Advances in Engineering Software 123 (2018) 38–50

Contents lists available at ScienceDirect

Advances in Engineering Software

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

Hull form design optimization of twin-skeg fishing vessel for minimum T

resistance based on surrogate model
⁎
Lin Yan, He Jingyi, Li Kai
School of Naval Architecture and Ocean Engineering, Dalian University of Technology, China

A R T I C LE I N FO A B S T R A C T

Keywords: Twin-skeg ship has better hydrodynamic performances than regular ship, however, it is still difficult to obtain an
Twin-skeg accurate relationship between skeg design and overall hydrodynamic performances. Resistance optimization is
Resistance optimization the major concern of developing twin-skeg ship. This paper proposes a combined approach for hull form design
CFD optimization of twin-skeg ship by using computational fluid dynamics (CFD) calculation and surrogate model.
Kriging modeling method
Main design parameters of skeg geometry and arrangement could be determined from the design domain by
Multi-objective optimization
using the proposed method. Parametric modeling technology is adopted for performing design evaluations in an
automatic manner with different design parameter combinations. A twin-skeg fishing vessel is selected as re-
search object. In the proposed method, the sample set for constructing surrogate models is generated by using
Optimal Latin Hypercube Sampling (OLHS) method, the corresponding responses are calculated through CFD
simulations, and then the surrogate models are constructed by using Kriging modeling method, which represent
the mathematical relationship between input design variables (skeg shape design variables) and output objective
functions (resistance values under four different working conditions). The functional analysis of variance
(ANOVA) is performed to investigate how much influence the design variables have on the objective functions.
Finally, a multi-objective evolutionary algorithm (NSGA-II) is used to obtain the optimal solution, which shows
5.4% average decrease in the total resistance than the original design. The CFD calculation results of the optimal
solution show that the proposed method can achieve minimum resistance design with high accuracy and low
time cost.

1. Introduction 1.1. Good resistance performance

Generally, the draft of fishing vessel is usually restricted due to the
shallow water of fishing port, which means the ratio of breadth to
draft should be high to obtain enough capacity and get better man-
euverability. And fishing vessel also is a kind of medium high-speed
ship with high fuel consumption. To lower the fuel consumption for
good energy saving and CO2 emission reduction, it is necessary to
design a fishing vessel type suitable for entering shallow water fishing
ports and carrying out tasks in the distant sea with good resistance
performance. Twin-skeg hull form is considered as one of the green
ship solutions by Swedish State Shipbuilding Experimental Tank
(SSPA) for ships with low ratio of length to breadth, high ratio of
breadth to draft, restricted draft or heavy loaded propellers [1,2].
Currently, twin-skeg hull form is mainly used in large ships such as
LNG carriers and container ships, which can benefit from irreplaceable
advantages:
The skeg can be considered as slender body which has excellent
resistance performance. In the twin-skeg hull form, the longitudinal
gradient of stern is small so that the boundary layer separation can be
controlled and the viscous pressure resistance can be reduced. The
central tunnel formed between the two skegs is smooth so that the run
body can be shortened, the maximum transverse section can move
backwards 10%–15% ship length, and then the wave resistance is
lower. Usually the total resistance of twin-skeg ship is less than con-
ventional twin-screw ship, even less than some single-screw ship
sometimes.
1.2. High propulsion efficiency
It can be proved by model test and numerical simulation that with
limited draught, the screw propulsion efficiency of twin-skeg twin-
screw ship is higher than that of conventional single-screw ship. That is
because in twin-screw ship, the load of each propeller is reduced, which

⁎
Corresponding author.
E-mail address: likai@dlut.edu.cn (K. Li).

https://doi.org/10.1016/j.advengsoft.2018.05.010
Received 27 February 2018; Received in revised form 26 April 2018; Accepted 20 May 2018
0965-9978/ © 2018 Elsevier Ltd. All rights reserved.
Y. Lin et al. Advances in Engineering Software 123 (2018) 38–50

is good for improving the open-water efficiency. And twin-skeg twin- Start
screw ship can make full use of the wake field, which can enhance the
hull efficiency by circumference wake.
Determine the design variables
1.3. Good usability and the ranges of stern skeg

