Optimization 2021

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Ocean Engineering 230 (2021) 109050

Contents lists available at ScienceDirect

Ocean Engineering
journal homepage: www.elsevier.com/locate/oceaneng

Hull optimization of an underwater vehicle based on dynamic


surrogate model
Weilin Luo *, Xiaoming Guo , Jiawei Dai , Taichun Rao
School of Mechanical Engineering and Automation, Fuzhou University, Fuzhou, Fujian, 350116, China

A R T I C L E I N F O A B S T R A C T

Keywords: A parallel multidisciplinary optimization design is proposed for the lines design of an underwater vehicle.
Underwater vehicle Resistance and energy consumption are concerned about in obtaining the optimized lines. Collaborative opti­
Collaborative optimization mization strategy is adopted to construct the optimization framework. The simulated annealing algorithm in
Resistance
combination of the modified method of feasible directions are taken as the optimization algorithms. CFD based
Energy consumption
CFD calculation
simulation is conducted to analyze the resistance of the underwater vehicle. To improve the optimization effi­
Dynamic surrogate model ciency, a surrogate model is incorporated into the optimization framework to approximate the CFD calculation.
Unlike the conventional static surrogate models used in the design of underwater vehicles, a dynamic surrogate
based on trust region and lower confidence bound is proposed. An adaptive balance parameter is taken in the
lower confidence bound approach to guarantee the tradeoff between the exploitation and exploration ability of
the dynamic surrogate model. Optimal Latin hypercube algorithm is employed as the method of design of ex­
periments. Radial basis function is employed to construct the surrogate model. SUBOFF model is used to
demonstrate the proposed optimization scheme. Lines of the underwater vehicle are determined and optimiza­
tion results of the resistance and energy consumption show the effectiveness of optimization scheme.

1. Introduction hydrodynamics, structure, automation, mechanics (mechanical engi­


neering), electrics, and communication etc. Conventional optimization
Underwater vehicle is an important tool for human beings to explore design of an underwater vehicle observes the methodology of series
and exploit ocean. Design of an underwater vehicle is a complex system design. Single discipline is considered at some stage of design. For
engineering project in which multiple performances should be taken example, at the stage of schematic design, the hydrostatic and hydro­
into account. Representative performances are hydrodynamic perfor­ dynamic performances are concerned while other disciplines like
mances including maneuverability, rapidity and seakeeping. Rapidity is structure, mechanics are ignored. As a result, it is apt to result in low
a main concern in the design of underwater vehicles. Resistance is an efficiency and loss of globally optimal solution by using series design.
important aspect reflecting the rapidity performance of an underwater Multidisciplinary design optimization (MDO) provides an effective way
vehicle. To reduce the resistance becomes comprehensive and chal­ to solve that problem. Using this design methodology, several closely
lenging work, which can be achieved by designing optimal lines. The related disciplines are taken into account simultaneously, which gua­
lines also exert a considerable influence on the other hydrodynamic rantees the overall optimal performance, improves the optimization
performances, i.e. maneuverability and seakeeping. How to obtain the efficiency as well. Since MDO was put forward in early 1980s
optimal lines is the main concern at the stage of schematic design of an (Sobieszczanski-Sobieski, 1982), it has been developed greatly and
underwater vehicle. Optimization design provides an approach and has applied widely in many areas such as aerospace engineering (Braun
been widely applied during past decades for example the studies in et al., 1997) (Valerie Michelle, 1999), automobile engineering (Yang
(Alam et al., 2014; Alvarez et al., 2009; Joung et al., 2012; Sun et al., et al., 2002) (Zhang et al., 2008) and marine engineering. For under­
2015). water vehicles, MDO has shown its feasibility at the stage of schematic
From the point of view of systems engineering, the design work of an design. For example, McAllister et al. (2002) presented a feasible MDO
underwater vehicle involves multiple disciplines such as approach that could maximize payload length of an autonomous

* Corresponding author. University School of Mechanical Engineering and Automation, Fuzhou, Fujian, 350116, China.
E-mail address: wlluo@fzu.edu.cn (W. Luo).

https://doi.org/10.1016/j.oceaneng.2021.109050
Received 8 December 2020; Received in revised form 24 February 2021; Accepted 18 April 2021
Available online 28 April 2021
0029-8018/© 2021 Elsevier Ltd. All rights reserved.
W. Luo et al. Ocean Engineering 230 (2021) 109050

Fig. 1. The framework of collaborative optimization.

