You are on page 1of 14

Aerospace Science and Technology 112 (2021) 106639

Contents lists available at ScienceDirect

Aerospace Science and Technology


www.elsevier.com/locate/aescte

Data-based approach for wing shape design optimization


Jichao Li ∗ , Mengqi Zhang
Department of Mechanical Engineering, National University of Singapore, Singapore 117575, Republic of Singapore

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

Article history: Aircraft design is a trade-off among different objectives and constraints, so multiple design rounds are
Received 20 September 2020 usually required. Aerodynamic shape optimization based on high-fidelity computational fluid dynamics is
Received in revised form 25 February 2021 consequently expensive, especially for wing design in the transonic regime. To address the interactive
Accepted 3 March 2021
demand in aircraft design, we present a data-based approach to real-time high-fidelity wing shape
Available online 16 March 2021
Communicated by Qiulin Qu
optimization. Accurate and fast data-based models are constructed to perform aerodynamic analyses
in lieu of costly computational fluid dynamic simulations. For the versatility of the data-based models,
Keywords: 135,108 training data that cover wing samples of different aerodynamic shapes, flight speeds, and flight
Aerodynamic shape optimization altitudes are used. The verification on 47,967 wings shows that mean relative errors of C L , C D , and
Data-based model C M compared to computational fluid dynamic simulations are all within 0.4%. The models are further
Aircraft wing design verified in multiple single-point, multi-point, and multi-objective wing design optimization problems.
The optimized wings have similar shapes to those obtained by computational-fluid-dynamics-based
optimization, and the differences in C D are merely one ∼ two counts. These results demonstrate the
effectiveness of the data-based approach to fast and high-fidelity wing design. This work showcases a
real-time high-fidelity optimization approach to high-dimensional nonlinear engineering problems using
data-based models.
© 2021 Elsevier Masson SAS. All rights reserved.

1. Introduction tion [11–14], are the most popular in wing shape design. The
CFD-based optimization method usually leverages the efficient ad-
Wing shape design optimization is critical to the aerodynamic joint method [15,16,4] and uses a gradient-based algorithm in
performance of aircraft. With the increasing emphasis on envi- the design optimization. This method has been successfully ap-
ronmental protection and flight economy, fuel burn reduction be- plied to aerodynamic shape design optimization of wings [2,17],
comes one of the leading factor in aircraft design. The major wing-body-tail configurations [18,19], and new conceptional air-
approach to improving cruise efficiency, in addition to using ad- craft [20,21]. In cases without an adjoint solver, surrogate-based
vanced aero engines and materials, is to perform wing shape de- optimization [3,22,23] can be a good alternative. In surrogate-
sign optimization. Thus, in the highly competitive aviation market, based optimization, analyses of aerodynamic functions are quickly
the success of a new commercial aircraft cannot be separated from performed by using cheap surrogate models. Then gradient-based
a well-designed wing. or gradient-free optimization algorithms can be used to solve the
Modern aerodynamic shape optimization techniques [1–3] design optimization problem [24]. Note that CFD analyses are still
leverage computational fluid dynamics (CFD) simulations. Com- required in surrogate-based optimization when constructing and
mercial aircraft design leads to wing shape optimization in the refining the surrogate models. One of the most popular choices
transonic regime, which requires high-fidelity CFD analyses to in surrogate-based optimization is the efficient global optimization
capture shock waves and laminar-turbulent transitions [4,5]. A (EGO) method [25], which has been applied to airfoil and wing
large number of shape design variables are usually required in shape design [26].
transonic wing design because the aerodynamic performance is Multiple optimization rounds are generally required in practi-
sensitive to small geometric modifications. Among different aero- cal wing design to comprise different constraints and objectives. It
dynamic optimization methods [6–8], two optimization methods, is highly useful to have a fast and high-fidelity wing design opti-
CFD-based optimization [9,2,10] and surrogate-based optimiza- mization approach. However, current design methods are iterative
and need to call the costly CFD solver in each iteration. Thus, wing
design optimization usually requires hours or days to run even on
* Corresponding author. high-performance computing (HPC) clusters. For airfoil design op-
E-mail address: cfdljc@gmail.com (J. Li). timization, Li et al. [27] developed a data-based approach that can

https://doi.org/10.1016/j.ast.2021.106639
1270-9638/© 2021 Elsevier Masson SAS. All rights reserved.
J. Li and M. Zhang Aerospace Science and Technology 112 (2021) 106639