The smaller propeller of twin-skeg twin-screw ship can achieve

Design of experiments
better cavitation and vibration performance. And the configuration of
(Selection of sample points)
twin-screw twin-rudder can also improve maneuverability, safety and
seakeeping performance.
Recent researches on twin-skeg ship hull form are mainly based on Perform CFD numerical
CFD numerical simulations. Sakamoto et al. conducted simulations of simulations at sample points
resistance and self-propulsion of a twin-skeg container ship for the se- under four working conditions
lection of the optimal location of propeller with electronic propulsion
system [1]. But the target of their research is not the optimization of the
shape of skeg. Kim et al. presented a study on optimization of main Construct surrogate models
dimensions and hull parameters of a twin-skeg LNG carrier based on the (using Kriging modeling method)
Increase test
SSPA's parametric studies, CFD calculations and model tests, and
achieved 13% reduction in fuel consumption comparing to a single-
sample points
Perform accuracy evaluation
screw design [2]. Their study is on the basis of abundant model test
of the surrogate models
data of SSPA, which is of great significance. However, when there is no
test data or lack of test conditions, a new approach is needed for the
optimization of the skeg form. Park et al. established an assess standard N
for the propulsion efficiency of twin-skeg ship according to the char- Accuracy satisfied?
acteristics study of its stern flow [3]. But only vertical inclination angle
and the distance between the two skegs, were used to express the skeg Y
shape and geometry arrangement. Chen et al. constructed a fully Proceed the functional
parameterized model of a 10000TEU twin-skeg container ship and analysis of variance
obtained 8.6% reduction in the total resistance by optimization of hy-
drodynamic performance [4]. In their study, vertical inclination angle
and distance between skegs were also used in the optimization process, Perform multi-objective
which cannot fully express the geometry features of the skegs. To in- optimization algorithm
Change optimization
vestigate the influence of skeg shape and geometry arrangement on the
algorithm
total resistance, the design and optimization parameters should fully
Algorithm N
express the skeg shape and geometry arrangement and therefore should
include vertical inclination angle, horizontal distance, length of the converges?
skeg slope, starting position and the fullness of the skeg. Y
It is well accepted that carrying out the research on the influence of
Obtain optimal
skeg shape and geometry arrangement on the total resistance of twin-
design solution
skeg ship is necessary. In this paper, a twin-skeg fishing vessel is se-
lected for minimizing the total resistance RT by optimization of skeg
parameters to get lower power and lower fuel consumption. The opti- End
mization of stern shape and geometry arrangement is based on CFD
calculations. To estimate the effective power curve, four different Fig. 1. The optimization procedure of the skeg shape and geometry arrange-
ment.
speeds are selected in the resistance calculation to assess the compre-
hensive resistance performance of one specified hull form design.
However, the numerical simulations between optimization iterations location and fidelity level of the next evaluation by maximizing an
are very time-consuming and impractical for this complex engineering augmented expected improvement function [9]. Meanwhile, Martin
problem. To alleviate the computational burden of the CFD calcula- carried out the study on the parameters of Kriging model [10]. Fur-
tions, Kriging modeling method is introduced in this study to substitute thermore, Kleijnen presented a detailed review of Kriging modeling
the true responses by four surrogate models for its accuracy and ro- method [11]. Kriging modeling method has been widely applied to
bustness with small data sets [5]. Based on the accuracy validation of engineering design optimization problems since Sacks et al. introduced
surrogate models, a multi-objective optimization algorithm is in- it into structural reliability problems [12]. Kriging modeling method
troduced to search for the optimal shape design solution with the best has also been actively used to aircraft and aerospace field. Jeong et al.
estimation of resistance performance. combined Kriging modeling method and genetic algorithm (GA) to a 2-
The Kriging modeling method, which is named after the South D airfoil design problem to optimize the lift-to-drag ratio [8]. Laur-
African geological engineer Krige DG, is mainly used for the determi- enceau and Sagaut first compared the global accuracy of different
nistic optimization problems by predicting the value of unknown point surrogate modeling strategies while the technique of sampling method
using stochastic processes [6–8], and it is the best choice to approx- varies, and verified the analysis by optimizing the shape of a transonic
imate highly nonlinear response functions when the number of variable airfoil [13]. In the automotive field, Lee and Kang, Chen et al. and Gao
dimensions is moderate [5]. To be noted that the reason why the study et al. conducted their researches on Kriging-based optimization of au-
is a deterministic optimization problem is that unlike actual physical tomotive door's structure, ride comfort and crashworthiness respec-
experiments, the results of CFD calculations stay the same when re- tively [14–16].
peating the calculation without any random error. Other relevant works The overall flowchart of the optimization process is presented in
on Kriging modeling study, such as multiple-fidelity sequential Kriging Fig. 1. The rest of this paper is organized as follows. In Section 2, after a
optimization (MFSKO) which use multiple-fidelity data and select the parameter study on the skeg of a twin-skeg fishing vessel, design of