underwater vehicle (AUV). Cao et al. (2007) proposed MDO to a con­ optimization process. For a complex optimization scenario, it requires
ceptual submarine design in which hydrodynamics and cost were quite a lot samples to achieve satisfactory accuracy. Therefore, the
considered. Vlahopoulos and Hart (2010) used the MDO to relating optimization efficiency cannot be guaranteed. Compared with static
affordability and performance in a conceptual submarine design. Alam approach, a dynamic surrogate model adaptively updates the samples
et al. (2012) addressed a new robust design optimization approach for with proper capacity. In terms of both efficiency and accuracy, a dy­
unmanned underwater vehicle design. Luo and Lyu (2015) applied MDO namic surrogate model is preferable. Pérez et al. (2002) proposed an
to the hydrodynamics optimization of an AUV. Bidoki et al. (2018) adaptive experiment design for the construction of a response surface
proposed a multidisciplinary feasible framework in combination of based surrogate model. Cheng and Wang (2012) integrated the trust
particle swarm optimization for the system and static design of an AUV. region to a mode pursing sampling based optimization framework to
Although MDO has been proven as an effectively way in the design of solve high dimensional optimization problem. Long et al. (2014)
underwear vehicles, some issues are to be further addressed. One is the addressed the trust region based surrogate model combined with low
optimization algorithm. Conventional optimization algorithms are and high dimensional optimization examples. Eason and Cremaschi
referred to as gradient-based methods such as successive quadratic (2014a) presented an adaptive sequential sampling for surrogate model
programming (SQP), method of feasible direction (MFD), large scale generation with artificial neural networks. Su et al. (2017) proposed a
generalized reduced gradient (LSGRG) and multifunction optimization Gaussian process-based dynamic surrogate model for complex engi­
system tool (MOST). Although rapid convergence rate can be achieved neering analysis. Yu et al. (2020) put forward a dynamic surrogate
by these algorithms, it is apt to result in locally optimal solution. To assisted evolutionary algorithm framework for structural optimization.
obtain globally optimal solutions, derivative-free methods are available Generally speaking, the research on dynamic surrogate models should
for example genetic algorithm (GA), particle swarm optimization (PSO), be paid more attention to and its feasibility in practical engineering
evolutionary algorithm (EA) and simulated annealing (SA). However, should be verified furthermore for example in the optimization design of
using these global optimization algorithms, the optimization efficiency marine vehicles.
cannot be guaranteed especially for large scale design scenarios. In this paper, an approach of MDO, collaborative optimization (CO),
Another issue in the application of MDO to underwater vehicles is the is applied to the lines design of an underwater vehicle. Reduction of
analysis of subdiscipline. Empirical formula is commonly used since it resistance and energy consumption is the optimization goal. A hybrid
provides a shortcut to evaluate the investigated performance in a sub­ optimization algorithm is built up in which a global optimization algo­
discipline. As a result, the optimization accuracy cannot be guaranteed. rithm- SA and a rapid gradient-based optimization algorithm-MMFD are
With the development of numerical analysis methods and computer combined with each other. To evaluate the resistance of the investigated
technology, some high-accuracy numerical computational models are underwater vehicle, CFD calculation based on RANS equation is per­
introduced into the optimization framework. A representative model is formed and compared against experimental results. Within the optimi­
based on CFD technique that is used to evaluate the hydrodynamics for zation framework, the CFD module is replaced by a dynamic surrogate
example (Liu et al., 2017; Liu and Luo, 2016; Luo and Lyu, 2015; model based on radial basis function (RBF). The samples used to
Vasudev et al., 2014). However, direct incorporation of CFD calculation construct the surrogate model are obtained by design of experiments
into a MDO platform degrades the optimization efficiency because (DOE) and updated by means of trust region and lower confidence
performing a CFD calculation is time-consuming and has to be repeated bound. Optimal Latin hypercube algorithm is adopted as the DOE
during an optimization process. To improve the optimization efficiency, method. The overall optimization is carried out on an Isight platform.
surrogate models are proposed to substitute the CFD calculation (Luo The SUBOFF model, which is from the National Defense Advanced
and Lyu, 2015), (Liu and Luo, 2016), (Gu et al., 2009), (Song et al., Research Project Agency (Groves et al., 1989) (Liu and Huang, 1998), is
2013). However, how to construct an appropriate surrogate model with investigated to demonstrate the optimization strategy proposed in the
good optimization efficiency and accuracy needs to be further studied. study. The main contribution of the study is the introduction of dynamic
Static surrogate model and dynamic surrogate model are two kinds of surrogate model in the lines design of underwater vehicles, instead of
surrogate model in the research on optimization design. It is noted that the conventional static surrogate model in literature. Another contri­
nowadays, static surrogate models are in common use while dynamic bution of the study is the optimization algorithm based on the combi­
surrogate models are paid much less attention to. For static surrogate nation of SA and MMFD, which aims to guarantee globally optimal
models, both the structure and samples are kept unchanged during an solution and optimization efficiency.

2
W. Luo et al. Ocean Engineering 230 (2021) 109050

The structure of CO has a two-level hierarchical form, involving a


system-level optimizer and several discipline-level optimizers, as shown
in Fig. 1, where f(z) is the optimization objective at system-level; x*ij is
the optimal design variable; zj is the design variable; J* (z) is the con­
sistency constraint; Ji (xi ) is the optimization objective at discipline-
level; xij is the design variable at discipline-level; z*j is the desired
design variable allocated from system-level; g(xij , xil ) is the discipline-
level constraint. In the CO based optimization procedure, desired
values of design variables are transferred from the system-level to the
discipline-level where optimization is carried out in parallel and inde­
pendently for each discipline with respect to the allocated desired design
variables from system-level. The optimized results are passed recur­
sively to system-level to calculate the consistency constraint. Mean­
while, the optimization is performed at the system-level. Such a two-
level optimization is conducted recurrently until the optimization goal
as well as the constraint is satisfied.
A standard CO can be described as Fig. 1. As can be seen, at the
system-level the consistency constraint is a kind of equality constraint,
which might result in locally optimal solution or low efficiency of
convergence because of the difficulty with Karush-Kuhn-Tucker neces­
sary condition. Measures to solve that problem are mainly referred to the
alleviation of the constraints and adoption of global optimization algo­
rithms. Nevertheless, it should be noted that convergence to globally
optimal solution is in general difficult for real application problems.
Slack factor or penalty function is one of two usual options to alleviate
the constraints for examples (Alexandrov and Lewis, 2002) (Lin, 2004).
In the study, slack factors are introduced. Moreover, adaptive simulated
annealing (ASA) is employed to guarantee the globally optimal solution.
ASA has been proven as an effective global optimizer (Kirkpatrick et al.,
1983), (Rafique et al., 2009). However, the main disadvantage is the
convergence rate especially for large scale complex design optimization.
Fig. 2. ASA-MMFD-CO optimization process.
Therefore, a gradient-based optimization algorithm, the modified
method of feasible direction (MMFD), is combined with ASA. MMFD is
2. Collaborative optimization
characterized by robustness and rapid convergence rate (Secanell and
Suleman, 2005), (Venter, 2010). A searching strategy based on ASA in
Optimization procedure is one of five important MDO components
combination of MMFD can be described as follows. The first step is to
(Sobieski and Haftka, 1997). Typical examples of MDO procedure are
quickly determine the range of globally optimal solution. ASA is
multidisciplinary feasible method (MDF) (Adelman and Mantay, 1991),
employed and a big slack factor is selected. Secondly, to precisely and
individual discipline feasible (IDF) (Balling and Wilkinson, 1997),
quickly locate the globally optimal solution, MMFD is used and a small
simultaneous analysis and design (SAND) (Lavelle et al., 1991),
slack factor is selected. The process of improved collaborative optimi­
collaborative optimization (CO) (Kroo et al., 1994), concurrent subspace
zation algorithm is depicted as Fig. 2.
optimization (CSSO) (Sobieski, 1989), analytical target cascading (ATC)
(Lin and Gea, 2013) and bi-level integrated system synthesis (BLISS)
3. Numerical simulation of hydrodynamics
(Sobieszczanski-Sobieski et al., 2000). Generally, the development of
MDO procedure has experienced three stages. At the first stage, the
CFD based numerical simulation provides an effective way to eval­
analysis and integration of disciplines are mainly concerned about. The
uate the hydrodynamic performance of marine vehicles. In the study,
representative approach is MDF. At the second stage, efforts are mainly
fluent module of ANSYS, a popular commercial CFD software package, is
devoted to the parallel analysis of disciplines and data management.
employed. The investigation model is taken as the underwater robot
Typical approaches are IDF and SAND. The third stage is characterized
project SUBOFF, conducted by the Defense Advanced Research Projects
by the involvement of expert knowledge. Approaches including CO,
Agency, since the project has already involved a wealth of experimental
CSSO, ATC and BLISS are developed at this stage. For a complicated
data. Moreover, some of the data have been released publicly. Re­
engineering project like an underwater vehicle, CO is preferable since it
searchers within the hydrodynamic community can use the data as a
is characterized by hierarchy, integration and parallel processing
reference to verify the validity of numerical simulation performed. The
(Hosseini et al., 2017).

