You are on page 1of 15

Engineering with Computers

https://doi.org/10.1007/s00366-019-00813-1

ORIGINAL ARTICLE

Multi‑objective optimization of a hatchback rear end utilizing


fractional factorial design algorithm
Sajjad Beigmoradi1 · Mehrdad Vahdati2

Received: 4 March 2018 / Accepted: 22 June 2019


© Springer-Verlag London Ltd., part of Springer Nature 2019

Abstract
Air flow has significant effects on fuel consumption, performance, and comfort. Decreasing drag coefficient enhances fuel
consumption and vehicle performance. Moreover, omitting or reducing the power of aerodynamic noise sources provides
passengers comfort. In this paper, optimization of a hatchback rear end is conducted considering drag and aerodynamic noise
objectives. To this end, five geometrical parameters of the hatchback rear end are chosen as design variables in two levels.
Numerical simulation is applied to survey air flow features around the models in the wind tunnel. To reduce the number of
runs, fraction factorial design algorithm is applied to generate layout of the simulations which decreased the number of case
studies to half. Main and interaction effects of these factors on drag coefficient and acoustic power of the rear end source are
derived using analysis of variance. Optimum level for each parameter is chosen considering simultaneous drag and noise
goals. Finally, characteristics of air flow and acoustic power around optimum model are discussed.

Keywords Multi-objective optimization · Drag reduction · Acoustic power level · Fractional factorial design

1 Introduction sequentially the car motion. Then, they have compared and
reported the numerical results with experimental tests. Beig-
Air flow around running vehicles can be studied from differ- moradi et al. [3] have investigated optimum parameters of
ent aspects such as fuel consumption, performance, stability, a simplified car model considering aerodynamic drag and
and aerodynamic noise. For this matter, researchers are moti- noise objects. They have used numerical simulations beside
vated to continue their studies on air flow around realistic or genetic algorithm (GA) for this matter. They have discussed
generic car from the previous decades up to now. flow characteristics around optimum models.
Aljure et al. [1] have evaluated ability of different turbu- Corallo ea al. [4] have investigated the interaction
lent models to simulate air flow around two simplified cars. between the longitudinal c-pillar vortices the air flow over
They have surveyed the structure of flow in detail around the the rear slant surfaces for a simplified car model with differ-
cars for different LES models. Winkler et al. [2] have used ent aspect ratios. They have found out that not only aspect
coupled dynamic and CFD simulation to evaluate behav- ratio effects on drag coefficient, but increasing the aspect
ior of a simplified car model during a transient crosswind. ratio causes separation of the flow completely and variation
They have claimed that full dynamic coupling is required of c-pillar vortex strength.
for large yaw angles of the vehicle. They have stated that Hanfeng et al. [5] have investigated the influences of dif-
static coupling overestimates the aerodynamic forces and ferent deflectors on aerodynamic drag of a simplified car
model. They have installed deflectors with different heights
* Sajjad Beigmoradi and width sizes at side and leading edge of slant. They have
s.beigmoradi@email.kntu.ac.ir observed that deflector at leading edge of slant changes the
Mehrdad Vahdati near wake of model similar to the rear end of the model
vahdati@kntu.ac.ir with larger slant angle. Moreover, they have claimed that for
smaller slant angle of model installing, the deflector at lead-
1
Mechanical Engineering, K. N. Toosi University ing edge is more efficient in drag reduction. Thacker et al.
of Technology, Tehran, Iran
[6] have studied the effect of cancelation flow separating
2
Mechanical Engineering, Faculty of K. N. Toosi University on aerodynamic drag. They have considered two different
of Technology, Tehran, Iran