39
Y. Lin et al. Advances in Engineering Software 123 (2018) 38–50

Table 1 - θS: vertical inclination angle of the skeg (the value is positive when
Main dimensions of hull. the skeg is outward-inclined, otherwise it is negative);
Symbol Main dimensions and hull coefficients Unit Value - lS: length of the skeg slope in x direction;
- xS: x-coordinate of skeg ending position;
LPP Length between perpendiculars m 78 - h: z-coordinate of skeg ending position;
B Breadth m 14.8
- BS: breadth between the center line of shaft and the farthest point of
D Depth m 6.35
T Draft m 2.9
skeg in y direction.
∇ Displacement m3 2275.57
DWT Design dead weight t 1400 Fig. 3 shows the detailed geometry information of these parameters.
Swet Area of the wetted surface m2 1266.61 And the values of parameters in the original design are listed in Table 2.
LPP/B Ratio of length between perpendiculars to breadth – 5.27
For observe the influence brought by changing parameters clearly,
B/T Ratio of breadth to draft – 5.103
CB Block coefficient, CB = Weight/(ρwater · LPP · B · T) – 0.68 the design parameters should be dimensionless. Therefore, the design
V Design navigating Speed knot 11.3 variables consist of one dimensional variable θS and five dimensionless
variables. The ranges of the six design variables are determined to avoid
unrealistic geometry, which are shown in Table 3.
experiment (DOE) is carried out, then the numerical simulation method
is introduced. Section 3, according to the data obtained in Section 2,
2.2. Deign of experiment
constructs and validates the surrogate models on the basis of Kriging
modeling method. Section 4 optimizes the surrogate functions of
The selection of sample points is very important in constructing
Section 3 and analyses the sensitivity of design variables to the total
surrogate model with certain accuracy, and the sample points should
resistance. Section 5 presents conclusions and topics for future research.
uniformly spread over the hypercube design space. Latin Hypercube
Sampling (LHS) method is often used to generate the sample points for
2. Parameter study for optimization and CFD calculation
Kriging model [11]. In this study, the Optimal Latin Hypercube Sam-
pling (OLHS) method is used to obtain the sample points due to that the
2.1. Optimization parameter study
points generated by using OLHS method is more uniformly distributed
than that of LHS method [17]. For the six design variables, 54 points are
A twin-skeg fishing vessel is selected as research object, whose main
created by using OLHS method. The original design point is also added
dimensions of the original design are shown in Table 1. Fig. 2 shows the
into the training sample set for the construction of surrogate models.
3D model of hull surface of the original design. The coordinate system
To prepare for the CFD calculation, different hull geometries are
of modeling is right handed, the origin point is the intersection point of
obtained through parameterized modeling by varying the design vari-
the 0th station plane, central longitudinal profile and base plane, the
ables of skeg according to the sample set.
positive direction of axis X is forward in the longitudinal direction, the
positive direction of axis Y is port side direction and the positive di-
rection of axis Z is upward. 2.3. CFD calculation configuration
To fully express the geometry shape and the arrangement of skegs,
the optimal design parameters should cover the aspects of starting and The coordinate system of CFD calculation is also right handed, but
ending position, spatial features surface, and fullness. The distance the origin point is the intersection point of the mid-section plane,
between two skegs, the height and x-coordinate of the shaft outlet can central longitudinal profile and undisturbed waterline surface. In order
determine the ending position of a skeg. The vertical inclination angle to reduce computing time, CFD calculation should be carried out by
can express the skeg spatial feature. The length of skeg can determine scaled model. The scaled model should have the same Fr as the real
the starting position of a skeg with the ending position. The maximum ship.
half width of skeg can express the fullness. These parameters are V
summarized as follows: Fr =
g ·LPP (1)
- b: distance between skegs in y direction; where V is the navigational speed (m/s), g is the acceleration of gravity

Fig. 2. Perspective view of hull geometry for original design.

40
Y. Lin et al. Advances in Engineering Software 123 (2018) 38–50

Fig. 3. Six geometry parameters of the skeg.

Table 2 To obtain effective power curve of ship, the resistance values under
Skeg parameters of the original design. different speeds should be calculated. In this paper, the pressure re-
Symbol Unit Value
sistance RP, shear resistance RS and total resistance RT under four
speeds are considered. The configuration of calculating speeds is shown
b m 6.6 in Table 4.
θS ○ 0 Trimmed cell mesher is used to mesh the calculation domain and
lS m 22.62
prism layer mesher is used to mesh the area near the hull surface. A
xS m 0.78
h m 0.967 grid-independency test is carried out based on the original design hull
BS m 2.76 and speed by using three different grid configurations, and the corre-
sponding meshes are shown in Fig. 5. The information of three grid
configurations and the test results are shown in Table 5.
Table 3 The computations were performed on a workstation with an Intel
Design ranges of skeg geometry parameters. Core Xeon-X5647 CPU with 2.93 GHz and 72 GB RAM. Compared to
Design variables Initial values Bound coarse grid, the changes in RT of fine grid is only 0.83% whereas the
grid number of the fine grid is 2.75 times, and the computing time is
Lower Upper 5.75 times of coarse grid. The computation cost of medium grid is also
x1 b/B 0.446 0.3 0.6
high while the computational results are closed to that of coarse grid.
x2 θS 0 −20 45 Therefore, the coarse grid can be used for CFD calculations in optimi-
x3 lS/LPP 0.29 0.25 0.35 zation. The local grid refinement of the bow and stern area are shown in
x4 xS/LPP 0.01 0 0.04 Fig. 6.
x5 h/T 0.333 0.25 0.4
x6 2BS/B 0.373 0.2 0.4
3. Applying surrogate model method for optimization