optimize arbitrary airfoils in seconds with the fidelity of CFD. Such Table 1
an approach has not been applied to wing design because much Bounds of key parameters in different flight missions.
more design variables are required to use, which would bring Mach number Altitude (m) Mass (kg)
too much difficulty to the training of an accurate aerodynamic Lower bound 0.82 9,000 180,000
model. For wings, the state-of-the-art surrogate aerodynamic mod- Upper bound 0.88 12,000 220,000
els could merely account for several primary fluid parameters such
as the angle of attack, the Mach number, and the Reynolds number.
It is of high interest for the aircraft design community to develop
data-based aerodynamic wing models that can fast analyze the ef-
fects of aerodynamic shape variables.
We seek to address the demand for real-time high-fidelity
wing shape design optimization by utilizing our recently developed
deep-learning-based wing mode parameterization method [3,28]
and deep-learning techniques. The wing mode parameterization
method defines a compact design space for wing shape optimiza-
tion by extracting global wing modes. Thus fewer shape design
variables are required. Moreover, wing sample shapes in the com- Fig. 1. Geometry and FFD control points of the CRM wing.
pact design space are much more realistic than those picked from
design spaces defined by conventional parameterization methods
We assume that the wings produce 87% lift of the aircraft based on
like the free-form deformation (FFD) [29]. This is valuable to the
the analysis in ADODG Case 4.2. The other lift is provided by the
training of aerodynamic models [27]. About 140,000 wings are
fuselage and tail. Thus, the lift required on each wing is:
gradually generated via a multi-round sampling strategy, and they
are further analyzed using a CFD solver. With these CFD analysis L = 0.87 × W × g × 0.5, (1)
results as the training data, we construct data-based models for
C L , C D , C M , and C P distributions using the deep neural network where g is the gravitational acceleration. Thus, the lift constraint
(DNN). Multiple tests are performed to determine suitable hyper- in the optimization is
parameters in DNN. The accuracy of the data-based models is veri-
fied in another dataset composed of 47,967 wings. Then, we couple
C Lcon = L /(0.5ρ U 2 S ), (2)
the models in a gradient-based optimization framework to perform where ρ , U , S are the air density, cruise speed, and wing area.
fast wing shape design. The performance of data-based optimiza- ρ and g are functions of the altitude; U is a function of the al-
tion is verified in single-point, multi-point, and multi-objective titude and Mach number. We define them by consulting the ICAO
wing shape design problems. We also show some innovative us- Standard Atmosphere. To ensure suitable cruise lifts in different
age of the data-based optimization method in wing design that missions, the angle of attack is allowed to vary in [0.5◦ , 3.5◦ ];
conventional optimization cannot handle. The relative wing twists to the CRM wing are allowed to vary in
The rest of this paper is organized as follows. First, we describe [−1.0◦ , 1.0◦ ].
the wing design optimization problem and discuss the selection As shown in Fig. 1, we use the FFD method implemented in
of design variables in Section 2. Then, in Section 3, we present pyGeo [29] to manipulate the wing shape. However, using FFD con-
the data-based models and verify their accuracy. In Section 4, we trol points as shape design variables would involve unnecessary
discuss the data-based wing design optimization framework and geometric freedom and thus leads to abnormal shapes in wing
compare its performance with CFD-based optimization in differ- samples. As stated in [27], this is the major obstacle to training
ent wing design problems. In Section 5, we show some innovative accurate aerodynamic models. To address this issue, we use the
usages of the data-based optimization in wing design. Finally, in global wing modal parameterization method proposed by Li and
Section 6, we end this paper with a summary of conclusions. Zhang [3] to extract compact wing shape design variables.
The modal parameterization method is an efficient approach to
2. Problem description aerodynamic shape design optimization [30]. Airfoil mode shapes
have been successfully used for airfoil [31,27] and wing section
We consider a transonic wing shape design optimization prob- shape design optimization [32]. Nevertheless, global wing mode
lem defined by the American Institute of Aeronautics and Astro- shapes were missing due to the absence of public geometric infor-
nautics (AIAA) Aerodynamic Design Optimization Discursion Group mation of wings. Li and Zhang [3] addressed this issue by propos-
(ADODG), where the Common Research Model (CRM) wing is opti- ing the deep-learning-based optimal sampling method to gener-
mized to reduce the drag in different flight missions. The prob- ate realistic wing samples in the desired design space. Then, the
lem is representative in commercial aircraft design. CFD-based wing mode shapes are derived using the singular value decom-
position (SVD) from the wing samples. The details of this process
optimization [2] and surrogate-based optimization [3] have been
are provided in the Appendix. We generate one thousand deep-
shown effective in single-point and multi-point design wing de-
learning-based wing samples to extract wing mode shapes and the
sign. However, both methods need to be coupled with the CFD
first eight modes are shown in Fig. 2. The contours represent the z
solver and the computational cost increases significantly with the
displacements of each mode applied to the baseline CRM wing.
rise of flight missions or design points considered.
The number of shape design variables to use is always a com-
This work aims to realize fast wing shape design optimization
promise in data-based aerodynamic optimization. Using too few or
using the data-based approach. In order to construct a generic
too many design variables would result in ineffective optimization
data-based aerodynamic model, a large range of possible flight or inaccurate data-based models, respectively. As the mathematic
missions of the CRM wing are taken into consideration. The bounds nature of SVD modes, the importance of high-order modes gradu-
of three key parameters of the flight missions are shown in Ta- ally decreases. Li and Zhang [3] used 40 wing modes and achieved
ble 1. Other parameters, for example, the Reynolds number and similar optimization results to those using 192 FFD control points.
cruise lift coefficient, can be computed based on the key parame- To further investigate the influence, we perform wing shape de-
ters. The mass shown in the table stands for the total aircraft mass. sign optimization using different numbers of wing modes at nine