13
Vol.:(0123456789)
Engineering with Computers

simplified car models: (1) with sharp edge at connection have observed good agreement between their numerical
of roof and slant angle and (2) rounded edge at connection simulations with the experiment. They have concluded that
of roof and slant angle. They have indicated that omitting modified CRF changes flow structure around the vehicle
separation on rear slant angle due to rounding the edge considerably and causes drag reduction about 20%. Tunay
causes for weaker drag. Minelli et al. [7] have utilized an et al. [18] have studied the flow attributes downstream of a
active flow control method for drag reduction of car model. simplified car model by numerical simulation and experi-
They have investigated the positions of flow detachments mental test. They have said that, according to the charac-
the around their model, and then, they have used actuation teristics of air flow on the rear slanted surface and in near
devices to compare their effects with unactuated flow. They wake zone, there is a significant variation in very short space
have reported significant reduction of the side recirculation in both streamwise and vertical direction of flow. McNally
bubble by employing actuation devices. et al. [19] have examined some parameters of active flow
Khaled et al. [8] have investigated some parameters of control technique to reduce drag coefficient for a simpli-
a simplified car model on aerodynamic drag. They have fied flat back car model. They have observed that, utilizing
performed some experiments for this matter, and at the micro jets, rear wake of model can be modified which results
end, they have reported their solutions for drag reductions. for drag reduction. Beigmoradi [20] has obtained optimum
Grandemange et al. [9] have optimized the simplified square levels of the rear end parameters for a simplified car model
back of car model. They have found out that top and bottom considering aerodynamic drag and noise. Beigmoradi has
chamfer angle play a significant role on aerodynamic drag. applied computational fluid dynamics to survey air flow
Therefore, changing these chamfer angle, they achieved opti- structure around the models beside Taguchi algorithm to
mized shape in aspect of aerodynamic drag. Volpe et al. [10] find optimum level of variables.
have studied evaluation of side air force and yaw moment As studying aerodynamic performance is a time and
based on unsteady development of flow for a simplified car cost-consuming procedure for both test and simulation,
model. They have concluded that rear part the leeward flank researchers are looking for solutions to decrease these
has the most sensitivity to crosswind. They have stated that expenses. Applying simplified models or using optimiza-
even though modification of rear end of the model impacts tion algorithms are some keys for this problem. In this work,
on lateral coefficient, it does not have significant influence application of fractional factorial algorithm for aerodynamic
on transient duration and force overshoots. Salati et al. [11, assessment of a hatchback is introduced that can reduce the
12] have studied the effect of some add-on parts on aero- number of computations significantly.
dynamic drag phenomenon for a heavy truck vehicle. They
have established some add-on devices on the top and side
of trailer at the front and rear ends to survey which part has 2 Background theory
the most impact on drag reductions. Bello-Millan et al. [13]
have examined the effect of yaw angle on drag coefficient 2.1 Flow equations
for a simplified car body utilizing experimental tests. They
have stated that drag coefficient increases considerably by Air flow around ground vehicles is supposed to be an incom-
rising yaw angle in constant Reynolds number. Beigmoradi pressible, 3D and, turbulent. One of the most time and cost
and Ramezani [14] have investigated the effect rear window economic methodology to compute characteristic of the air
angle on the aerodynamic and acoustic parameters for a sim- flow in complex industrial problems is Reynolds Averaged
plified car model. They have achieved critical angle for rear Navier–Stokes (RANS) models. Variables of Navier–Stockes
window in aspect of drag and aero-acoustic noise. Keogh equation are decomposed into mean and fluctuating terms in
et al. [15] have simulated air flow around a simplified car Reynolds averaging approach. Therefore, vector and scaler
model during cornering maneuver in three different radial quantities can be written as ui = ū i + u�i and 𝜑 = 𝜑̄ + 𝜑� ,
conditions. They have stated that, at a constant Reynolds, respectively. Substituting these terms as variables into the
cornering with five car length radius increases drag coef- instantaneous continuity and momentum equations and
ficient in comparison with straight-line condition. Vahdati averaging time gives the ensemble-averaged continuity and
et al. [16] have optimized a hatchback car in aspect of drag momentum equations in form of expressions (1) and (2),
coefficient. They have utilized numerical method to investi- respectively:
gated aerodynamic characteristics around the model. They
𝜕𝜌 𝜕
have achieved optimum configuration for the car by employ- + (𝜌ui ) = 0, (1)
𝜕t 𝜕xi
ing fractional factorial algorithm. They have concluded that
this technique has a good ability to predict drag coefficient.
Kim et al. [17] have surveyed the effect of cab roof fairing
(CRF) on drag reduction for a heavy truck vehicle. They

13
Engineering with Computers

𝜕
(𝜌u ) +
𝜕
(𝜌ui uj )
2.2 Acoustic power equations
𝜕t i 𝜕xj
𝜕p
[ (
𝜕ui 𝜕uj 2 𝜕ul
)] Proudman [22] used Lighthill’s acoustic analogy to obtain
𝜕
=− + 𝜇 + − 𝛿 a formula for acoustic power generated by isotropic turbu-
𝜕xi 𝜕xj 𝜕xj 𝜕xi 3 ij 𝜕xl (2)
lence without mean of air flow. More recently, Lilley [23] re-
𝜕
+ (−𝜌u�i u�j ), derived the formula considering the retarded time difference,
𝜕xj which was neglected in Proudman’s original formula. Both
formulations yield acoustic power due to the unit volume of
where ui and xi are the velocity and displacement compo-
isotropic turbulence (in W/m) as
nents in the i = x, y, z directions, ρ is density, ρ is pressure,
𝛿ij is Kronecker delta, and 𝜇 is molecular viscosity. To solve ( )
u3 u5
Eq. (2) it is required to model unknown additional Reynolds PA = 𝛼𝜌0 , (5)
l a50
stress term (𝜌u′i u′j ). To this end, realizable k − ε is used for
modeling turbulence flow around the car model. Realizable
where u and l are the turbulence velocity and length scales,
model can fulfil mathematical margins on the normal stress,
correspondingly, and a0 is the speed of sound. In Eq. (5), 𝛼
regardless of turbulent flows physics. Shih et al. [21]
is a constant. In terms of turbulence kinetic energy, k , and
appealed that it also can deliver superior performance for
turbulence dissipation rate, 𝜀, Eq. (5) can be redrafted as
flows including rotation, boundary layers under strong
adverse pressure gradients, separation, and recirculation. PA = 𝛼𝜀 𝜌0 𝜀Mt5 , (6)
Transport equations for k and ε in the realizable k − ε are where
estimated as Eqs. (3) and (4), respectively:

2k
𝜕
(𝜌k) +
𝜕
(𝜌kui ) Mt = . (7)
𝜕t 𝜕x a0
[(i ) ]
=
𝜕
𝜇+
𝜇t 𝜕k (3) The rescaled constant, 𝛼𝜀, is set to 0.1 based on the cali-
𝜕xi 𝜎k 𝜕xi bration of Sarkar and Hussaini [24] in using direct numerical
+ Ck + Cb − 𝜌𝜀 − Qm + Sk , simulation of isotropic turbulence.
Acoustic power can also be stated in dB, which is calcu-
𝜕
(𝜌𝜀) +
𝜕
(𝜌𝜀ui ) lated from Eq. (8):
𝜕t 𝜕xi ( )
PA
(8)
�� � �
𝜕 𝜇 𝜕𝜀 LP = 10 log ,
= 𝜇+ t Pref
𝜕xi 𝜎𝜀 𝜕xi
(4)
𝜀2 w h e r e Pref i s t h e r e fe r e n c e a c o u s t i c p owe r
+ 𝜌D1 S𝜀 − 𝜌D2 √ ( Pref = 10−12 W/m3 ).
k + v𝜀
𝜀 The Proudman’s formula offers an approximate meas-
+ D1𝜀 D3𝜀 Cb + S𝜀 ,
k urement of the local contribution to total acoustic power
where per unit volume in a given turbulence field. Indeed, Proud-
man’s formula is a steady state model of broad band noise
source among other broad band models that predicts acous-
[ ]
𝜂 k