(m/s2) and LPP is the length between perpendiculars (m).
In this study, the scaling ratio α is set as 0.1. The calculation domain
is shown in Fig. 4, in which L means the length of ship model
(L = α × LPP), the setting of the computational domain is as follows:
Longitudinal: −5 L ≤ X ≤ 2.5 L
Transversal: 0 ≤ Y ≤ 2.5 L
Vertical: −2 L ≤ Z ≤ L
3.1. Construction of surrogate models
Based on statistical theory, Kriging modeling method has better
reflection of the overall trend of the data points and produces zero error
in the estimation of sample points comparing to other surrogate model
methods.
Suppose a sample set X = [x1, x2, …, xn]T, n is the total number of
sample points, xi is a m-dimensional vector that represents the ith
sample point, so X is a n × m design matrix. The outputs are Y = [y(x1),

41
Y. Lin et al. Advances in Engineering Software 123 (2018) 38–50

Fig. 4. Computational domain and coordinate system.

Table 4 y(x2), …, y(xn)]T, where y(xi) can be used to express resistance at point
Configuration of calculating speeds. xi. The following model is used to model deterministic computer re-
Navigational speed Fr
sponse as

Full-scale (knot) Full-scale (m/s) Model-scale (m/s)

y (x ) = f T (x ) β + Z (x ) (2)

9 4.630 1.464 0.167 where x is m-dimensional design variable vector of untried point, y(x) is
10 5.144 1.627 0.186 a predicted value at point x, f (x) = [f1(x), …, fs(x)]T is the polynomial
11 5.659 1.789 0.205 function of x and β = [β1, …, βs]T is the regression coefficient vector, Z
12 6.173 1.952 0.223
(x) is a random process with mean zero and covariance

cov[Z (xi ), Z (xj )] = σ 2R (xi , xj ) (3)

Fig. 5. Grid configurations for the grid-independency test: coarse grid, medium grid and fine grid.

42
Y. Lin et al. Advances in Engineering Software 123 (2018) 38–50

Table 5 Table 6
Grid information and grid-independency test results. Common types of correlation functions.
Grid type Coarse Medium Fine Name R(dk)

Grid base size (m) 0.1125 0.09 0.07 Linear R(dk) = max (0, 1 − θk|dk|)
Grid number 525,615 846,339 1,446,046 Exponential R(dk) = exp ( − θk|dk|)
Computing time (h) 4 12 23 Generalized exponential R(dk) = exp ( − θk|dk|p),p ∈ [1, 2]
RP (N) 27.325 26.150 25.407 Gaussian R(dk) = exp ( − θk|dk|2)
RS (N) 56.177 56.929 57.406
RT (N) 83.502 83.079 82.812
|RP-RP, Corse|/RP, Corse – 4.30% 7.02%
where xik − xjk is the spatial distance between the two sample points in
|RS-RS, Corse|/RS, Corse – 1.34% 2.19%
|RT-RT, Corse|/RT, Corse – 0.51% 0.83%
the kth dimension. Table 6 shows several common types of SCF, where
dk = xik − xjk. θ is an unknown covariance parameter vector. θk is the
kth parameter corresponding to the kth variable, which implies the
where σ2 is the process variance and R(xi,xj) is the spatial correlation importance of input xk. p implies the smoothness of the SCF, the ex-
function (SCF) between the sample points xi and xj. ponential and Gaussian correlation functions have p = 1 and p = 2
respectively.
m The Gaussian function is the most commonly used SCF, which is a
R (xi , xj ) = ∏ R (xik − xjk ) special case of the generalized exponential function, it provides a very
k=1 (4)

Fig. 6. Local grid refinement of the bow and stern area.

43
Y. Lin et al. Advances in Engineering Software 123 (2018) 38–50

Table 7
Parameters of Kriging models.
Fr y θ1 × 102 θ2 × 102 θ3 × 102 θ4 × 102 θ5 × 102 θ6 × 102 β × 102 MLE

0.167 yP 29.131 124.594 1.856 9.169 12.408 4.168 15.333 99.855

yS 19.808 33.421 45.071 2.708 0.605 32.312 −13.378 42.108
yT 28.025 125.103 1.595 2.247 19.697 6.821 18.657 98.059
0.186 yP 20.455 167.518 1.931 0.122 8.389 18.850 15.737 62.983
yS 25.027 30.980 41.517 8.788 0.016 31.055 −11.077 21.819
yT 25.998 157.655 0.159 0.031 6.962 22.065 15.735 63.802
0.205 yP 20.751 152.699 7.191 0.109 4.352 18.217 18.516 63.581
yS 22.313 49.132 78.841 0.045 0.017 10.028 −8.962 24.701
yT 28.082 144.820 1.052 0.119 7.908 19.772 17.061 64.805
0.223 yP 24.214 169.829 6.080 0.066 4.075 13.662 16.078 64.922
yS 22.884 60.850 80.213 0.024 0.748 3.931 −11.317 26.494
yT 24.693 140.629 6.423 0.069 5.459 16.790 17.923 66.892