2
J. Li and M. Zhang Aerospace Science and Technology 112 (2021) 106639

Fig. 2. Global mode shapes of the CRM wing. Linear combination of these orthogonal modes is an efficient approach to the wing shape parameterization. (For interpretation
of the colors in the figure(s), the reader is referred to the web version of this article.)

design points specified in ADODG Case 4.5. The design points rep- in IDWarp1 to deform the volume mesh. In the CFD analysis of
resent typical flight missions with different M and C Lcon at an each wing sample, we use ADflow [34,35] to solve the Reynolds-
altitude of 11,740 m. The geometric freedom of the design space averaged Navier–Stokes equations with a Spalart–Allmaras turbu-
increases with the number of wing modes used. Consequently, as lence model [36]. The C P distributions on some wing samples
shown in Fig. 3, more drag reductions are obtained by using more solved by ADflow are shown in Fig. 4. The domains with M > 1.0
wing modes. Nevertheless, using 40-50 wing modes almost fully are highlighted to show the strength of shock waves. These CFD
taps the potential of the modal shape parameterized method, and analyses indicate the complex aerodynamics in the wing shape de-
further increasing the number of modes does not significantly im- sign space.
prove the optimization results. This argument holds at different The process of choosing samples from the design space is called
design points. Thus, we chose to use 50 wing modes as the shape the design of experiment (DoE). We use a multiple round DoE
design variables in this work. strategy to gradually generate wing samples, which is along with
the training and validation of data-based aerodynamic models.
3. Data-based aerodynamic models Compared with the adaptive sampling algorithms based on the
expected improvement (EI) criterion [12,37], this multiple-round
To optimize the cruise efficiency, aerodynamic drag and lift are sampling strategy is simple and robust. In the first round, the Latin
required to be modeled. The pitching moment is also required to hypercube sampling (LHS) method is used to generate samples for
constrain the trim drag. The pressure distribution on the wing sur- the training of aerodynamic models; The incremental LHS method
face can provide valuable guidance to wing designers. We show
is used to generate samples for the validation of the models. In
how to construct data-based models of these aerodynamic func-
the next round of DoE, all data (both training and validation) in
tions in this section.
the previous round are used as the training data, and a new set of
wing samples are generated by the incremental LHS, which is used
3.1. Training data
as the validation data. The accuracy of data-based models gener-
ally increases with the rise of the number of training data, and
The training data is composed of CFD analyses on samples cho-
the process is repeated until the specified accuracy is satisfied. As
sen from the design space, which contains 60 parameters, the
shown in Table 2, five DoE rounds are performed in this work. Ex-
Mach number, the altitude, the angle of attack, seven twist angles,
cept for few samples of which CFD simulations do not converge,
and 50 wing modes. We use the coarse L3 mesh for the CRM wing
we eventually have 135,108 and 47,967 CFD samples for training
provided by Hwang et al. [33] to reduce the computational cost in
and validation, respectively.
this research. Previous work based on the mesh [33,3] has shown
to be effective to tap the potential of aerodynamic shape optimiza-
tion. To automatically generate CFD meshes for different wings,
1
we use the efficient analytic inverse-distance method implemented https://github.com/mdolab/idwarp.