D1 = max 0.43, , 𝜂=S ,S= 2Sij Sij .
𝜂+5 𝜀 tic power level by applicable estimation, which is the most
common attitude in industrial fields reducing time and cost
In these equations, Ck denotes the creation of turbulence
of computations.
kinetic energy due to the mean velocity gradients. Cb is the
Power level is nominated in this work as a demonstrative
creation of turbulence kinetic energy due to buoyancy. Qm
of generated noise, since this response is the most conveni-
represents the share of the fluctuating dilatation in compress-
ent parameter to control noise sources in near-field stud-
ible turbulence to the overall dissipation rate. Sk and Sε are
ies. Decreasing power of sources yields in overall inner and
user-defined source terms. D2, D3𝜀, and D1ε are constants. σk
outer noise reduction, as the best and most effective solution
and σε are the turbulent Prandtl numbers for k and ε, respec-
in this regard.
tively. µt is eddy viscosity that for realizable model can be
computed from Shih et al. [21].

13
Engineering with Computers

2.3 Design of experiment method A half-one fraction for a problem with five design vari-
ables would be acquired by equating variable E to the ABCD
Design of experiment (DOE) method is a series of struc- interaction. This approach is depicted in Table 1.
tured tests or simulations on design factors to find out the The generator I = ABC…N is then solved for the missing
effect of each variable on output responses. The aim of this column (N) so that N = ABC…(N − 1) defines the product of
method is obtaining optimal configuration of the system and plus and minus signs to use in each row to generate the lev-
a formulated relationship between design factors and out- els for the nth factor. It should be noted that any interaction
put responses to predict the behavior of the system. There effect could be used to generate the column for the nth factor.
are some techniques in the field of DOE which choosing However, using any effect other than ABC…(N − 1) will not
among them depends on system characteristics, time and produce a design of the highest possible resolution. In Table 1,
cost restrictions, and designer proficiency. One of these tech- each word in treatment combination column expresses that
niques is fractional factorial design which can estimate the parameter is at high level, while other parameters at low level.
effect of main design factors and their low-order interactions For instance, abe is a response when A, B, and E variables are
by proper choosing of the runs. at their high level, while C and D parameters are at their low
In n design factors’ problem, that each factor has two level.
level, there is need for ­2n runs to survey the effect of After running all runs, statistical analysis should be per-
main and interactions of factors. As the number of factors formed to estimate the effect of each parameter and their inter-
increases, the number of runs that is required for a complete actions on output response. The effect of the interactions of N
factorial design, rapidly outgrows time and costs of runs. If parameters can be written as AB…N and defined by Eq. (9)
designer reasonably omit high-order interactions [25], infor- [25]
mation on the main variables and low-order interactions may
2
be achieved by running only a fraction of the complete fac- AB … N = (contrastAB…N ), (9)
2n
torial design. These fractional factorial designs are among
the most extensively applied types of designs for product where the contrast of AB…N is defined as
and process design, process enhancement, and industrial/
ContrastAB…N = (a ± 1)(b ± 1) … (n ± 1). (10)
business research. A one-half-fraction of the 2n design of the
highest resolution may be constructed by writing down a full To do analysis of variance (ANOVA) of runs, calculation of
factorial design including of the runs for a full 2n−1 factorial some terms is essential. These parameters are sum of square
and then adding the nth factor by identifying its high and of parameters, total sum of square, sum square of error, and
low levels with the plus and minus signs of the highest order mean square which are calculated from Eqs. (11) to (14),
interaction ABC…(N − 1). respectively:

Table 1  Configuration of runs Run A B C D E = ABCD Treatment


for one-half-fraction design combination
with five parameters response

1 − − − − + e
2 + − − − − a
3 − + − − − b
4 + + − − + abe
5 − − + − − c
6 + − + − + ace
7 − + + − + bce
8 + + + − − abc
9 − − − + − d
10 + − − + + ade
11 − + − + + bde
12 + + − + − abd
13 − − + + + cde
14 + − + + − acd
15 − + + + − bcd
16 + + + + + abcde

13
Engineering with Computers

1
SSAB…N = (contrastAB…N )2 , (11)
2n

k
�∑ �2
yi
(12)

SST = y2i − ,
i=1
DOFtotal


SSerror = SStotal − SSi , (13)

SSi
MSi = , (14)
DOF
where yi and DOF are output response and degree of free- Fig. 1  Configuration of rear end and its design variables
dom correspondingly. For p factors with two levels, DOF is
one for each parameter/interaction and 2p − 1 for total DOF.
Defining F = MS i and comparing with F0 [25], the effect
MS
error Table 2  Range of variations for design variables
of each parameter can be achieved.
Moreover, regression equation based on the main effects Spoiler Spoiler Fifth door Boat tail Diffuser
angle (𝛼 ) length (L) height (H) angle (𝛽 ) angle (𝜃 )
and low-order interaction is (°) (mm) (mm) (°) (°)
p p p
∑ ∑ ∑ Low level − 20 220 − 50 0 0
yi = 𝛽0 + 𝛽i xi + 𝛽ij xi xj . (15) High level 20 290 50 20 30
i=1 i=1 j=1