Table 8
Additional checking points and corresponding simulation results.
No. Design variables Fr RP RS RT

x1 x2 x3 x4 x5 x6

1 0.512 0.342 0.307 0.002 0.358 0.227 0.223 33.699 62.187 95.886
2 0.466 9.117 0.280 0.003 0.291 0.211 0.205 23.213 52.094 75.278
3 0.573 −19.489 0.310 0.020 0.348 0.308 0.186 22.968 43.232 66.941
4 0.414 −16.412 0.258 0.033 0.278 0.261 0.167 17.235 34.818 52.053
5 0.392 39.446 0.304 0.033 0.251 0.256 0.223 31.837 61.004 93.437
6 0.377 42.428 0.284 0.010 0.299 0.282 0.205 23.842 51.414 75.256
7 0.375 −10.173 0.311 0.022 0.290 0.286 0.186 20.291 42.777 63.572
8 0.409 44.647 0.350 0.008 0.397 0.345 0.167 13.238 34.624 47.662
9 0.440 −3.704 0.348 0.024 0.333 0.279 0.223 34.472 61.454 95.926
10 0.588 −5.237 0.266 0.033 0.266 0.332 0.205 27.612 51.939 79.622
11 0.442 −17.937 0.294 0.031 0.391 0.253 0.186 21.576 42.789 64.165
12 0.570 −12.923 0.255 0.020 0.272 0.344 0.167 17.580 35.059 53.038
13 0.321 11.112 0.283 0.026 0.378 0.275 0.223 31.302 60.996 92.034
14 0.459 14.335 0.272 0.031 0.293 0.323 0.205 24.345 51.418 74.762
15 0.554 −2.118 0.299 0.024 0.366 0.369 0.186 19.369 42.825 62.194
16 0.427 −12.145 0.272 0.005 0.389 0.357 0.167 16.171 35.001 51.171

Table 9 smooth and infinitely differentiable surface and is defined with only
Fitting accuracy of Kriging model. one vector parameter θ. However, the study is not a continuously dif-
NRMSE R2 eavg ferentiable problem which means there is no need to control the
smoothness of the surrogate models, so the exponential type SCF is
RP 1.95% 92.86% 1.53% chosen to construct the surrogate models because it only controls the
RS 0.99% 97.42% 0.48% range of influence of each input variable. Therefore, Eq. (4) can be
RT 1.51% 95.53% 0.89%
written as Eq. (5):

Fig. 7. Relative errors scatter of responses.

44
Y. Lin et al. Advances in Engineering Software 123 (2018) 38–50

Fig. 8. Sensitivity analysis results of design variables.

m n
⎡ ⎤ (2πσ 2)− 2 (Y − Fβ )T R−1(Y − Fβ ) ⎤
R (xi , xj ) = exp ⎢− ∑ θk x ik − x jk ⎥ L (θ, β, σ 2) = exp ⎡−
R ⎢
⎣ 2σ 2 ⎥
⎦ (6)
⎣ k=1 ⎦ (5)

The parameters β and σ2 can be estimated by maximizing the like- where F = [f(x1),⋅⋅⋅, f(xn)]T is a n × s matrix including the value of f(x)
lihood of samples. The likelihood function is evaluated at each of the experimental points, R is an n × n correlation
matrix as follows:

45
Y. Lin et al. Advances in Engineering Software 123 (2018) 38–50

Start Table 10
Comparison of the optimization result and the original design.
Original Optimized
Generation of the
initial population x1 0.446 0.440
x2 0 31.509
x3 0.293 0.281
Fitness assessment x4 0.022 0.007
by surrogate models, x5 0.333 0.352
x6 0.373 0.306
initial the Pareto set
RT,9 (N) 50.214 47.121
RT,10 (N) 62.256 58.909
RT,11 (N) 77.005 73.220
Selection RT,12 (N) 95.236 90.378

Crossover
By substituting Eq. (8) to ln L(θ, β, σ2) function, the correlation
parameters can be calculated by Eq. (9):
Mutation Maximize ln L (θ ) = −[n ln 2πσ 2̂ + ln R + 1]/2
No
s. t . θ>0 (9)
Fitness assessment by surrogate Eq. (2) can be written as Eq. (10):
models, update Pareto set
y (̂ x ) = β + r T (x ) R−1(Y − Fβ ) (10)

where r(x) is the correlation vector that represents the correlation be-
Meet the Convergence Standard? tween an untried point x and all sample points.
r (x ) = [R (x , x1), R (x , x2), ...,R (x , xn )]T (11)
Yes
The coefficients and parameters needed for constructing the surro-
End
gate models are summarized in Table 7.
Fig. 9. Flowchart of applying NSGA-II.
3.2. Accuracy evaluation