Fig. 3. SUBOFF model.

3
W. Luo et al. Ocean Engineering 230 (2021) 109050

Fig. 4. Suboff and flow field.

SUBOFF model is shown as Fig. 3, in which Loa is the overall length; L the Afterbody cap (L + Lfo + Laf ≤ x ≤ Loa )
parallel body length; Lfo the forebody length; Laf the afterbody length; Lca { }1
the afterbody cap length; R the radius of the parallel middle body. R(x) = 0.1175R 1 − (3.2x − 44.733333)2 2 (5)
As an axisymmetric rotation body, the mathematical model of the
SUBOFF model can be determined by radius R(x) in the (o-xyz) Cartesian
coordinate system. 3.1. Computing domain and meshing
Forebody (0 ≤ x ≤ Lfo ):
{ To perform numerical simulation, initially it is necessary to deter­
R(x) = R 1.126395101x(0.3x− 1)4 + 0.44287470x2 (0.3x− 1)3 +1 mine the computing domain and generate grids. In general, a large-scale
domain is very burdensome for calculation, moreover the requirements
}1/2.1 for workstation also increase greatly. Too small calculation domain
− (0.3x− 1)4 (1.2x + 1) , (1)
cannot simulate well the underwater vehicle’s movement, even results
Parallel middle body (Lfo ≤ x ≤ L + Lfo ): in inaccurate calculations. Therefore, the determination of an appro­
priate computational domain is a crucial step in CFD calculation. In this
R(x) = R, (2)
study, the computational domain is selected as shown in Fig. 4. The
Afterbody (L + Lfo ≤ x ≤ L + Lfo + Laf ): velocity entrance is hemispherical with the same diameter as the length
of SUBOFF. The aft body is a cylinder with the same diameter as the

⎧ ( )
1 ( ) 4 ⎫1/2
⎪ 2 2 2 3 2 ⎪
⎨ rh + rh⋅ko⋅ε (x) + 20 − 20rh − 4rh⋅ko − 3 kl ε (x) + − 45 + 45rh + 6rh⋅ko + kl ε (x) ⎪
⎪ ⎬
R(x) = R ( ) (3)
⎪ ⎪
⎩ +( 36 − 36rh2 − 4rh⋅ko − kl)ε5 (x) + − 10 + 10rh2 + rh⋅ko + 1 kl ε6 (x)
⎪ ⎪

3

afore body, while the length twice that of SUBOFF.


where rh is the coefficient of minimal radius of the rear body; ko the tail Besides the computational domain, the quality of grid also exerts an
fat index; kl the tail smoothing index; the length factor ε(x) is expressed important influence on the calculation results, with respect to the
by: convergence and precision. In the study, unstructured grid is adopted
( ) since unstructured grid has the advantages of less human intervention
Lfo + L + Laf − x
ε= , (4) and high degree of automation, which helps an automatic optimization
Lfo implementation. It is noted that one of the tasks of the study is to build

Fig. 5. Boundary meshing.

4
W. Luo et al. Ocean Engineering 230 (2021) 109050

Fig. 6. Suboff surface mesh distribution.

Fig. 7. Grid quality test results.

Fig. 8. Velocity distribution.

up an automatic optimization platform that will be explained subse­ and a fluctuating component. The RANS equation takes the form as
quently. In mesh generation, the closer to the model, the smaller the size (Yakhot et al., 1992):
of grid is. Moreover, mesh refinement is carried out where the cross- ⎫
∂ui
section sharply alters, especially at the end. No slip wall is defined in =0 ⎪


∂xi ⎬
boundary conditions. The wall function method is employed in dealing (6)
( )
with near wall. The initial height of the near-wall grid is selected as y+
0 =
∂ui ∂ui ∂p ∂ ∂ui ′ ′



ρ + ρuj = − +μ − ρui uj + Si i, j ∈ (1, 2, 3) ⎭
45. Meshing results are shown in Fig. 5 and Fig. 6, and the quality check ∂t ∂xj ∂xi ∂xj ∂xj
is reported in Fig. 7. As can be seen from the check results, the grid
quality is mostly concentrated in the range of 0.7–1 or so, no less than where ui and uj are velocity terms; p is the pressure term; ρ is the density
0.3, which demonstrates the validity of meshing. of flow; μ is the dynamic viscosity; Si is a generalized source term; ρui uj is
′ ′

the Reynolds stress.


3.2. Numerical simulation To solve the RANS equation, an additional turbulence model should
be involved. Several turbulence models are available such as SA model,
CFD simulation strategies mainly refer to DNS (direct numerical standard k-ω model, SST k-ω model, standard k-ε model, RNG k-ε model,
simulation), LES (large eddy simulation) and RANS (Reynolds-average Realizable k-ε model, RSM model etc. In the study, SST k-ω is selected,
Navier Stokes) simulation. Owing to its calculation efficiency, RANS which has been confirmed as an appropriate model for SUBOFF by
simulation is in common use. In the study, this method is adopted to comparing different turbulence models applied to calculation of hy­
evaluate the resistance of SUBOFF. Using this method, the standard N–S drodynamic forces of SUBOFF (Bai et al., 2010).
equation is transformed into a Reynolds-averaged governing equation In the numerical simulation, forward movement at a constant speed
by replacing a flow variable with the combination of a mean component is considered for the investigated SUBOFF to demonstrate the CFD based