In Eq. (15), 𝛽i and 𝛽ij are regression coefficients which are


the half of the effects (Eq. 9) for each main effect or interac- 3.2 Mesh generation
tion. In addition, xi and yi are design variable i and output
response, respectively. Numerical simulation is known as the most powerful tool
in primary phase of the car development. It significantly
decreases the cost and time of product development by
reducing number of physical prototypes. In this paper,
3 Modeling and simulation CFD tool is applied to do numerical simulation of the air
flow around the car. ANSYS FLUENT R15 code, which
3.1 Case study is a boundary volume method base, is employed to solve
airflow around the model under RANS attitude. Compu-
In this research, enhancement of aerodynamic performance tational domain with 25, 091 mm × 4288 mm × 4288 mm is
for a facelifted C-segment hatchback car is surveyed. Vehi- determined as virtual wind tunnel and the car model is
cle length, height, width, and wheel base are, respectively, positioned in 5270 mm from the flow inlet. The blockage
4070 mm, 1760 mm, 1430 mm, and 2650 mm. Face lift is ratio of the models is less than 5.7%. 360,000 first-order
performed on rear end of the car to convert sedan to hatch- tria CFD mesh is employed to cover the surface of the
back. Freezing other panels and parts except rear end is car model with the size from 1 to 20 mm. Five boundary
determined as one of the study’s hard point. To this end, layers of Penta elements are produced near the car surface
five design variables of rear end are chosen to be optimized. to afford an accurate computation of the flow character-
These parameters are: (1) rear spoiler length (L); (2) rear istics near the wall by considering the y + value within
spoiler angle (𝛼 ); (3) fifth door height (H); (4) rear lamp boat an acceptable range (30–300) for k − 𝜀 turbulence model.
tail angle (𝛽 ); and (5) rear diffuser angle (𝜃). The range for Unstructured tetrahedral elements are placed between
variations of these parameters is determined by industrial boundary layers and structured hexahedral elements. Size
design department. Figure 1 and Table 2 depict design vari- boxes around the car and the rear end are created to con-
ables and their variation’s range, respectively. trol the size of elements in near the car model. Finer mesh
around and at the rear end of the car model is required to
increase accuracy of flow calculation around the model
and detecting wakes at this region. By moving away from

13
Engineering with Computers

Fig. 2  Position of car model in wind tunnel (up) and meshing around the model

the model, the size of hexahedral element increases by


growing ratio 1.2. Maximum length of elements in the size
box around the car is 30 mm, and in the rear end, size box
is 70 mm. Figure 2 shows the position of the car model in
computational domain and meshing around it.
Morphing technique is utilized with the aim of chang-
ing the geometric parameters of the rear end. This method
can be used to modify the geometry of the model without
significant variation in mesh quantity. To this end, dif-
ferent numbers of morph boxes are created around each
parameter to cover requirements for the modifications. A
total number of morph boxes are 60.

3.3 Boundary condition Fig. 3  Ahmed bluff body in different view (mm)

As the body is symmetry, half of car and tunnel are modeled


for the flow simulations to reduce time and cost of consump- formation of vortices. To evaluate aerodynamic noise around
tion. 38.89 m/s is set for input flow which is top speed of the car, broadband noise source model is activated.
the vehicle considering 0.025% of turbulent intensity and
viscosity ratio. Stationary wall with no motion is consid-
ered for the body and rotational moving walls are set for 4 Results and discussion
tires. Moreover, moving wall and no motion with zero shear
stress is defined for road and tunnel side walls correspond- 4.1 Validation of turbulent model
ingly. Density and viscosity of the air are 1.19 kg/m3 and
1.79 × 10−5 kg/m s, respectively. Velocity of sound in the Ahmed bluff body [26] (Fig. 3) is a generic car model that
air and reference acoustic power is considered 340 m/s and provides a good perspective for investigating aerodynamic
10−12 W . The k − 𝜀 realizable turbulence model is hired for behavior of hatchbacks. Variations of rear slant angle (𝛼 )
air simulations considering standard wall function, which deliver different drag coefficients that express importance
has an appropriate ability in estimating flow separation and of rear end of hatchbacks on aerodynamic performance.
To find out accuracy of the turbulence model, numerical

13
Engineering with Computers

simulations are conducted for Ahmed model in primary


phase and numerical results are compared with experi-
ments. Dimensions of the simulated wind tunnel were
6044 mm × 2740 mm × 1660 mm regarding to the experi-
mental setup data [27] and Ahmed’s model is placed at the
distance of 1 m from the air flow inlet.
It is observed that the computed drag coefficient using
k − 𝜀 realizable turbulence model with the standard wall
function is in good agreement with the experimental results,
as reported in Ref. [27]. The numerically obtained drag coef-
ficients are compared with experimental results [27] for
various backlight angles and are depicted in Fig. 4. Rela-
tive error of numerical and experimental is conducted from
Fig. 4  Comparison of numerically and experimentally obtained drag Eq. (16):
coefficients for different slant angles
|CDN − CDT |
Err = | (16)
| × 100.
CDN
Table 3  Difference between Backlight angle Difference (%)
numerically and experimentally (°)
drag coefficient
In Eq. (16), CDN and CDT are drag coefficient from numer-
0 1.1
ical simulations and experimental tests, respectively. The
5 1.2
results of difference analysis are illustrated in Table 3. The
10 2
results indicate that the average deviation between the simu-
15 6.7
lation and experimental results is 3.4%, which is acceptable
20 9.7
in aspect of engineering view.
25 2
30 3.3
4.2 Fractional factorial results
35 2.9
40 1.5
To obtain minimum drag and noise, five design factors are
selected for multi-objective optimization. Fractional fac-
torial design based on design of experiment is applied to