The Kriging model produces zero error at sample points because it

reproduces the exact observations at sample points, therefore, only the
capability of predicting the values at untried points should be assessed.
The metrics used here to assess the fitting accuracy are normalized root
mean square error (NRMSE), R-square (R2) and average relative error
(eavg). A moderate number of additional verification points that have
not been used in the construction of the surrogate models are needed.
16 additional checking points with four different navigation speeds are
generated by using OLHS method for the accuracy evaluation. The si-
mulation results of the checking points are shown in Table 8.
The accuracy evaluation metrics are calculated by using Eqs.
(12)–(14):

∑iq= 1 (yi − yi )̂ 2
q
NRMSE =
̂ − ymin
ymax ̂ (12)
q
∑i = 1 yi − yi ̂
R2 = 1 − q
∑i = 1 yi − y (13)
Fig. 10. The effective power curve of the optimal solution. q
1 yi − yi ̂
eavg =
q
∑ yi
i=1 (14)
⎡ R (x1, x2) ⋯ R (x1, xn ) ⎤
R=⎢ ⋮ ⋱ ⋮ ⎥ where q is the number of the checking points, yi is the simulation re-
⎢ R (x , x ) ⋯ R (x , x ) ⎥ sponse at each checking point and y is the average value of the re-
⎣ n 1 n n ⎦ (7)
sponses of all the checking points; yi ,̂ ymax ̂ ̂
and ymin are the corre-
The unknown parameters β and σ2 can be estimated as [18] sponding surrogate values at the checking points and their maximum
and minimum respectively. Table 9 shows the accuracy results. The
β = (F TR−1F)−1F TRTY normalized root mean square error (NRMSE) and the average relative
σ 2̂ = (Y − Fβ ) R−1(Y − Fβ )/ n error (eavg) of the total resistance are less than 2%, and the R2 value of
T
(8)
the total resistance is higher than 95%. As shown in Fig. 7, the nu-
In this study, a constant regression model is used, so f(x) = 1, F = 1. merical simulation results and predicted values are closely distributed

46
Y. Lin et al. Advances in Engineering Software 123 (2018) 38–50

Fig. 11. Comparison of the free surface between the original and optimized design.

around the line RACTUAL = RPREDICTED, which means the surrogate
model have a good prediction of the twin-skeg vessel's total resistance
when the geometry parameters are varied.
4. Parameter sensitivity analysis and hull form optimization
0.186, 0.205, 0.223) are chosen as the optimization objective functions
for the skeg shape optimization, constraints should maintain the design
dead weight (DWT) and keep the variations of displacement ∇ in an
acceptable range. The optimization problem of twin-skeg fishing vessel
can be written as Eq. (15):

The total resistances under four working conditions (Fr = 0.16,

47
Y. Lin et al. Advances in Engineering Software 123 (2018) 38–50

Fig. 12. Comparison of streamlines around skeg between the original and optimized design.

min F (x ) = {f1 (x ), f2 (x ), f3 (x ), f4 (x )} where RT,9, RT,10, RT,11 and RT,12 are total resistances under four
f1 (x ) = RT ,9 (x ) working conditions respectively, LPP0, B0, T0, DWT0 and ∇0 are the
f2 (x ) = RT ,10 (x ) main dimensions, dead weight and displacement of the original design.
f3 (x ) = RT ,11 (x )
f4 (x ) = RT ,12 (x ) 4.1. Parameter sensitivity
x = [x1, x2 , x3 , x 4 , x5 , x 6]
s. t . h1 = LPP − LPP 0 = 0 In order to identify how much the influence of each design variable
h2 = B − B0 = 0 have on design objective, the functional analysis of variance (ANOVA)
h3 = T − T0 = 0 method is used [8]. One of the four working conditions is chosen to
h4 = DWT − DWT0 = 0 show the functional ANOVA result. Fig. 8 shows the sensitivity analysis
results on the resistances with normalized design variables under the
∇ − ∇0
g= ≤ 1% condition when the full-scale ship speed is 10 knots.
∇0 According to Fig. 8, the sensitivities of the design variables on the
x1 ∈ [0.3, 0.6] pressure and the total resistance are very similar, which means the
x2 ∈ [−20, 45] changes of the skeg geometry parameters mainly affect the pressure
x3 ∈ [0.25, 0.35] resistance RP. The sensitivity of the vertical inclination angle θS on total
x 4 ∈ [0, 0.04] resistance RT is much higher than that of other design variables, RT
x5 ∈ [0.25, 0.4] increases when the skeg is inward-inclined and reaches its minimum
x 6 ∈ [0.2, 0.4] (15) when the vertical inclination angle is around 73% of its range of value,
which means θS = 27.5°. This is because a certain degree of outward-