5
W. Luo et al. Ocean Engineering 230 (2021) 109050

Table 1 rameters involve the maximal radius R, the coefficient of minimal radius
Comparison of CFD values and experimental values of resistance. of the rear body rh, the parallel body length L, tail fat index ko, and tail
velocity(m/s) CFD(N) Experimental value(N) Error smoothing index kl. As can be seen from the equations (1)~(5), the
shape of the model is determined by the above five parameters. For the
3.045511 80.924021 87.4 7.4%
5.14444 227.97766 242.2 5.8% objective-resistance, the main effects are described in the left plot of
6.091022 311.202 332.9 6.5% Fig. 9 (where blue bars indicate a positive correlation; while red bars
7.161067 429.15894 451.5 4.9% indicate a negative correlation). As can be seen, the effect of velocity v is
8.231111 565.58719 576.9 1.9% predominant. Other positive effects are v2 , R2 , R, R − v, L2 , rh2 sequen­
9.151967 698.16092 697.0 0.1%
tially. For the objective-energy consumption, mass is selected as the
objective in parameter sensitivity analysis because of the close rela­
simulation. Fig. 8 presents the velocity distribution and pressure varia­ tionship between mass and energy consumption. The main effects are
tion along with the model. depicted as the right plot in Fig. 9. As can be seen, the variable R exerts a
A comparison of resistance between experiments provided by the major influence on mass. Other positive effects are sequentially rh,R2 ,kl,
Defense Advanced Research Projects Agency and simulation is shown in ko, rh2 . In the later disciplinary optimization (Section 4), the design
Table 1 below. variables will be determined based on the results of parameter sensi­
It can be seen from the above table that the error between the CFD tivity analysis.
simulation value and the experimental value is within the allowable
range of the project (8%).
Although CFD calculation provides a precise tool to calculate the 4.2. Surrogate model
resistance of SUBOFF, it is not appropriate to directly embed the CFD
module into an optimization platform. The main difficulty is time- Surrogate models are constructed using a data-driven, bottom-up
consuming calculation and resultant low efficiency of optimization, approach. The flowchart of approximation is shown as Fig. 10. The
especially when the number of grids is huge. To guarantee both the process involves three major steps which may be interleaved iteratively:
calculation accuracy and optimization efficiency, a surrogate model can collection of sample data, selection of a surrogate model and
be used to approximate the numerical CFD simulation model.

4. Dynamic surrogate model

4.1. Design of experiments

A surrogate model can be viewed as a regression model. It is con­


structed based on limited important samples that can be obtained by
using design of experiment (DOE). DOE provides an effective way to
arrange experiments. It usually involves the selection of input variables,
output variables, and control variables. Using DOE, the number and the
spatial distribution of key samples needed to construct the approximate
model (i.e. surrogate model) can be determined. The study of DOE is an
important topic in optimization design. Several DOE methods are
available such as full factorial design, parameter study, data file,
orthogonal arrays, central composite design, Latin hypercube, and
optimal Latin hypercube etc. In this study, optimal Latin hypercube is
employed owing to the guaranteed space-filling quality.
To determine the important variables, parameter sensitivity analysis
is performed. By this means, the number of design variables can be
reduced. As a result, the DOE efficiency can be improved, meanwhile the
Fig. 10. Flowchart of a surrogate model.
risk of regression error lowers. For the model SUBOFF, the lines pa­

Fig. 9. Sensitivity analysis of design variables.

6
W. Luo et al. Ocean Engineering 230 (2021) 109050

Fig. 11. Error analysis of resistance.

optimization of model parameters, and verification of the surrogate 4.3. Lower confidence bound
model.
The scientific challenge of surrogate modelling is the generation of Since the adoption of a surrogate model in an optimization process
an accurate model using as few simulation evaluations as possible. aims to replace the time-consuming numerical simulation, the optimi­
Several kinds of surrogate models are available such as polynomial zation results largely depends on the approximation accuracy of the
response surface model, radial basis function (RBF) model, kriging surrogate model constructed, which can be improved by a dynamic
model, support vector machine (SVM) model, space mapping model surrogate model. In the dynamic model, samples are updated by using
(Bandler et al., 2004), artificial neural network (ANN) model (Eason and some prescreening method. Commonly used prescreening methods refer
Cremaschi, 2014b), Bayesian network model (Chivatá Cárdenas, 2019), to expected improvement (EI), probability of improvement (PI), most
and random forest model (Dasari et al., 2019). Comparatively, the likely improvement (MI) and lower confidence bound (LCB). For a
polynomial response surface model is preferred in most cases. However, global optimization problem, EI, PI and LCB are preferable. Compara­
the main difficulty with this method is the determination of the order tively, the adaptability of LCB for different optimization scenarios is
and cross-terms of the polynomial function especially for complex better than EI and PI (Liu et al., 2012). In the study, LCB is employed to
nonlinear problems. In the study, RBF surrogate model is adopted. update samples in the surrogate model. The objective function of LCB is
Owing to the ability of interpolation of scattered multivariate data, RBF defined as:
based surrogate model can follow complicated shapes and accurately
σ (x),
flb (x) = ̂y (x) − ŵ (7)
represent complex data (Queipo et al., 2005). A RBF model can be
viewed as a sum of weighted basis functions. The number of weights
where ̂ y (x) is the prediction result by a surrogate model; ̂σ (x) is the
depends only on the number of samples rather than the dimension of
standard deviation; w is the balance parameter to control the tradeoff
design variables. As a result, the calculation efficiency can be improved,
between exploration ability and exploitation ability of LCB method. In
which is good to high dimensional and multivariate problems (Zhou
detail, a larger w implies better exploration ability or global optimum
et al., 2013).
while a smaller w means better exploitation ability or local optimum.
To validate the surrogate model, a classical model validation metric
Instead of a constant like w = 2 commonly used, in the study an adaptive
is employed, i.e. the predictive coefficient of determination, or R2 value.
balance parameter is proposed as:
This validation method analyzes the correlation between the surrogate ⎧
model outputs and expected outputs. A value close to 1 indicates a well- ⎪
⎪ 3, α > 3̂σ (x)

designed surrogate model. In the study, Figs. 11 and 12 show the vali­ w=
2, σ (x) ≤ α ≤ 3̂
2̂ σ (x)
(8)
dation results of surrogate models with respect to resistance and energy ⎪


1, σ (x) ≤ α < 2̂
1̂ σ (x)
consumption respectively. As seen, the R2 value in resistance approxi­ min{xd }, α < 1̂ σ (x)
mation is 0.98711, and 0.96904 for energy consumption approximation,
which implies the high confidence of the surrogate models designed in where α is the error between preidicted value and true value, i.e. α =
the study. y (x) − y(x). xd is the Euclidean distance between two sample points. It
̂

7
W. Luo et al. Ocean Engineering 230 (2021) 109050

Fig. 12. Error analysis of energy consumption.