3
J. Li and M. Zhang Aerospace Science and Technology 112 (2021) 106639

Fig. 3. C D reductions in shape optimization using a different number of models at nine typical flight missions. It shows that 50 wing modes contain almost enough geometric
merits for wing shape design.

Fig. 4. C P distributions of some wing samples indicate that the design space contains complex aerodynamic phenomena.

Table 2
Number of wing samples in each round. All of them are analyzed by CFD.

Round 1 Round 2 Round 3 Round 4 Round 5


Number of samples 40,000 26,845 27,975 41,340 48,260

The gradient of aerodynamic coefficients to each design vari- 3.2. Data-based models
able can be quickly solved by using the adjoint method. It has
been shown that the accuracy of DNN-based aerodynamic models Data-based aerodynamic models are cheap surrogate models
can be improved by using the gradient information [38]. However, that fit functions between the inputs and the aerodynamic coeffi-
many commonly-used CFD codes still do not have an adjoint suite. cients. Typical data-based methods include the kriging model [39]
For generality, we do not perform adjoint computations and merely and deep neural networks (DNN) [40,38,41,42]. Since the training
train the data-based models by CFD results. In the CFD analysis of of kriging leads to an inverse computation of the covariance ma-
each sample, the C L , C D , C M , and C P are recorded. C L , C D , and C M trix (the matrix order equals the number of training points), it is
are the metrics used in data-based optimization. C P is involved to not suitable for large datasets. Although Li et al. [27] managed to
provide a quick visual evaluation of each design. handle large datasets with tens of thousands of samples in airfoil

4
J. Li and M. Zhang Aerospace Science and Technology 112 (2021) 106639

Fig. 5. Investigation of DNN hyperparameters in modeling aerodynamic coefficients. (Solid and dashed lines represent the loss functions in the validation and training datasets,
respectively.)

design by using the mixture of kriging models, we find the fitted and one count for the others means 0.001. Note the x-axis in Fig. 6
functions are likely to be not smooth at the boundaries of different is a logarithmic scale. The probability density function (PDF) of the
experts. In airfoil design, it hardly brings difficulty to optimiza- errors (not absolute values) is Gaussian-distributed with the mean
tion because mixture models are defined by the two dominant approximately zero. The relative errors are computed by
modes and the searching is usually inside the same expert due
to geometric constraints. However, in wing shape design optimiza-
tion, it is much more common to see that the optimizer searches ε = |( y NN − y CFD )/ y CFD | × 100%. (3)
across different experts and gradient-based optimization fails be-
fore convergence. Thus, the approach of using a mixture of kriging It can be seen that C D and C M are more difficult to model than
models is not ideal in wing design problems. We chose DNN as the C L from the PDFs of the relative errors. For all aerodynamic co-
data-based model in this work due to its high capability of fitting efficients, the averaged absolute errors in the 47,967 validation
nonlinear functions and handling large datasets. Actually, even in samples are all less than one count.
1/ndim
the airfoil design problem, Bouhlel et al. [38] showed DNN could The sampling sparsity ratio (nsample ) of these models is approx-
be a better choice. imately 1.217. Compared with the sampling sparsity ratios in the
We use the mean absolute error (MAE) as the loss function in data-based airfoil analysis [27] (2.027 and 2.825 in the subsonic
the training of DNN models. The Adam gradient-based optimiza- and transonic models, respectively), there is a significant decrease.
tion algorithm for stochastic objective functions is used to train Nevertheless, it does not negatively affect the model accuracy in
DNN models due to its high efficiency [43]. The tanh activation this work. The mean relative errors for C L , C D , and C M are 0.20%,
function is used in all neurons to ensure a differentiable model. 0.35%, and 0.36%, respectively. The accuracy of these data-based
The hyperparameters of DNN (number of layers and number of wing models is close to that of data-based airfoil models in [27],
neurons in each layer) have a big influence on its accuracy. Gen- which means that it is promising to produce the same optimiza-
erally, more layers and neurons can provide more fitting freedom tion results as CFD-based optimization.
and may lead to higher accuracy. We investigate their influence in We investigate the pressure distribution by monitoring C P on
modeling the aerodynamic coefficients, and the results are shown six sectional airfoils of the wing, which can be expressed as a
in Fig. 5. The solid and dashed lines represent the loss functions vector. Modeling the C P values either independently or simulta-
in the validation and training datasets, respectively. For C L models, neously is difficult. Inspired by works using the proper orthogonal
the accuracy first increases with the rise of neuron and layer num- decomposition (POD) method [44–47] and kriging models in ap-
bers. However, when the layer number is increased to N layer = 6, proximating high-dimensional data, we use POD to reduce the di-
we find the DNN models tend to be over-fitting. The accuracy on mension and alternatively model the dominant POD coefficients to
the validation dataset does not increase although the model seems reconstruct C P . We refer to this approach as the DNN-POD model.
to be more accurate on the training dataset. C D and C M models We use conservative DDN models with N layer = 3 and N neuron = 70
encounter such issues when N layer = 5 and N layer = 4, respectively. to avoid over-fitting. The validation of the C P models in predicting
Based on the investigation, we chose DNN models of three, four, some test wings at three angles of attack (1◦ , 2◦ , and 3◦ ) is shown
and five layers for C L , C D , and C M . One hundred neurons are used in Fig. 7, where the C P curves of DNN-POD and CFD are solid and
in each layer. The error distributions of the chosen models in the dashed lines, respectively. 50 POD modes are used in this test and
analyses of validation samples are shown in Fig. 6. The absolute the data-based model manages to produce almost the same C P
errors are shown in counts, where one count for C D means 0.0001 distributions to CFD analyses.

