms of SSLD, respectively. The new models substantially improve the than Re N 2 × 105, respectively. Finally, it must be stated that a constant
fit to experimental data. Consequently, for the accurate estimation of value of the drag coefficient (≈0.2) is acceptable for higher Reynolds
the smooth sphere drag coefficient of engineering problems, Eqs. (22) numbers (i.e. Re N 106) (Almedeij [10], Potter et al. [4]).
and (23) can be utilized for problems with Reynolds numbers in the As mentioned previously, the developed equations were calibrated
using the data of Voloshuk and Sedunow [8], and Brown and Lawler
[5]. The data of Morsi and Alexander [11] along with analytical solution
of Stokes regime will be used to validate the proposed models.
Table 6
Validation of the developed models using the experimental data of Morsi and Alexander
Fig. 5 represents the comparison between the validation data of the
[11], and the analytical solution of Stokes regime. drag coefficient and the associated calculation results. The figure shows
that the results of both models are comparable with the measurement
Re CD −5% of CD Eq. (22) Eq. (23) +5% of CD
data.
0.002 12,000 11,400.00 12,008.86 12,034.71 12,600.00 The experimental values of the drag coefficient and ±5% of them are
0.004 6000 5700.00 6005.70 6017.32 6300.00
compared with the estimated values by the proposed models in Table 6.
0.007 3429 3257.14 3432.91 3438.94 3600.00
0.01 2400 2280.00 2403.80 2407.74 2520.00 The results indicated that the models have no significant error com-
0.04 600 570.00 602.85 603.54 630.00 pared to the experimental data. All of the estimated drag coefficients
0.07 343 325.71 345.57 345.88 360.00 by both models are within the bounds of ±5% of validation data except
0.1 240 228.00 242.66 242.84 252.00 for 3 data points. If the bounds are increased until ±7%, 3 data points
0.2 120 114.00 122.59 122.63 126.00
0.3 80 76.00 82.57 82.57 84.00
will also be covered.
0.5 49 46.55 50.56 50.53 51.45
0.7 36.5 34.68 36.83 36.80 38.32 6. Discussion
1 26.5 25.18 26.54 26.50 27.82
2 14.4 13.68 14.50 14.48 15.12
In this section, two issues about the multi-gene GP procedure will be
3 10.4 9.88 10.46 10.44 10.92
5 6.9 6.56 7.12 7.10 7.24 discussed: 1) Compatibility of multi-gene GP approach with the natural
7 5.4 5.13 5.60 5.58 5.67 of the problem and 2) evaluation of the level of the accuracy of the
10 4.1 3.90 4.36 4.36 4.31 developed models.
20 2.55 2.42 2.73 2.76 2.68 For the first issue, as mentioned previously, over 100 models were
30 2 1.90 2.12 2.14 2.10
50 1.5 1.42 1.58 1.57 1.58
developed with different forms of the equations using multi-gene GP
70 1.27 1.21 1.31 1.30 1.33 procedure. Although Eqs. (22) and (23) have the lowest errors than
100 1.07 1.02 1.08 1.07 1.12 the others, most of the developed models are better than the best
200 0.77 0.73 0.77 0.77 0.81 existing correlations. For example, Eq. (24) presents another model for
300 0.65 0.62 0.65 0.66 0.68
the estimation of the drag coefficient:
500 0.55 0.52 0.56 0.55 0.58
700 0.5 0.48 0.51 0.50 0.52
^ ¼ 24:0285=Re þ 9:4104Re=ðRe þ 2:8620Þ2 þ 0:0277½ ln ðRe þ 32:6397Þ2
C
1000 0.46 0.44 0.47 0.46 0.48 D
2 5
2000 0.42 0.40 0.41 0.41 0.44 −0:0128½ ln ðRe þ 52269Þ −0:4320 ln ðRe þ 2:5064Þ þ 3:5991 for Reb2  10 :
3000 0.4 0.38 0.40 0.40 0.42 ð24Þ
5000 0.385 0.36 0.39 0.39 0.40
7000 0.39 0.37 0.40 0.40 0.41
10,000 0.405 0.38 0.41 0.41 0.42 The values of SSLD, RMSLD and SRE of Eq. (24) are 0.0960, 0.0141,
20,000 0.45 0.43 0.44 0.44 0.47 and 11.5100, respectively. By considering Table 4, Eq. (24) has better
30,000 0.47 0.44 0.45 0.45 0.49 accuracy than the best existing model of other studies. The key issue is
50,000 0.49 0.46 0.46 0.46 0.51
that Stokes' law (i.e. 24/Re) can be observed in this model. This fact
18 R. Barati et al. / Powder Technology 257 (2014) 11–19

Fig. 6. Illustration of non-smooth curve of Eq. (25).