Table 4  Drag coefficient and maximum of acoustic power results for structured runs
Spoiler angle Spoiler length Door height Boat tail angle Diffuser angle Drag coefficient Max of acoustic
power level (dB)

− 20 220 − 50 0 30 0.282 38.6971


20 220 − 50 0 0 0.350 49.0479
− 20 290 − 50 0 0 0.296 38.6629
20 290 − 50 0 30 0.417 49.0860
− 20 220 50 0 0 0.303 43.8849
20 220 50 0 30 0.392 38.7264
− 20 290 50 0 30 0.277 43.8409
20 290 50 0 0 0.372 49.0473
− 20 220 − 50 20 0 0.300 49.0633
20 220 − 50 20 30 0.390 59.4926
− 20 290 − 50 20 30 0.287 43.9165
20 290 − 50 20 0 0.370 64.4807
− 20 220 50 20 30 0.285 38.7041
20 220 50 20 0 0.356 64.5245
− 20 290 50 20 0 0.295 43.8329
20 290 50 20 30 0.411 64.6029

13
Engineering with Computers

Fig. 5  Normal plot of the effects 


for drag coefficient (𝛼 = 0.05)

 A

 AE
E

AB
 B (IIHFW7\SH

Percent
 1RW6LJQLILFDQW
 6LJQLILFDQW

 )DFWRU 1DPH
 $ 6SRLOHU$QJOH
% 6SRLOHU/HQJWK
 & WK'RRU+HLJKW
' %RDW7DLO$QJOH

( 'LIIXVHU$QJOH


         
Effect

minimum time and cost of computations. The best layout drag coefficient. On the other hand, according to the analy-
in this approach is when the effect of main factors and their sis results, boat tailing of rear lamp, changing of fifth door
low-order interactions can be evaluated, while they are not height, and their interactions do not have great influence on
aliased with each other. Therefore, half-fraction with resolu- drag coefficient.
tion V meets this condition for a problem with five design One the most advantages of fractional factorial design is
factors [25]. Table 4 demonstrates the configuration of the simultaneous survey of the effect of main factors and their
16 runs for designed simulations and their corresponding interactions. According to ANOVA, spoiler angle is one of
results for drag and noise. In the next step, analysis of vari- the factors, which has major impact on aerodynamic drag.
ance (ANOVA) on the results is performed to achieve the According to Figs. 6 and 7, setting spoiler angle in low level
influence of each factor on drag coefficient and acoustic delivers minimum drag. For spoiler length parameter, even
power. though Fig. 6 displays that choosing this factor at low level
Normal plot of the factors effect on drag is shown in gives low drag, spoiler angle and length interaction in Fig. 7
Fig. 5. This plot provides a good approximation to find key discards this point. Regarding to Fig. 7, considering spoiler
factors in unreplicated problems [28]. In this pattern, the length at high level decreases drag coefficient. In this situ-
effective parameters do not fall in the vicinity of the line. ation, interaction effect has superiority to main effect plot.
Regarding to Fig. 5, spoiler angle, spoiler length, diffuser Regarding to Fig. 7, variation of drag in diffuser–spoiler
angle, interactions between spoiler angle–diffuser angle, angle interaction plot is significant and setting diffuser
and also spoiler angle–spoiler length have major effect on angle at high level creates smaller drag coefficient. There is

Spoiler Angle Spoiler Length Door Height Boat Tail Angle Diffuser Angle
0.38

0.36
Mean of Drag

0.34

0.32

0.30

-20 20 220 290 -50 50 0 20 0 30

Fig. 6  Main effects’ plot for drag

13
Engineering with Computers

Spoiler Angl * Spoiler Leng Spoiler


0.40
Length
220
0.35
290
0.30
Spoiler Angl * Door Height Spoiler Leng * Door Height Door
0.40
Height
0.35 -50
Mean of Drag

50
0.30
Spoiler Angl * Boat Tail An Spoiler Leng * Boat Tail An Door Height * Boat Tail An Boat Tail
0.40
Angle
0.35 0
20
0.30
Spoiler Angl * Diffuser Ang Spoiler Leng * Diffuser Ang Door Height * Diffuser Ang Boat Tail An * Diffuser Ang Diffuser
0.40
Angle
0
0.35

0.30
-20 20 220 290 -50 50 0 20
Spoiler Angle Spoiler Length Door Height B oat Tail Angle

Fig. 7  Interaction plot for drag (fitted means)

a conflict for the trend of plot in main and interaction figures, model with key terms. Regression model that is obtained
but priority of interaction effect to main effect caused for from modified ANOVA is written as
choosing diffuser angle at top level. Regarding to Fig. 6, fifth
CD = 0.30025 − (0.000605 × 𝛼) + (0.000118 × L) + (0.000414 × 𝜃)
door and boat tail angle do not have considerable influence
+ (0.000009 × 𝛼 × L) + (0.000047 × 𝛼 × 𝜃).
on drag coefficient, but in aspect of their interactions with
(17)
spoiler angle (Fig. 7), choosing fifth door height and boat
Equation (17) includes the crucial terms of main effect
tail angle at top and low level, respectively, generates lower
and their interactions. As it is seen in this equation, coef-
drag coefficient.
ficient of spoiler angle has maximum value in comparison
After estimation of the major factors, modifying model
with other coefficients. In can be interpreted that variation of
is performed to obtain regression model according to cru-
spoiler angle has the most important role on drag coefficient
cial terms. To this end, unimportant terms are omitted from
and interaction of spoiler and diffuser angle ranks second.
ANOVA and the analysis is repeated to obtain regression
It shows that to have minimum drag, spoiler angle should