5
J. Li and M. Zhang Aerospace Science and Technology 112 (2021) 106639

Fig. 6. Error distributions of DNN models in 47,967 validation wing samples.

Fig. 7. Data-based model produces similar C P distributions (dashed lines) to CFD analyses (solid lines) on arbitrary test wings at different angles of attack.

4. Wing shape design optimization demands, we develop a uniform data-based optimization frame-
work by referring to the multi-point optimization framework de-
Aerodynamic design of aircraft wings is a complex problem that veloped by the Multidisciplinary Design Optimization Laboratory.
may lead to single-point optimization, multi-point optimization, In this framework, we couple the data-based aerodynamic models
and multi-objective optimization problems. To address different with optimization and geometric constraint modules to perform

6
J. Li and M. Zhang Aerospace Science and Technology 112 (2021) 106639

merely one hundred iterations, which is much more efficient. Dif-


ferent gradient-based optimization tests converge to results with
the same drag coefficients, which implies that multimodality is not
a concern in this problem. Thus, we chose SLSQP to perform wing
design optimization in this work.
The XDSM (eXtended Design Structure Matrix) diagram [49] of
the data-based wing design optimization framework is shown in
Fig. 9. We use pyGeo [29] to evaluate geometric constraints, i.e.
volume and thickness constraints, and their derivatives. The deriva-
tives of aerodynamic functions in the DNN models are solved by
using the automatic differentiation function implemented in Ten-
sorFlow.
Although multi-objective optimization can be solved by using
evolutionary algorithms such as the non-dominated sorting genetic
algorithm II (NSGA-II) [50], it can be significantly inefficient in
high-dimensional problems. To address the issue, we alternatively
solve a series of multi-point design optimization with different
Fig. 8. Comparison of gradient-based and gradient-free optimization algorithms in a
wing drag minimization problem. combination weights [51,52] using SLSQP (Fig. 9). Different com-
bination weights have a different emphasis on the objective func-
tions, and thus the solutions of these multi-point optimizations
fast wing shape design optimization. One to N (N can be thou-
could represent the Pareto frontier line. We compare the perfor-
sands) aerodynamic functions can be simultaneously considered to
mance of this strategy and NSGA-II in a two-objective wing design
solve single-point, multi-point, and massively multi-point wing de-
optimization problem. The optimization is to maximize the lift to
sign optimization problems. The combination weights of different
drag ratio (L / D) at two Mach numbers: M = 0.83 and M = 0.87.
aerodynamic functions can be adjusted to solve the Pareto fron-
The altitude and angle of attack are 11,740 m and 2.0◦ , respec-
tier for multi-objective optimization. We explain the framework in
tively. For simplicity, no aerodynamic constraints are involved. The
Section 4.1 and compare the optimization results with CFD-based
details of the two-objective optimization problem are summarized
optimization in single-point (Section 4.2), multi-point (Section 4.3),
in Table 3.
and multi-objective wing design problems (Section 4.4).
The NSGA-II algorithm implemented in pyOptSparse [48] is
used in this comparison. We use a population size of 100 by de-
4.1. Optimization algorithm
fault, and the optimization results with three maximum generation
numbers (100, 1000, and 10000) are investigated. It can be seen
The computational cost of aerodynamic analyses using the DNN in Fig. 10 the performance of NSGA-II increases with the rise of
models is rather low. In the validation test on 47,967 wing sam- the generation number. Nevertheless, SLSQP (with 15 two-point
ples, the total computational time is 7.26 seconds. Apart from optimizations) achieves better results than them. Meanwhile, the
the model loading time, the aerodynamic analyses (C L , C D , and computational cost of NSGA-II increases linearly with the number