can be inferred from Eq. (4) and Eq. (5) that, if the error is large, the is. Consequently, the trust region can be extended, as reflected by the
solution searching emphasizes more in a global range while as the error regulation of parameters in Eq. (7): usually r1 < r2 < 1, c1 < c2.
decreases the solution searching emphasizes more in a local range and
the improvement of the solution accuracy. By using the adaptive balance 4.5. Verification of dynamic surrogate model
parameter, the global optimal solution as well as its accuracy can be
guaranteed. To demonstrate the effectiveness of the proposed dynamic sug­
grogate model, a representative test example in MDO field, a speed
4.4. Trust region reducer (Kodiyalam, 1998), is taken as the verification model. The
optimization problem is formulated as:
To improve the stability of solution and convergence rate, an adap­ ⎧

tive trust region is used in dynamic surrogate model. Since a surrogate ⎪ min : f = 0.7854⋅x1 ⋅x2 ⋅( 3.3333⋅x2 + 14.9334⋅x3 − 43.0934)



⎪ 2 3
model is a good approximation only in a neigbourhood of xk at k itera­ ⎪ − 1.5079⋅x ⋅( x2 + x2 ) + 7.477⋅( x3 + x3 ) + 0.7854⋅( x ⋅x2 + x ⋅x2 )


⎪ 1 4 6 5 7
y (xk ) ≈ y(xk ), the iteration step in an optimzation process
tion, i.e. ̂ ⎪ 6 7 6 7
⎪ s.t. : g = 27/( x ⋅x2 ⋅x ) − 1.0 ≤ 0; g = 397.5/( x ⋅x2 ⋅x2 ) − 1.0 ≤ 0



should be restricted in a “trust region”, i.e. ⎪ 1 1 2 3 2 1 2 3
⎪ g = 1.93⋅x3 /( x ⋅x ⋅x4 ) − 1 ≤ 0; g = ( 1.93⋅x3 )/( x ⋅x ⋅x4 ) − 1 ≤ 0



⎪ 3 4 2 3 6 4 5 2 3 7
⎪ /
‖s‖ ≤ δk , (9) ⎪


√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
( 2 2
)/( 2 3 ) 6
( 3
)
g5 = 745 ⋅x4 x2 ⋅x3 + 16.9⋅10 110⋅x6 − 1 ≤ 0
where s denotes the iteration step. ‖ ⋅‖ is the notation of Euclidean norm; ⎪

⎪ √ ̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
( )/( 2 2 ) ̅ /
( )


δk is the trust radius and updates according to the approximation ac­ ⎪

⎪ g6 = 7452 ⋅x25 x2 ⋅x3 + 157.5⋅106 85⋅x37 − 1 ≤ 0

curacy as follows (Alexandrov et al., 1998): ⎪




⎪ g7 = x2 ⋅x3 /40 − 1 ≤ 0; g8 = x1 /x2 − 12.0 ≤ 0
⎧ ⎪

⎨ c1 ‖s‖, r < r1 ⎪
⎪ g = 5 − x1 /x2 ≤ 0; g10 = (1.5⋅x6 + 1.9)/x4 − 1 ≤ 0

⎪ 9
δk = min(c2 ‖s‖, Δ* ), r > r2 (10) ⎪


‖s‖, otherwise ⎩ g11 = (1.1⋅x7 + 1.9)/x5 − 1 ≤ 0

(12)
where r1, r2, c1, c2 are constants; Δ* is an upper bound of trust radius; r is
a factor describing how well the surrrogate model approximates a high- where the objective f is the weight of the speed reducer, xi is a design
fidelity model (CFD numerical model in the study) in a first-order sense, variable, gi is a constant. The theoretical optimum of f is 2994.
defined as: Several other surrogate models are used to be compared with the
proposed dynamic surrogate model, including a dynamic radial basis
y(xk ) − y(xk + s)
r= . (11) function (DRBF) model (Peng et al., 2011), a lower confidence bound
y(xk ) − ̂
y (xk + s)
dynamic radial basis function (LCB-DRBF) model (Dennis and Torczon,
In general, the closer to 1 r is, the more accurate the surrogate model 1997), and a trust region-dynamic radial basis function (TR-DRBF)

8
W. Luo et al. Ocean Engineering 230 (2021) 109050

Table 2
Optimization of a speed reducer using different surrogate models.
F ​ = 0.6Fd + 0.4Ne (13)

Surrogate model Number of iteration Optimal result of objective Error where Fd denotes the resistance and Ne denotes the energy consumption
Proposed 113 2994.2 0.006% which can be theoretically determined by
DRBF 432 3187.5 6.5%
LCB-DRBF 287 2725.6 9.0% ρCx ΩT v3
Ne = (14)
TR-DRBF 181 3035.3 1.4% 2ηp

where ρ is the fluid density; Cx the resistance coefficient; ΩT the area of


mdoel (Long et al., 2014). Table 2 lists the optimization results and
wetted surface of the underwater vehicle; ηp the propeller efficiency
comparison results. As can be seen, the optimization efficiency and ac­
(Baowei et al., et al.). In the study, the propeller efficiency is assumed as
curacy by using the proposed model are better than the other models.
a constant being 0.96.
It is noted that through Eq. (14), it can be inferred that Ne is strongly
5. Isight integration optimization
linked to Fd. Nevertheless, in the study they are treated as different
5.1. Construction of optimization platform disciplines. The reason is that firstly, as seen from the parameter sensi­
tivity analysis (in subsection 3.1), different effects are determined for
With the development of computer technology, some integrated resistance and energy consumption, respectively. In detail, the
optimization platforms are created such as Isight, Dakota, ModelCenter, maximum radius R and the parallel length L are considered the main
VisualDOC etc. In the study, Isight platform is adopted to conduct the effects for resistance. While for energy consumption, the maximum
design optimization of the lines of SUBOFF model. radius R, the coefficient of minimal radius of the rear body rh, the tail fat
Isight originates from GE’s powerful CAO (computer aided optimi­ index ko, and the tail smooth index kl are considered the main effects.
zation) platform. It provides a library of components to accomplish the Thus, in a CO framework, there is difference between resistance disci­
simulation process of design optimization. The main components refer pline and energy-consumption discipline, in terms of both objective and
to simcode component, calculator component and optimization design variables. Secondly, RBF surrogate models are constructed for
component. Specific software include three-dimensional modeling resistance and energy consumption respectively. As aforementioned, the
software creo3.0, pre-processing software icem15.0 and simulation design variables are different for resistance and energy consumption,
software fluent15.0. A specific software can be automatically executed which implies different surrogate models for resistance and energy
through simcode component batch file. For example, in the study 3D consumption although energy consumption is strongly linked to
modeling of SUBOFF is conducted by using creo3.0. When the modeling resistance.
is finished, Isight sends the command to icem to start the pre-simulation At discipline level, according to the results of parameter sensitivity
(i.e. meshing) and then output the mesh file. Next, Isight starts Fluent analysis in section 3.1, the maximum radius R and the parallel length L
batch file and automatically runs Fluent to carry out numerical simu­ are selected as the design variables in the resistance discipline and
lation. After calculation, the results (such as the force, moment, and correspondingly the objective function can be designed as:
pressure impacted on SUBOFF) are passed to the next component, i.e. f1 = (R − R1 )2 + (L − L1 )2 (15)
the optimization component. In that way, new lines of SUBOFF can be
obtained and the modeling-calculation-optimization process will be for a given velocity. For energy-consumption discipline, the maximum
repeated until the pre-defined objective function is achieved. Isight radius R, the coefficient of minimal radius of the rear body rh, the tail fat
platform integrates various kinds of software by building blocks; com­ index ko, and the tail smooth index kl are selected as the design variables
bines all design flows into a unified, organic and logical framework; and the objective function can be designed as:
automatically runs simulation and restarts the design flow. As a result, it
gains over traditional optimization design especially in terms of f2 = (R − R2 )2 + (rh − rh2 )2 + (ko − ko2 )2 + (kl − kl2 )2 (16)
efficiency. It is noted that in Eqs. (15) and (16), the variables with subscripts 1
In the study, at system-level, the optimization goal is designed as: and 2 represent the design variables at discipline-level, while the