48
Y. Lin et al.
Fig. 13. Comparison of velocity distribution of wake field in the propeller plane.
inclination makes the gradient in the stern lines smaller, then the
streamlines in the stern area get smoother.
The total resistance is also sensitive to b/B and h/T. When b/B in-
creases and/or h/T decreases, the stern body become bigger, the hull
form will be like to a non-streamlined object whose boundary layer will
be separated, then the viscous pressure resistance will increase. In the
consideration of rational propeller arrangement, a proper reduction in
b/B and increase in h/T will be conducive to decrease the total re-
sistance. As to the shear resistance RS, it is sensitive to b/B, the lS/LPP,
BS and θS, every change that causes the increase of the wetted surface
area will make the RS increase.
4.2. Optimization process
The effective power curves of different design points may intersect
with each other, which means the optimal design point cannot be de-
termined by single objective optimization. Therefore, the multi-objec-
tive optimization of the four working conditions is conducted instead of
the optimization of the resistance at the design speed.
The non-dominated sorting genetic algorithm II (NSGA-II) is
49
Advances in Engineering Software 123 (2018) 38–50
adopted here to solve this multi-objective problem. The NSGA-II algo-
rithm reduces the complexity of the sorting of the non-inferior solutions
in genetic algorithm and has the advantages of high processing speed
and good convergence [19]. Fig. 9 is the procedure of the optimization
process of the NSGA-II. In the application of NSGA-II in this paper, the
population size is 120, number of generations is 200, the crossover
probability is 0.7, and the mutation probability is 0.03. The effective
power curve of the optimal solution is selected among the Pareto set by
the minimum distance selection method [20], which is shown in
Fig. 10. It shows 5.4% average decrease in the total resistance in
comparison with the original design. The optimal parameters and re-
sistance are shown in Table 10. The optimized vertical inclination angle
is 31.509°, which agrees with the result of the functional ANOVA
analysis. By comparing the original and the optimized design variables,
it can be found that the reduction in total resistance is mainly caused by
the changes in the vertical inclination angle, the starting position of the
skeg and the half breadth of the skeg.
To examine the resistance performance of the optimal solution
based on Kriging model, numerical simulation of the optimized design
under the corresponding full-scale design speed (11.3 knots) is carried
Y. Lin et al.
out, and the total resistance RT is 78.824 N, which shows 5.6% decrease
than the original design. It can be observed from Fig. 11 that in the
optimized design, waves generated by the hull are much less than the
original design. It is mainly because the skegs are outward inclined and
have a better performance to restrain the wave-making.
Fig. 12 shows the comparison of streamlines around skeg between
the original and optimized design. In the original design, some of the
streamlines flow from one side to the other side of the skeg, which
cause greater pressure difference between the two sides of a skeg. The
streamlines of optimized design show less crossover than the original
design, which is also conducive to lowering the resistance.
Fig. 13 shows the comparison of the velocity distribution of wake
field in the propeller plane between the original and optimized design.
The velocity vectors describe the projections of the spatial velocity
vectors on YZ plane. The value of a projection vector reflects that how
much velocity will be left in the forward direction of the ship. High
value of one projection vector means few velocity left in the forward
direction of the ship. It can be observed from the comparison that after
optimization the velocity vectors around the propeller center become
shorter, which means that the optimized design has less velocity loss in
the YZ plane than the original design.
5. Conclusions
This paper presents a combined approach for hull form design op-
timization of twin-skeg ship by using CFD calculation and surrogate
model. Main design parameters of skeg geometry and arrangement
could be determined from the design domain by using the proposed
method. A parametric modeling process is adopted to perform design
evaluations in an automatic manner for different design parameter
combinations.
A twin-skeg fishing vessel is selected as research object. Aiming on
reducing the total resistance in the preliminary stage of ship design,
based on CFD calculations, the proposed optimization approach con-
structs surrogate models by using Kriging modeling method to lower
calculation cost. To better fit the hull form optimization problem,
multiple speeds are considered to obtain the effective power curve.
After carefully study on skeg design and CFD mesh configuration, de-
sign parameters for parameterized modeling and CFD calculation en-
vironment are determined, a sample set is created by using OLHS
method, the corresponding CFD calculations of all sample points are
performed, and then the calculated results are used to construct sur-
rogate models to predict ship resistance performance with skeg geo-