C M ) of each wing sample merely cost 1.1 × 10−4 second. Com- of generations, and the largest cost in this test is one million eval-
pared with the computational cost used in the construction of uations of the objective functions. SLSQP is much more efficient,
the data-based models, estimating aerodynamic functions using which merely costs about 2000 objective and derivatives analyses.
the models is rather cheap, and the on-line/off-line ratio is less This investigation shows that performing a series of gradient-based
than 1. × 10−11 in this work. Such cheap models could be coupled multi-point optimization can approximate the Pareto frontier line
with gradient-free algorithms, which may have a better capabil- in wing shape design, and this strategy is much more efficient than
ity in global searching than gradient-based ones. However, we find
using an evolutionary algorithm.
the performance of two typical gradient-free algorithms, the ge-
netic algorithm (GA) and the particle swarm optimization (PSO)
4.2. Single-point design optimization
method, are much ineffective than gradient-based optimization.
The comparison of different optimization algorithms in a wing
drag minimization problem is shown in Fig. 8. Drag minimiza- We first verify the data-based optimization framework in a drag
tion is performed for a mission with a flight altitude of 11,250 minimization problem shown in Table 4. The optimization problem
m at M = 0.85. The problem is subject to two aerodynamic con- is based on the ADODG Case 4.1, and the flight altitude is 11,740 m,
straints, a lift constraint with C L = 0.5 and a moment constraint M = 0.85, and C Lcon = 0.5. We use strict thickness constraints other
with C M > −0.17. No other constraints are specified for simplicity. than a loose volume constraint to ensure practical design [17].
All the optimization algorithms used in this comparison are imple- The computational cost in evaluating geometric constraints
mented in pyOptSparse [48].2 We use the sequential least squares could be significant given the low cost of data-based aerodynamic
programming (SLSQP) algorithm in gradient-based optimization, models. Nevertheless, the data-based optimization is done within
and multiple tests are performed to investigate whether gradient- one minute (33 seconds in this case) on a personal computer (PC)
based optimization may be stuck to local minima. The starting using one CPU processor with 1.7 GHz. We also perform CFD-based
points of different SLSQP tests are generated by the LHS method. optimization using the efficient adjoint solver in ADflow, which
Optimization using GA does not show a clear convergence trend costs 18.4 CPU hours using a 2.6 GHz processor. Data-based op-
even after one million calls of the aerodynamic analysis models. timization is about five thousand times faster than CFD-based op-
SPO is more efficient than GA in this test. Nevertheless, SPO still timization. We evaluate the optimized wings using CFD analyses,
requires half a million evaluations to find a similar solution to which are compared in Fig. 11. Both results are almost shock-free
that achieved by SLSQP. Optimizations using SLSQP converge after and have the same drag. The wing obtained by the data-based op-
timization has a slight shock wave near the wing-tip and produces
0.003 less lift. Nevertheless, the shapes of the two optimized wings
2
https://github.com/mdolab/pyoptsparse. are almost the same.

7
J. Li and M. Zhang Aerospace Science and Technology 112 (2021) 106639

Fig. 9. XDSM diagram of the data-based wing design optimization.

Table 3
Two-objective wing shape design problem for comparison of NSGA-II and SLSQP.

Functions Description Quantity


Maximize L/D Lift-to-drag ratio at M = 0.87 (Mission 1) 1
L/D Lift-to-drag ratio at M = 0.83 (Mission 2) 1
Total objective functions 2
With respect to xshape Wing shape modes 50
γ Wing twists 7
Total design variables 57

Table 5
Comparison of different single-point optimization results.