Fig. 13. Isight optimization platform.

9
W. Luo et al. Ocean Engineering 230 (2021) 109050

Fig. 14. Optimization process of the objective and design variables.

10
W. Luo et al. Ocean Engineering 230 (2021) 109050

Table 3 5.2. Optimization results


Optimization results.
initial range optimal value by optimal value by The stopping rule for optimization is set as: when the difference
value SSM DSM between current optimization result of objective ⃒ and the
⃒ historical op­
Kl 44.6244 [43, 45] 44.4819 44.4123 timum of objective is less than 0.001, i.e. if ⃒Fop − F(k)⃒ < 0.001, opti­
Ko 10 [9, 11] 11 10.5101 mization stops. It is noted that the historical optimum Fop can be
Rh 0.1175 [0.1,0.2] 0.1749 0.1683 detected automatically by the Isight software. Moreover, the optimum is
L (ft) 10.6459 [10, 11] 10.0577 10.0577 guaranteed by the proposed hybrid optimization algorithm, i.e. ASA +
R (ft) 0.8333 [0.8033, 0.8625 0.8572
0.8633]
MMFD. As aforementioned, ASA guarantees globally optimal solution
Fd (N) 977.1870 884.4318 865.7660 while MMFD guarantees the optimization efficiency. Fig. 14 presents the
(reduced by (reduced by optimization processes and results. The velocity of the underwater
9.4%) 11.4%) vehicle is selected as 19knot (9.7922 m/s). As can be recognized from
Ne (N⋅m/s) 8787.7435 8516.8326 8426.1457
Table 1, in general the accuracy of CFD calculation increases with an
(reduced by (reduced by
3.2%) 4.1%) increase of velocity. Compared with the velocity in the last row in
Number of 941 700 Table 1, the selected velocity 9.7922 m/s in optimization increases by
iteration around 7%. In the plots in Fig. 14, the black points represent feasible
solutions; red points denote infeasible solutions; while green point in­
dicates the optimal solution.
variables without subscripts represent desired design variables allocated
Table 3 lists the initial values, range and optimal values of design
from system-level.
variables and objectives in which kl, ko and rh are non-dimensional
The Isight structure is shown as Fig. 13. The modules including Creo,
variables. It should be noted that the initial values and setting of
Icem, Fluent, and Matlab are incorporated. In the Task block (the top
design space are based on the fact that for the investigated underwater
block) one can define the optimization goals, design variables and
vehicle a series of experiments with different lines were performed. The
constraints. Optimization of resistance-discipline (denoted by
design space reflects the range of lines. Both the lower bound and upper
optimization-1 in Fig. 13) and optimization of energy-consumption-
bound of the range of a design variable have corresponding experi­
discipline (denoted by optimization-2 in Fig. 13) are executed in par­
mental results, so do the initial values of design variable. CFD calcula­
allel. As seen, in an optimization loop data flow starts from Creo for
tion and optimization within the experimental range make the results
modeling, to Icem for meshing, to Fluent for hydrodynamic calculation,
more credible.
to Matlab for dynamic surrogate model construction and subsequent
Comparison between static surrogate model (SSM) and the proposed
optimization. Calculation results are fed back to the Task block. Design
dynamic surrogate model (DSM) is also presented in Table 3. As can be
variables might be modified if the optimization goals are not achieved.
recognized, both the resistance (Fd) and energy consumption (Ne)
Afterwards, updated design variables would be transferred to the Creo
descend on the basis of optimal lines, no matter which surrogate model
module and optimization process would be repeated. The detailed pro­
is employed. Moreover by comparison, the optimization effect based on
cedure of optimization can be described as
DSM outperforms SSM, in terms of both minimization of objectives and
number of iteration.
(1) According to the mathematical model of SUBOFF, a 3D model is
The optimized hull shape is shown in Fig. 15 and compared with the
constructed by using Creo. Corresponding model file with an
shape with original lines. A distinct difference can be recognized by
extension of.stp can be obtained at this stage.
observing the sterns. It is noted that the investigated SUBOFF hull can be
(2) Based on the obtained model file, meshing is performed by using
described by mathematical expression (Groves et al., 1989). The vari­
Icem. Corresponding mesh file with an extension of.msh can be
ables in the mathematical model are the design variables in the opti­
obtained at this stage.
mization problem described in the manuscript, including the maximal
(3) Based on the obtained mesh file, CFD is carried out by using
radius R, the coefficient of minimal radius of the rear body rh, the par­
Fluent. The calculation results are recorded in a data file with an
allel body length L, tail fat index ko, and tail smoothing index kl.
extension of.txt.
Therefore, optimization of those design variables means the lines design
(4) Based on the CFD calculation results, surrogate models are con­
of SUBOFF.
structed in Matlab. Optimization of the SUBOFF lines is con­
ducted in combination of obtained surrogated models.
(5) The above steps are repeated before optimization stopping rule is
satisfied.

Fig. 15. Hull shape of SUBOFF.