metry variations. The functional ANOVA is introduced to investigate
how much influence each design variable has on the objective function.
The final optimized design, which is obtained by using multi-objective
optimization algorithm NSGA-II as solving strategy, shows 5.4%
average reduction in the total resistance in comparison with the ori-
ginal design. The CFD calculation results of the optimal solution ob-
tained by NSGA-II show that the proposed method can achieve
minimum resistance design with high accuracy and low time cost.
As for the future works, the hull-propeller-rudder interaction of the
twin-skeg hull form will be studied by using dynamic grid method.
Furthermore, by combining other disciplines in the shipbuilding field,
multi-disciplinary optimization study can be carried out based on the
study in this paper.
50
Advances in Engineering Software 123 (2018) 38–50
Acknowledgment
This work was supported by the National Natural Science
Foundation of China (Grant No. 51509033).
Supplementary materials
Supplementary material associated with this article can be found, in
the online version, at doi:10.1016/j.advengsoft.2018.05.010.
References
[1] Sakamoto N, Kawanami Y, Uto S, Sasaki N. Estimation of resistance and self-pro-
pulsion characteristics for low L/B twin-skeg container ship by a high-fidelity RANS
solver. J. Ship Res. 2013;57(1):24–41https://doi.org/10.5957/JOSR.57.1.110032.
[2] Kim K, Tilling F, Bathfield N, Liljenberg H. Hydrodynamic optimization of twin-skeg
LNG ships by CFD and model testing. J Naval Architect Ocean Eng
2014;68(12):1806–18https://doi.org/10.1016/j.camwa.2014.10.010.
[3] Park DW, Chun HH. Design practice for the stern hull form of a twin-skeg ship. J
Mar Sci Technol 2009;14(3):310–21https://doi.org/10.1007/s00773-009-0046-5.
[4] Chen JP, Wei JF, Jiang WJ. Optimization of a twin-skeg container vessel by para-
metric design and CFD simulations. Int J Naval Architect Ocean Eng
2016;8(5):466–74https://doi.org/10.1016/j.ijnaoe.2016.05.008.
[5] Simpson TW, Poplinski JD, Koch PN, Allen JK. Metamodels for computer-based
engineering design: survey and recommendations. Eng Comput
2001;17(2):129–50https://doi.org/10.1007/PL00007198.
[6] Krige DG. A statistical approach to some basic mine valuations and allied problems
at the witwatersrand. J South Afr Inst Min Metall 1951;52(6):119–39http://hdl.
handle.net/10520/AJA0038223X_4792.
[7] Simpson T, Mistree F, Korte J, Mauery T. Comparison of response surface and kri-
ging models for multidisciplinary design optimization. In: 7th AIAA/USAF/NASA/
ISSMO Symposium on Multidisciplinary Analysis and Optimization; 1998 Sept 02-
04, St. Louis, Missouri, Document ID: AIAA-98-4755. https://doi.org/10.2514/6.
1998-4755.
[8] Jeong SK, Mitsuhiro M, Kazuomi Y. Efficient optimization design method using
kriging model. J Aircraft 2005;42(2):413–20https://doi.org/10.2514/1.6386.
[9] Huang D, Allen TT, Notz WI, Miller RA. Sequential kriging optimization using
multiple-fidelity evaluations. Struct Multidis Optim 2006;32(5):369–82https://doi.
org/10.1007/s00158-005-0587-0.
[10] Martin JD. Computational improvements to estimating kriging metamodel para-
meters. J Mech Des 2009;131(8):084501https://doi.org/10.1115/1.3151807.
[11] Kleijnen JPC. Kriging metamodeling in simulation: A review. Eur J Oper Res
2009;192(3):707–16https://doi.org/10.1016/j.ejor.2007.10.013.
[12] Sacks J, Schiler SB, Welch WJ. Design for computer experiment. Technometrics
1989;31(1):41–7https://doi.org/10.1080/00401706.1989.10488474.
[13] Laurenceau J, Sagaut P. Building efficient response surfaces of aerodynamic func-
tions with kriging and cokriging. AIAA J 2008;46(2):498–507https://doi.org/10.
2514/1.32308.
[14] Lee KH, Kang DH. Structural optimization of an automotive door using the kriging
interpolation method. Proc Inst Mech Eng Part D J Automob Eng
2007;221(12):1525–34https://doi.org/10.1243/09544070JAUTO403.
[15] Chen SM, Shi TZ, Wang DF, Chen J. Multi-objective optimization of the vehicle ride
comfort based on kriging approximate model and NSGA-II. J Mech Sci Technol
2015;29(3):1007–18https://doi.org/10.1007/s12206-015-0215-x.
[16] Gao DW, Zhang N, Feng JZ. Multi-objective optimization of crashworthiness for
mini-bus body structures. Adv Mech Eng 2017;9(7):1–11https://doi.org/10.1177/
1687814017711854.
[17] Bates S, Sienz J, Toropov V. Formulation of the optimal latin hypercube design of
experiments using a permutation genetic algorithm. In: 45th AIAA/ASME/ASCE/
AHS/ASC Structures, Structural Dynamics & Materials Conference; 2004 Apr 19-24,
Palm Springs, California, Document ID: AIAA-2004-2011. https://doi.org/10.2514/
6.2004-2011.
[18] Xia B, Pham MT, Zhang YL, Koh CS. A global optimization algorithm for electro-
magnetic devices by combining adaptive taylor kriging and particle swarm opti-
mization. IEEE Trans Magn 2013;49(5):2061–4https://doi.org/10.1109/TMAG.
2013.2238907.
[19] Deb K, Pratap A, Agarwal S, Meyarivan T. A fast and elitist multiobjective genetic
algorithm: NSGA-II. IEEE Trans Evol Comput 2002;6(2):182–97https://doi.org/10.
1109/4235.996017.
[20] Barakat S, Bani-Hani K, Taha MQ. Multi-objective reliability-based optimization of
prestressed concrete beams. Struct Saf 2004;26(3):311–42https://doi.org/10.1016/
j.strusafe.2003.09.001.