C CFD
D C DNN
D C LCFD C LDNN CFD
CM DNN
CM
Opt 1 214.7 215.0 0.483 0.480 -0.170 -0.170
Opt 2 243.3 244.1 0.537 0.534 -0.170 -0.170
Opt 3 278.9 279.9 0.591 0.587 -0.170 -0.170
Opt 4 203.8 204.9 0.450 0.449 -0.170 -0.170
Opt 5 229.0 229.0 0.500 0.497 -0.170 -0.170
Opt 6 261.0 260.9 0.550 0.547 -0.170 -0.171
Opt 7 199.8 200.9 0.420 0.418 -0.170 -0.170
Opt 8 222.8 224.0 0.466 0.462 -0.170 -0.171
Opt 9 254.5 256.8 0.513 0.507 -0.170 -0.171

Table 6
Nine-point optimization problem statement.

Functions Description Quantity


9
Minimize i =1 w i C iD Drag coefficient 1
Fig. 10. SLSQP is more efficient and effective than NSAGA-II in the two-objective
wing shape design optimization. With respect to αi Angle of attack 9
xshape Wing shape modes 50
γ Wing twists 7
Table 4 Total design variables 66
Single-point optimization problem statement. Subject to C Li = c i Lift constraint 9
C M ≥ −0.17 Moment constraint at C L = 0.5 1
Functions Description Quantity
t ≥ t initial Thickness constraints 750
Minimize CD Drag coefficient 1 Total constraints 760
With respect to α Angle of attack 1
xshape Wing shape modes 50
γ Wing twists 7
Total design variables 58 drag count. Compared to the computational cost of 10 ∼ 30 CPU
Subject to C L = C Lcon Lift constraint 1 hours with high-performance processors, the cost of data-based
C M ≥ −0.17 Moment constraint 1 optimization (∼ 1 minute on a modest PC) is a significant reduc-
t ≥ t initial Thickness constraints 750
tion. The data-based optimization approach can provide aircraft
Total constraints 752
designers comparable wing designs to CFD-based optimization al-
most in real-time.
We further test the data-based optimization in single-point de-
sign optimization for other flight missions, which are based on the 4.3. Multi-point design optimization
ADODG Case 4.5. The flight Mach numbers and C Lcon are shown in
Fig. 3. As shown in Table 5, we compare the results by data-based Single-point overstates the benefit at the design point at the
optimization with those solved by CFD-based optimization. The cost of off-design performance. Thus, multi-point optimization [53,
differences in optimized C D are almost within one count except 54] is required in practical aircraft wing design. As shown in Ta-
for the ninth optimization problem, which has a more significant ble 6, we perform a nine-point design optimization of the CRM
error because the mission is at the boundaries of the design space. wing to verify the proposed approach. The design points and com-
Nevertheless, the average difference in these tests is merely 0.9 bination weights are defined by referring to the ADODG Case 4.5.

8
J. Li and M. Zhang Aerospace Science and Technology 112 (2021) 106639

Fig. 11. Data-based optimization is comparable to CFD-based optimization in single-point design of the CRM wing.

Table 7 4.4. Multi-objective design optimization


Comparison of nine-point optimization results.