11
W. Luo et al. Ocean Engineering 230 (2021) 109050

6. Conclusions Balling, R., Wilkinson, C., 1997. Execution of multidisciplinary design optimization
approaches on common test problems. AIAA J. 35 (1), 178–186.
Bandler, J.W., Cheng, Q.S., Dakroury, S.A., et al., 2004. Space mapping: the state of the
In the study of application of MDO to an underwater vehicle, con­ art[J]. IEEE Trans. Microw. Theor. Tech. 52 (1), 337–361.
cerns are mainly focused on the construction of hybrid optimization Baowei S Wei D, Zhiyong G, et al. Multidisciplinary design optimization of torpedo
strategy in a CO framework and surrogate models to approximate the general design based on collaborative optimization[J]. Torpedo Technol., 17(6):7-
11.
CFD calculation. Through the optimization results it can be concluded Bidoki, M., Mortazavi, M., Sabzehparvar, M., 2018. A new approach in system and tactic
that: design optimization of an autonomous underwater vehicle by using
Multidisciplinary Design Optimization. Ocean. Eng. 147, 517–530.
Braun, R.D., Moore, A.A., Kroo, I.M., 1997. Collaborative architecture for launch vehicle
(1) A combination of global optimization algorithm and gradient- design. J. Spacecraft Rockets 34 (4), 478–486.
based optimization can improve the optimization efficiency and Cao, A.X., Zhao, M., Liu, W., et al., 2007. Application of multidisciplinary design
guarantee globally optimal solution as well. optimization in the conceptual design of a submarine. J. Ship Mech. 11 (3), 373–382.
Cheng G, Wang G. Trust Region Based MPS Method for Global Optimization of High
(2) Optimal Latin hypercube based DOE combined with RBF model Dimensional Design Problems, Proceeding of 20th AIAA/ASME/AHS Adaptive
provides an effective way to construct a surrogate model to Structures Conference. April 2012, Honolulu, Hawaii, America, pp.1-8.
approximate the CFD calculation for an underwater vehicle like Chivatá Cárdenas, I., 2019. On the use of Bayesian networks as a meta-modelling
approach to analyse uncertainties in slope stability analysis[J]. Georisk 13 (1),
SUBOFF. 53–65.
(3) A dynamic surrogate model strategy updates the samples and Dasari, S.K., Cheddad, A., Andersson, P., 2019. Random Forest Surrogate Models to
regenerates timely during an optimization process, which can Support Design Space Exploration in Aerospace Use-Case[C]//IFIP International
Conference on Artificial Intelligence Applications and Innovations. Springer, Cham,
improve both the optimization accuracy and efficiency,
pp. 532–544.
compared with static surrogate models. Dennis, J.E., Torczon, V., 1997. Managing approximation models in optimization[J].
Multidiscipl. Design Optimiz.: State-of-the-Art 5, 330–347.
In the study, only the main body is considered. Appendages like fins Eason, J., Cremaschi, S., 2014a. Adaptive sequential sampling for surrogate model
generation with artificial neural networks. Comput. Chem. Eng. 68 (68), 220–232.
will be taken into account in calculating hydrodynamics in the next Eason, J., Cremaschi, S., 2014b. Adaptive sequential sampling for surrogate model
work. In addition, more hydrodynamic performances such as maneu­ generation with artificial neural networks[J]. Comput. Chem. Eng. 68, 220–232.
verability and seakeeping will be involved in the CFD module. Valida­ Groves, N., Huang, T.T., Chang, M.S., 1989. Geometric Characteristics of DARPA
SUBOFF Models (DTRC Models Nos. 5470 and 5471), pp. 1–75. Technical Report
tion of the CFD simulation work by field test is also a next plan. DTRC/SHD 1298-01.
Comparison of different combined optimization algorithms will also be Gu, H., Yang, L., Hu, Z., et al., 2009. Surrogate Models for Shape Optimization of
studied. Moreover, construction of the objective at system-level will be Underwater Glider.//2009 International Conference on Computer Modeling and
Simulation. IEEE, pp. 3–6.
studied. It is noted that in the study the balance between resistance and Hosseini, M., Nosratollahi, M., Sadati, H., 2017. Multidisciplinary design optimization of
energy consumption, as shown in Eq. (13), is not supported by theory or UAV under uncertainty[J]. J. Aero. Technol. Manag. 9 (2), 169–178.
experiments, but by trials. In next work, such an issue will be studied, Joung, T.H., Sammut, K., He, F., et al., 2012. Shape optimization of an autonomous
underwater vehicle with a ducted propeller using computational fluid dynamics
which aims to theoretically determine the weights of disciplines when analysis. Int. J. Naval Architect. Ocean Eng. 4 (1), 45–57.
defining the system-level objective. Kirkpatrick, S., Gelatt, C.D., Vecchi, M.P., 1983. Optimization by simulated annealing[J].
Science 220 (4598), 671–680.
Kodiyalam, S., 1998. Evaluation of Methods for Multidisciplinary Design Optimization
CRediT authorship contribution statement (MDO), Phase I[R], NASA/CR-1998-208716.
Kroo, I., Altus, S., Braun, R., et al., 1994. Multidisciplinary Optimization Methods for
Weilin Luo: Conceptualization, Methodology, Formal analysis. Aircraft Preliminary Design//5th Symposium on Multidisciplinary Analysis and
Optimization. Panama City Beach, Florida, pp. 697–707.
Xiaoming Guo: Writing – original draft, Investigation. Jiawei Dai:
Lavelle, T., Plencner, R., Seidel, J., 1991. Concurrent Optimization of Airframe and
Visualization, Methodology. Taichun Rao: Data curation, Software. Engine Design parameters[C]//4th Symposium on Multidisciplinary Analysis and
Optimization, p. 4713.
Lin, J.G.G., 2004. Analysis and enhancement of collaborative optimization for
Declaration of competing interest multidisciplinary design[J]. AIAA J. 42 (2), 348–360.
Lin, P.T., Gea, H.C., 2013. A gradient-based transformation method in multidisciplinary
The authors declare that they have no known competing financial design optimization. Struct. Multidiscip. Optim. 47 (5), 715–733.
Liu, H.L., Huang, T.T., 1998. Summary of DARPA SUBOFF Experimental Program Data.
interests or personal relationships that could have appeared to influence NAVAL SURFACE WARFARE CENTER CARDEROCK DIV BETHESDA MD
the work reported in this paper. HYDROMECHANICS DIRECTORATE.
Liu, K., Luo, W., 2016. Design of the lines of underwater vehicles based on collaborative
optimization. J. Mar. Sci. Technol. 21 (4), 709–714.
Acknowledgement Liu, B., Zhang, Q., Fernández, F.V., et al., 2012. Self-adaptive Lower Confidence Bound:
A New General and Effective Prescreening Method for Gaussian Process Surrogate
Authors are thankful for the support by China Fujian Provincial Model Assisted Evolutionary algorithms[C]//2012 IEEE Congress on Evolutionary
Computation. IEEE, pp. 1–6.
Department of Science and Technology, Grant 2020R0169. Liu, X., Yuan, Q., Zhao, M., et al., 2017. Multiple objective multidisciplinary design
optimization of heavier-than-water underwater vehicle using CFD and
References approximation model. J. Mar. Sci. Technol. 22 (1), 135–148.
Long, T., Guo, X., Peng, L., Liu, L., 2014. Optimization strategy using dynamic radial
basis function metamodel based on trust region. J. Mech. Eng. 50 (7), 184–190.
Adelman, H.M., Mantay, W.R., 1991. Integrated multidisciplinary design optimization of
Luo, W., Lyu, W., 2015. An application of multidisciplinary design optimization to the
rotorcraft[J]. J. Aircraft 28 (1), 22–28.
hydrodynamic performances of underwater robots. Ocean. Eng. 104 (23), 686–697.
Alam, K., Ray, T., Anavatti, S.G., 2012. A new robust design optimization approach for
McAllister, C., Simpson, T., Kurtz, P., et al., 2002. Multidisciplinary design optimization
unmanned underwater vehicle design. Proc. IME M J. Eng. Marit. Environ. 226 (3),
testbed based on autonomous underwater vehicle design. In: //9th AIAA/ISSMO
235–249.
Symposium on Multidisciplinary Analysis and Optimization, p. 5630.
Alam, K., Ray, T., Anavatti, S.G., 2014. Design and construction of an autonomous
Peng, L., Liu, L., Long, T., 2011. Optimization strategy using dynamic radial basis
underwater vehicle. Neurocomputing 142, 16–29.
function metamodel[J]. J. Mech. Eng. 47 (7), 164–170.
Alexandrov, N.M., Lewis, R.M., 2002. Analytical and computational aspects of
Pérez, V.M., Renaud, J.E., Watson, L.T., 2002. Adaptive experimental design for
collaborative optimization for multidisciplinary design [J]. AIAA J. 40 (2), 301–309.
construction of response surface approximations. AIAA J. 40 (12), 2495–2503.
Alexandrov, N.M., Dennis, J.E., Lewis, R.M., et al., 1998. A trust-region framework for
Queipo, N.V., Haftka, R.T., Shyy, W., et al., 2005. Surrogate-based analysis and
managing the use of approximation models in optimization[J]. Struct. Optim. 15 (1),
optimization[J]. Prog. Aero. Sci. 41 (1), 1–28.
16–23.
Rafique, A.F., LinShu, H., Zeeshan, Q., et al., 2009. Multidisciplinary Design of Air-
Alvarez, A., Bertram, V., Gualdesi, L., 2009. Hull hydrodynamic optimization of
Launched Space Launch Vehicle Using Simulated annealing[C]//Annual Conference
autonomous underwater vehicles operating at snorkeling depth. Ocean. Eng. 36 (1),
on Artificial Intelligence. Springer, Berlin, Heidelberg, pp. 719–726.
105–112.
Secanell, M., Suleman, A., 2005. Numerical evaluation of optimization algorithms for
Bai, T., Liang, Z., Zhou, Y., 2010. Comparison and application of turbulence modes in
low-Reynolds-number aerodynamic shape optimization [J]. AIAA J. 43 (10),
submarine maneuvering hydrodynamic forces computation[J]. Chin. J. Ship Res. 5,
2262–2267.
22–28.