upholds that multi-gene GP procedure can adapt itself with the natural Results of the numerical and experimental validations indicated that
of the problem. However, for Eqs. (22) and (23), other functions appear two new correlations which are accurate, simple, and explicit illustrated
instead of 24/Re. For example, a form of tangent hyperbolic function the best representation of available experimental data for both sub-
plays the same role for the calculation of the drag coefficients in low critical and turbulent regions. The developed models by multi-gene GP
Reynolds number (Stokes' law region). substantially (up to almost 70%) improved the fit to experimental data
For the second issue, interestingly, it is observed that the value of than the best existing correlations in terms of SSLD. The correlations
SSLD can be decreased [less than the values of Eqs. (22) and (23)] by were also verified using another set of experimental data points. The
increasing the maximum depth of trees in the multi-gene GP procedure. results showed that all of the estimated drag coefficients were within
However, the obtained correlations by higher number of depth of trees the bounds of ±7% of experimental values.
have a non-smooth curve. For example, Eq. (25) presents one of these Finally, it can be said that the proposed models are useful tools for
models computer programs because of the high accuracy and simplicity of
them. Moreover, for future researches, it is hoped that the procedure
^ ¼ 24:03=Re þ 1:3255e tanhðReÞ þ 0:5756 ln ðRe þ 12:3504Þ−0:3910 ln ðRe þ 30:2029Þ
CD of this study can be successfully adopted for other problems in the
−0:0062 cosð0:13993ReÞ−3:3243 h f tanhðReÞ þ tanh½ sinhðReÞg i
10 10 field of particle transport.
þ ½237119 ln ðReÞ þ 164358= 5  10 sinðReÞ þ 4:8396  10
n o n o
9 14
− 1:5  10 sin½0:4343 ln ðReÞ = 10 ½ cosðReÞ þ tanhðReÞ

 Appendix A. Supplementary data
−6 −5
þ sin½ ln ðReÞ= ln ð10Þ 1:04  10 Re þ 1:60  10
5
þ 14:91 sin½ tanhðReÞ= logðRe þ 9:24Þ þ 0:3996 for Reb2  10 : Supplementary data to this article can be found online at http://dx.
ð25Þ doi.org/10.1016/j.powtec.2014.02.045.

The values of SSLD, RMSLD and SRE of Eq. (25) are 0.0777, 0.0127, References
and 10.3246, respectively. By considering Table 4, Eq. (25) has the
[1] R.L.C. Flemmer, C.L. Banks, On the drag coefficient of a sphere, Powder Technol. 48
lower errors than Eq. (22) and other correlations. However, as it can (1986) 217–221.
be seen in Fig. 6, the curve of Eq. (25) is non-smooth which is not [2] F. Kreith, Mechanical Engineering Handbook, CRC Press LLC, 1999.
acceptable. In other words, for each model with a lower error than [3] B.R. Munson, D.F. Young, T.H. Okiishi, W.W. Huebsch, Fundamentals of Fluid
Mechanics, John Wiley & Sons, Inc., New York, 2009.
Eq. (22), the curve of the model becomes non-smooth. Therefore, by [4] M.C. Potter, D.C. Wiggert, B.H. Ramadan, Mechanics of Fluids, SI Version, Cengage
considering available experimental data of drag coefficient, it can be Learning, 2011.
said that Eqs. (22) and (23) are almost the best possible smooth [5] P.P. Brown, D.F. Lawler, Sphere drag and settling velocity revisited, J. Environ. Eng.
129 (2003) 222–231.
standard curves which can be constructed. [6] N.S. Cheng, Comparison of formulas for drag coefficient and settling velocity of
spherical particles, Powder Technol. 189 (2009) 395–398.
7. Conclusions [7] M.D. Mikhailov, A.P. Freire, The drag coefficient of a sphere: an approximation using
Shanks transform, Powder Technol. 237 (2013) 432–435.
[8] V.M. Voloshuk, J.S. Sedunow, The Processes of Coagulation in Dispersed Systems,
In the present study, the existing correlations of drag coefficient Nauka, Moscow, 1971.
were discussed, critically. Then, a reliable and complete set of historical [9] K. Ceylan, A. Altunbaş, G. Kelbaliyev, A new model for estimation of drag force in the
flow of Newtonian fluids around rigid or deformable particles, Powder Technol. 119
data was collected for the development and validation of correlations
(2001) 250–256.
for the estimation of the smooth sphere drag coefficient. An effective [10] J. Almedeij, Drag coefficient of flow around a sphere: matching asymptotically the
procedure (i.e. multi-gene Genetic Programming) was used to develop wide trend, Powder Technol. 186 (2008) 218–223.
drag coefficient models through optimizing both parameters and [11] S.A. Morsi, A.J. Alexander, An investigation of particle trajectories in two-phase flow
systems, J. Fluid Mech. 55 (1972) 193–208.
structure of models. Because the procedure is stochastic, the multi- [12] W.W. Rubey, Settling velocity of gravel, sand, and silt particles, Am. J. Sci. 148 (1933)
gene GP was run over 100 times for selecting appropriate models. 325–338.
 R. Barati et al. / Powder Technology 257 (2014) 11–19 19