Fig. 8  Normal plot of the effects 


for acoustic power (𝛼 = 0.05)
 A

 D
AD

BE

Percent

 (IIHFW7\SH
 1RW6LJQLILFDQW
 6LJQLILFDQW

)DFWRU 1DPH

$ 6SRLOHU$QJOH
 % 6SRLOHU/HQJWK
& 'RRU+HLJKW
 E ' %RDW7DLO$QJOH
( 'LIIXVHU$QJOH

       
Effect

13
Engineering with Computers

take its lowest value. For acoustic power response, as it is effects of the spoiler angle and considering both plots shows
shown in Fig. 8, the effect of two interactions is significant: that positioning this factor at low level generates acoustic
(1) spoiler and boat tail angle and (2) spoiler length and sources with smaller power.
diffuser angle. According to Fig. 8, it seems that only door According to Figs. 8 and 9, second important factor
height and its interactions do not have major effect on aero- in aspect of aerodynamic noise is boat tail angle. Paying
dynamic noise. Considering distance of the spoiler angle attention to Figs. 9 and 10, positioning boat tail angle in
effect from the line, it can be estimated that not only this lower level creates smaller acoustic power on the rear end
factor has the most contribution on drag, but it also plays a of the model. Regarding to Figs. 8 and 10, interaction of
significant role on increasing the power of acoustic sources the spoiler and boat tail angles takes third rank in aspect
at rear end of the car. of the importance on acoustic source power. Results from
Studying the plot of main effects confirms that spoiler Fig. 10 approve that positioning spoiler and boat tail angle
angle has the most variation on maximum acoustic power at lower level reduces aerodynamic noise. For spoiler length
during movement from low level to high level. Regarding factor, positioning this factor at its low level generates lower
to Fig. 10, there is no conflict between main and interaction aerodynamic noise. Therefore, it is a confliction to choose

Spoiler Angle Spoiler Length Door Height Boat Tail Angle Diffuser Angle

54
Mean of Noise

51

48

45

42
-20 20 220 290 -50 50 0 20 0 30

Fig. 9  Main effects’ plot for acoustic power

Spoiler Angl * Spoiler Leng Spoiler


 Length
220
 290


Spoiler Angl * Door Height Spoiler Leng * Door Height Door
 Height
-50
Mean of Noise

 50


Spoiler Angl * Boat Tail An Spoiler Leng * Boat Tail An Door Height * Boat Tail An Boat Tail
 Angle
0
 20


Spoiler Angl * Diffuser Ang Spoiler Leng * Diffuser Ang Door Height * Diffuser Ang Boat Tail An * Diffuser Ang Diffuser
 Angle
0
 30


       
Spoiler Angle Spoiler Length Door Height B oat Tail Angle

Fig. 10  Interaction plot for acoustic power (fitted means)

13
Engineering with Computers

optimal level considering both drag and noise objectives. both drag and noise objectives is required to obtain an
Interaction of spoiler length and diffuser angle shows that optimal configuration of the rear end. To this end, desir-
locating this factor at upper level is desirable in aspect of ability function for each objective is calculated from
drag and noise. Even though door height does not have regression model, and then, maximum value for compos-
major impact on the acoustic power, considering this factor ite desirability function is achieved employing reduced
in combination of spoiler length and angle, choosing upper gradient algorithm. As both objectives are supposed to
level for this factor can deliver better results in aspect of the have same significance, the weight for computing com-
aero-acoustic attribute. position desirability function is considered equal for drag
After identification of the significant factors, revising and lift. Considering 0.1 weight for both drag and noise
model is conducted to obtain regression model according individual desirability function gives 1, 0.997 and 0.999
to key terms. To this end, minor factors are omitted from for noise, drag and combined desirability values that
ANOVA and the analysis is repeated to obtain regression show proposed optimal values meet regression models
model with significant terms. Objective function for acoustic perfectly. Predicted drag and maximum acoustic power
power level that is acquired from modified ANOVA is writ- for the optimum model is 0.281 and 38.66 dB accord-
ten as Eq. (18): ingly that occurs in a model with − 20°, 239 mm, 50 mm,
Noise = 48.726 + (6.150 × 𝛼) + (4.852 × 𝛽)
0°, and 30° for spoiler angle, spoiler length, door height,
(18) boat tail angle, and diffuser angle, respectively (Fig. 11).
− (1.592 × 𝜃) + (3.548 × 𝛼 × 𝛽) + (2.270 × L × 𝜃).

4.3 Multi‑objective optimization results 4.4 Air flow pattern around optimal model

According to the previous section, setting spoiler length To validate predicted drag and noise results for optimal
at upper level even though delivers minimum drag force, model, numerical simulation for the proposed configuration
but this position for spoiler boosts power of acoustic is performed. According to simulation, it is observed that
sources at rear end. Therefore, compromising between drag and maximum of acoustic power level at rear end for

Fig. 11  Optimization results’ plot

13
Engineering with Computers

Fig. 12  Pressure around the spoiler (pascal)