12
W. Luo et al. Ocean Engineering 230 (2021) 109050

Sobieski, J., 1989. A Step from Hierarchic to Non-hierarchic System. NASA-CP-3031. Venter, G., 2010. Review of Optimization Techniques [J]. Encyclopedia of Aerospace
Part I. NASA, Virginia, pp. 51–78. Engineering.
Sobieski, J., Haftka, R., 1997. Multidisciplinary aerospace design optimization: survey of Vlahopoulos, N., Hart, C.G., 2010. A Multidisciplinary design optimization approach to
recent developments. Struct. Optimiz. 14 (1), 1–23. relating affordability and performance in a conceptual submarine design. J. Ship
Sobieszczanski-Sobieski, J., 1982. A Linear Decomposition Method for Large Prod. Design 26 (4), 273–289.
Optimization Problems. A Blueprint for Development. NASA TM-83248. Yakhot, V., Orszag, S.A., Thangam, S., et al., 1992. Development of turbulence models
Sobieszczanski-Sobieski, J., Agte, J.S., Sandusky, R.R., 2000. Bilevel integrated system for shear flows by a double expansion technique[J]. Phys. Fluids 4 (7), 1510–1520.
synthesis. AIAA J. 38 (1), 164–172. Yang, R.J., Gu, L., Tho, C.H., et al., 2002. Reliability-based Multidisciplinary Design
Song, L., Wang, J., Yang, Z.Y., 2013. Research on shape optimization design of Optimization of a Full Vehicle System.//43rd AIAA/ASME/ASCE/AHS/ASC
submersible based on Kriging model. J. Ship Mech. 17, 8–13. Structures, Structural Dynamics, and Materials Conference, p. 1758.
Su, G., Peng, L., Hu, L., 2017. A Gaussian process-based dynamic surrogate model for Yu, M., Li, X., Liang, J., 2020. A dynamic surrogate-assisted evolutionary algorithm
complex engineering structural reliability analysis. Struct. Saf. 68, 97–109. framework for expensive structural optimization. Struct. Multidiscip. Optim. 61 (2),
Sun, C., Song, B., Wang, P., 2015. Parametric geometric model and shape optimization of 711–729.
an underwater glider with blended-wing-body. Int. J. Naval Architect. Ocean Eng. 7 Zhang, Y., Li, G.Y., Sun, G.Y., et al., 2008. Application research on multidisciplinary
(6), 995–1006. design optimization of the full vehicle light weight. China Mech. Eng. 19 (7),
Valerie Michelle, Manning, 1999. Large-scale Design of Supersonic Aircraft via 877–881.
Collaborative Optimization. Ph.D. Thesis. Stanford University. Zhou, L.R., Yan, G.R., Ou, J.P., 2013. Response surface method based on radial basis
Vasudev, K.L., Sharma, R., Bhattacharyya, S.K., 2014. A multi-objective optimization functions for modeling large-scale structures in model updating[J]. Comput. Aided
design framework integrated with CFD for the design of AUVs. Methods Oceanogr. Civ. Infrastruct. Eng. 28 (3), 210–226.
10, 138–165.

13

You might also like