[13] F. Engelund, E. Hansen, Monograph on Sediment Transport in Alluvial Streams, [27] J.H. Holland, Adaptation in Natural and Artificial System, University of Michigan
Monograpsh Denmark Technical University, Hydraulic Lab, Denmark, 1967. Press, Ann Arbor Mich, 1975.
[14] H. Rouse, Fluid Mechanics for Hydraulic Engineers, Dover, New York, N.Y., 1938 [28] J.R. Koza, Genetic Programming: On the Programming of Computers by Means of
[15] A. Nasrollahi, S.A.A. Salehi Neyshabouri, G. Ahmadi, M.M. Namin, Numerical simula- Natural Selection, MIT Press, Cambridge, Mass, 1992.
tion of particle saltation process, Part. Sci. Technol. 26 (2008) 529–550. [29] M.P. Hinchliffe, M.J. Willis, H.G. Hiden, M.T. Tham, B. McKay, G. Barton, Modelling
[16] R. Clift, J.R. Grace, M.E. Weber, Bubbles, Drops, and Particles, Academic, New York, chemical process systems using a multi-gene genetic programming algorithm,
1978. Late Breaking Paper, GP'96, Stanford: USA, 1996.
[17] R. Clift, W.H. Gauvin, The motion of particles in turbulent gas streams, Proc. [30] H.G. Hiden, Data-Based Modelling using Genetic Programming, PhD Thesis Dept.
Chemeca, 70, 1970, pp. 14–28. Chemical and Process Engineering, University of Newcastle, UK, 1998.
[18] R. Turton, O. Levenspiel, A short note on the drag correlation for spheres, Powder [31] D.P. Searson, GPTIPS: Genetic Programming & Symbolic Regression for MATLAB,
Technol. 47 (1986) 83–86. User Guide, 2009.
[19] A. Haider, O. Levenspiel, Drag coefficient and terminal velocity of spherical and non- [32] M.H. Baziar, Y. Jafarian, H. Shahnazari, V. Movahed, M. Amin Tutunchian, Prediction
spherical particles, Powder Technol. 58 (1989) 63–70. of strain energy-based liquefaction resistance of sand–silt mixtures: an evolutionary
[20] A.R. Khan, J.F. Richardson, The resistance to motion of a solid sphere in a fluid, Chem. approach, Comput. Geosci. 37 (2011) 1883–1893.
Eng. Commun. 62 (1987) 135–150. [33] H. Shahnazari, M.A. Tutunchian, Prediction of ultimate bearing capacity of shallow
[21] P.K. Swamee, C.S.P. Ojha, Drag coefficient and fall velocity of nonspherical particles, foundations on cohesionless soils: an evolutionary approach, KSCE J. Civ. Eng. 16
J. Hydraul. Eng. 117 (1991) 660–667. (2012) 950–957.
[22] A. Terfous, A. Hazzab, A. Ghenaim, Predicting the drag coefficient and settling veloc- [34] R. Barati, Parameter estimation of nonlinear Muskingum models using Nelder–
ity of spherical particles, Powder Technol. 239 (2013) 12–20. Mead simplex algorithm, J. Hydrol. Eng. ASCE 16 (2011) 946–954.
[23] F.A. Morrison, An Introduction to Fluid Mechanics, Cambridge University Press, New [35] R. Barati, Application of Excel Solver for parameter estimation of the nonlinear Mus-
York, 2013. kingum models, KSCE J. Civ. Eng. 17 (2013) 1139–1148.
[24] A. Johari, G. Habibagahi, A. Ghahramani, Prediction of soil–water characteristic curve [36] W.H. Graf, Hydraulics of Sediment Transport, Water Resources Publications, Little-
using genetic programming, J. Geotech. Geoenviron. 132 (2006) 661–665. ton, Colorado, 1984.
[25] H.M. Azamathulla, A.A. Ghani, N.A. Zakaria, A. Guven, Genetic programming to pre- [37] B.C. Yen, Sediment Fall Velocity in Oscillating Flow, University of Virginia, Depart-
dict bridge pier scour, J. Hydraul. Eng. 136 (2009) 165–169. ment of Civil Engineering, 1992.
[26] H. Hakimzadeh, V. Nourani, A.B. Amini, Genetic programming simulation of dam
breach hydrograph and peak outflow discharge, J. Hydrol. Eng. (May 18 2013),
http://dx.doi.org/10.1061/(ASCE)HE.1943-5584.0000849 (Posted ahead of print).