Pressure magnitude on symmetry line of rear spoiler


for minimal drag and optimal case is depicted in Fig. 13.
Even though, for the model with minimum drag, the air
pressure at vicinity of spoiler junction with roof is nega-
tive, but it has bigger value than optimal model. Moving
toward the tip of the spoiler in this case, pressure grows
up and close to the optimum model. Regrading to Fig. 13,
pressure magnitude for minimum drag case is about 233 Pa
which is a little smaller than optimum model with 250 Pa.
It shows that longer spoiler is helpful for streamlining air
flow at the rear end, and as a result, it can be applied to
decrease drag force. Moreover, pressure variation for a
model with minimum drag is about 594 that is smaller than
optimal model with 748 Pa. As the source of these pres-
sure variation is the length of spoiler, it can be concluded
that for a constant angle, increasing the length of spoiler
decreases pressure variation alongside of the spoiler.
Fig. 13  Pressure distribution along spoiler Simulation results shows that variation of diffuser angle
changes the velocity pattern at the rear end. It can be seen
that locating diffuser angle at upper level besides the spoiler
the model are 0.280 and 38.7 dB, respectively. It can be can angle generates a wake at the lower area of the rear end.
calculated that numerical simulation results and predicted Regarding to Fig. 14, posing diffuser angle at lower level
values from optimization have about 0.2% and 0.1% devia- for minimum noise destroys the velocity pattern at the rear
tion based on numerical simulation for drag and noise that end, which is formed at minimum drag cases and impacts
is negligible. By studying air flow around optimum model, on the bigger area at this zone. The size of the affected zone
it is seen that pressure gradient has the most contribution to at the rear end of minimum noise case can be identified the
total drag. For the optimal model, pressure and viscous drag source of increasing drag in comparison with optimal case.
are 0.256 and 0.024, respectively, that indicate more than Moreover, minimum drag and noise cases in Fig. 14 also
90% of total drag caused by pressure. Optimization results show that setting rear door at different heights does not show
show that rear spoiler plays an important role in drag and lift major change on the velocity pattern around the rear window
coefficient and it can be concluded that this part has consid- or separation of air flow at the rear end. This point agrees
erable effect on pressure gradient of the rear end. Regarding with ANOVA results that showed rear door height and does
to Fig. 12, negative pressure zone for minimum drag case is not have significant effect on drag and noise objectives.
smaller in comparison with optimal case. It seems that varia- According to Fig. 14, separation of air flow occurs at the
tions of pressure alongside spoiler for this case are smoother tip of the rear spoiler. For optimum case, which has shorter
due to longer spoiler. spoiler, separation happens earlier in comparison with two

13
Engineering with Computers

Fig. 14  Velocity contour (m/s) around the car with: minimum drag (up); minimum noise (middle); optimal model (down)

other cases and this causes for relatively bigger wake size at and direction of vortices are more complex than minimum
the rear end in comparison with minimum drag case. In addi- drag car. In this case, two x-axis beside one y-axis vortices
tion, it is concluded that even though the size of the wake are generated at the rear end. According to Fig. 15, y-axis
at the rear end can be changed by variation of the spoiler vortex is circulating on the rear window. Therefore, it can be
length, but the form of vortex is fairly similar. concluded that on rainy days, driver has limitations in rear
Not only rear end factors can change the size of the wakes, view due to circulate droplets on the window.
but they also can generate different shapes of vortices. As it In this study, maximum acoustic power level on the rear
is shown in Fig. 15, for the car model with minimum drag, window is attempted to be minimize for two main reasons:
two y-axis vortexes are created in upper and bottom regions (1) rear glass has minimum distance from rear end to the
of the rear end. The size of bottom vortex is bigger in com- passenger’s ear and (2) treating airborne noise path from
parison with the upper one. It shows that the effect of lower outer surface of rear glass to the cabin is an expensive
vortex in aerodynamic drag is more than upper vortex. The cumbersome task. Therefore, minimizing acoustic power
configuration of the vortices for optimal model is similar of the source can be conducted as an effective solution for
to minimum drag, but the size of the upper vortex is bigger this problem. Acoustic power level around optimal model
in comparison with the corresponding wake in minimum is shown in Fig. 16. As it is depicted, acoustic power-
drag model. For the model with minimum noise, the shape level contours at the rear end have maximum values at the

13
Engineering with Computers

Fig. 15  Streamlines based on velocity (m/s) around the car with minimum drag (left); minimum lift (middle); optimal model (right)

5 Conclusion

In this paper, optimal shape for rear end of a hatchback


model is obtained considering aerodynamic drag and noise
objectives. Fractional factorial design based on design of
experiment is chosen to make different configurations of
simulation. Contribution of five design factors and their
interaction on drag and noise is surveyed and regression
equation for each object is derived. Multi-objective opti-
mization based on maximizing composite desirability
function is conducted and their results are reported. It is
concluded that spoiler angle has the most crucial role in
Fig. 16  Acoustic power level (dB)
drag and noise. Interaction between spoiler and diffuser
angle ranked second important factor for drag coefficient,
but for noise boat tailing, the rear lamp has significant
bottom of window. Therefore, any defect in the assembly effect. Simulations show how spoiler length effects on aer-
of the glass to body structure causes for transferring aero- odynamic performance by changing pressure variation at
dynamic noise to the cabin. Moreover, rear wiper is placed the rear end zone. It is observed that event though spoiler
in the region with maximum acoustic power. Therefore, length has influence on vortex size, but the shapes of the
any leakage around this component facilities passing noise rear vortex are constant due to change of spoiler length.
to the cabin. Maximum of acoustic power level, which is Moreover, acoustic source at rear end due to air flow is
displayed in Fig. 16, is about 38.7 dB which is a little more identified. It is estimated that since acoustic power near the
than the case with minimum noise. However, in both cases, wiper position is maximum, any leakage in connection of
maximum of acoustic power occurs in the same place at rear wiper to the glass transfers airborne noise to the cabin.
the bottom region of the rear window. In upper section of
the rear window, acoustic power level has its minimum Acknowledgements The authors would like to express their gratitude
value due to the lowest flow velocity. to Automotive Industries Research & Innovation Center (AIRIC) of
SAIPA, for valuable support.

13
Engineering with Computers

References 14. Beigmoradi S, Ramezani A (2013) Effect of the backlight


angle on the aero-acoustics of the c-pillar. Int Rev Model Simul
6(3):988–993
1. Aljure DE, Lehmkuhl O, Rodriguez I, Oliva A (2014) Flow
15. Keogh J, Barber T, Diasinos S, Doig G (2016) The aerodynamic
and turbulent structures around simplified car models. Comput
effects on a cornering Ahmed body. J Wind Eng Ind Aerodyn
Fluids 96:122–135
154:34–46
2. Winkler N, Drugge L, Trigell AS, Efraimsson G (2016) Cou-
16. Vahdati M, Beigmoradi S, Batooei A (2018) Minimising drag
pling aerodynamics to vehicle dynamics in transient crosswinds
coefficient of a hatchback car utilising fractional factorial design
including a driver model. Comput Fluids 138:26–34
algorithm. Eur J Comput Mech 27(4):322–341
3. Beigmoradi S, Hajabdollahi H, Ramezani A (2014) Multi-
17. Kim JJ, Lee S, Kim M, You D, Lee SJ (2017) Salient drag reduc-
objective aero acoustic optimization of rear end in a simplified
tion of a heavy vehicle using modified cab-roof fairings. J Wind
car model by using hybrid robust parameter design, artificial
Eng Ind Aerodyn 164:138–151
neural networks and genetic algorithm methods. Comput Fluids
18. Tunay T, Yaniktepe B, Sahin B (2016) Computational and experi-
90:123–132
mental investigations of the vortical flow structures in the near
4. Corallo M, Sheridan J, Thompson MC (2015) Effect of aspect
wake region downstream of the Ahmed vehicle model. J Wind
ratio on the near-wake flow structure of an Ahmed body. J Wind
Eng Ind Aerodyn 159:48–64
Eng Ind Aerodyn 147:95–103
19. McNally J, Fernandez E, Robertson G, Kumar R, Taira K, Alvi F,
5. Hanfeng W, Yu Z, Chao Z, Xuhui H (2016) Aerodynamic drag
Yamaguchi Y, Murayama K (2015) Drag reduction on a flat-back
reduction of an Ahmed body based on deflectors. J Wind Eng Ind
ground vehicle with active flow control. J Wind Eng Ind Aerodyn
Aerodyn 148:34–44
145:292–303
6. Thacker A, Aubrun S, Leroy A, Devinant P (2012) Effects of sup-
20. Beigmoradi S (2015) Aerodynamic drag and noise minimization
pressing the 3D separation on the rear slant on the flow structures
of rear end parameters in a simplified car model utilizing robust
around an Ahmed body. J Wind Eng Ind Aerodyn 107:237–243
parameter design method. SAE technical paper 2015-01-1360
7. Minelli G, Hartono EA, Chernoray V, Hjelm L, Krajnović S
21. Shih TH, Liou WW, Shabbir A, Yang Z, Zhu J (1995) A new k − ε
(2017) Aerodynamic flow control for a generic truck cabin using
Eddy-viscosity model for high Reynolds number turbulent flows—
synthetic jets. J Wind Eng Ind Aerodyn 168:81–90
model development and validation. Comput Fluids 24(3):227–238
8. Khaled M, El Hage H, Harambat F, Peerhossaini H (2012) Some
22. Proudman I (1952) The generation of noise by isotropic turbu-
innovative concepts for car drag reduction: a parametric analysis
lence. Proc R Soc Lond Ser A Math Phys Sci 214(1116):119–132
of aerodynamic forces on a simplified body. J Wind Eng Ind Aero-
23. Lilley GM (1993) The radiated noise from isotropic turbulence
dyn 107:36–47
revisited, NASA contract report 93-75. NASA Langley Research
9. Grandemange M, Cadot O, Courbois A, Herbert V, Ricot D, Ruiz
Center, Hampton
T, Vigneron R (2015) A study of wake effects on the drag of
24. Sarkar S, Hussaini MY (1993) Computation of the sound gener-
Ahmed’s squareback model at the industrial scale. J Wind Eng
ated by isotropic turbulence. DTIC document
Ind Aerodyn 145:282–291
25. Montgomery D (2012) Design and analysis of experiments, 8th
10. Volpe R, Ferrand V, Da Silva A, Le Moyne L (2014) Forces and
edn. Wiley, Hoboken, New Jersey. www.wiley​.com/go/permis​ sions​
flow structures evolution on a car body in a sudden crosswind. J
26. Ahmed SR, Ramm G (1984) Salient features of the time-averaged
Wind Eng Ind Aerodyn 128:114–125
ground vehicle wake. SAE technical paper, p 840300
11. Salati L, Schito P, Cheli F (2017) Wind tunnel experiment on a
27. Strachan R, Knowles K, Lawson N (2007) The vortex structure
heavy truck equipped with front-rear trailer device. J Wind Eng
behind an Ahmed reference model in the presence of a moving
Ind Aerodyn 171:101–109
ground plane. Exp Fluids 42(5):659–669
12. Salati L, Cheli F, Schito P (2015) Heavy truck drag reduction
28. Lenth RV (1989) Quick and easy analysis of unreplicated factori-
obtained from devices installed on the trailer. SAE Int J Commer
als. Technometrics 31:469–473
Veh 8(2015-01-2898):747–760
13. Bello-Millán FJ, Mäkelä T, Parras L, del Pino C, Ferrera C (2016)
Publisher’s Note Springer Nature remains neutral with regard to
Experimental study on Ahmed’s body drag coefficient for different
jurisdictional claims in published maps and institutional affiliations.
yaw angles. J Wind Eng Ind Aerodyn 157:140–144

13

You might also like