C CFD
D C NN
D C LCFD C LNN CFD
CM NN
CM To analyze the trade-off between different objective functions,
Design point 1 222.2 222.4 0.483 0.481 -0.161 -0.161 multi-objective optimization problems arise in aircraft design. The
Design point 2 249.5 249.8 0.537 0.536 -0.176 -0.176 Pareto frontier solved by multi-objective optimization is helpful for
Design point 3 283.8 285.6 0.591 0.592 -0.190 -0.192
the designer to comprise the objective functions and carry out
Design point 4 208.8 209.4 0.450 0.449 -0.155 -0.155
Design point 5 231.2 231.4 0.500 0.498 -0.170 -0.170 the best choice [55]. We investigate a two-objective optimization
Design point 6 260.1 260.6 0.550 0.548 -0.187 -0.187 problem to show the effectiveness of the proposed data-based ap-
Design point 7 206.4 206.8 0.420 0.419 -0.155 -0.155 proach. The objective functions are C D at two flight missions with
Design point 8 231.5 232.0 0.466 0.466 -0.173 -0.173
M = 0.87 and M = 83, respectively. The aircraft mass and flight al-
Design point 9 266.8 268.0 0.513 0.513 -0.191 -0.192
titude are 219,500 kg and 11,740 m. The required lift coefficients
at the specified flight altitude are 0.525 and 0.577. The details
The optimized wing is analyzed using CFD for verification. The of the two-objective optimization problem are summarized in Ta-
optimized shapes and C P counters at the fifth design point with ble 8.
M = 0.85 and C Lcon = 0.5 are shown in Fig. 12. CFD-based opti- To efficiently solve this two-objective design problem, we per-
mization results are provided as well for comparison. As a com- form a series of two-point design optimization with various com-
promise of the performances at other design points, the optimized bination weights. CFD-based optimization is also performed as a
wings are not shock-free. Thus the drag at this design point is comparison of data-based optimization. As shown in Fig. 13, data-
increased compared to that in the single-point optimization. Data- based optimization provides solutions close to the Pareto frontier
based approach effectively captures this trade-off and converges line solved by CFD-based optimization. The drags of the opti-
to a similar wing to that solved by CFD-based optimization. The mized wings are 10 ∼ 20 counts smaller than the CRM wing at
differences in C D are within one count at most design points. A the two missions. It approximately takes 600 CPU hours to solve
detailed comparison is shown in Table 7. The computational cost the Pareto frontier line by CFD-based optimization which is rather
of CFD-based optimization, in this case, is 270 CPU hours. Data- time-consuming. Data-based optimization costs several minutes
based approach significantly reduces the cost, which merely takes and provides solutions that are merely of one ∼ two counts dif-
2 minutes in this problem. ference.

9
J. Li and M. Zhang Aerospace Science and Technology 112 (2021) 106639

Fig. 12. Data-based optimization achieves similar results to CFD-based optimization in multi-point CRM wing design.

Table 8
Multi-objective optimization problem statement.

Functions Description Quantity


Minimize CD Drag at M = 0.87 (Mission 1) 1
CD Drag at M = 0.83 (Mission 2) 1
Total objective functions 2
With respect to α Angle of attack 2
xshape Wing shape modes 50
γ Wing twists 7
Total design variables 59
Subject to C L = 0.525 Lift constraint at Mission 1 1
C L = 0.577 Lift constraint at Mission 2 1
t ≥ t initial Thickness constraints 750
Total constraints 752
Fig. 13. Data-based multi-objective optimization effectively pushes the CRM wing to
the Pareto frontier.

5. Aerodynamic wing design in an innovative approach


In this section, we showcase the usage of the data-based model
With the cheap data-based optimization framework, wing de- in a dummy wing design problem with a total aircraft mass of
sign can be performed in an innovative way. For example, the ben- W = 220, 000 kg. The propulsion forces required to offset the aero-
efits of operating the aircraft at different altitudes and Mach num- dynamic drag on the wing at different Mach numbers and altitudes
bers can be investigated by performing a series of aerodynamic are compared. This may help to choose the suitable cruise flight
shape optimization. Such an investigation can provide designers condition in aircraft design.
guidance for aircraft design, but it is not computationally afford- We first compute the drag forces on the CRM wing to sup-
able when CFD-based optimization is used. In addition, we could port such an aircraft at different missions, which are shown in
also easily compromise the influences of different flight missions Fig. 14(a). To generate the counter, 21 × 21 = 441 missions with
by performing massively multi-point design optimization using the different flight altitudes and Mach numbers are investigated by
data-based model. the aerodynamic models. Since the air density and the speed of

10
J. Li and M. Zhang Aerospace Science and Technology 112 (2021) 106639

Fig. 14. The drag forces of the baseline wing and optimized wings at different cruise altitudes and Mach numbers to support an aircraft mass of W = 220, 000 kg.

sound differ with the change of the altitude, the lift coefficient point, multi-point, and multi-objective wing design problems. The
constraint for each mission is computed by Eq. (2). Choosing opti- optimized wing shapes are almost the same, which merely lead to
mal operating conditions based on the analyses of the CRM wing drag discrepancies within 1 ∼ 2 counts.
may not be preferable because the drag forces can be further re- The computational cost of aerodynamic analyses by the data-
duced by performing shape optimization. It is too time-consuming based models is 1.1 × 10−4 second per wing. With such compu-
to use CFD-based optimization due to the high computational cost. tationally cheap models, wing shape optimization problems can be
However, the investigation can be done within hours on a mod- solved almost in real-time, and thus wing design can be performed
est PC using data-based optimization. Using the data-based model, in an innovative approach. We showcase an example to investigate
the minimum drag forces considering wing shape modification are the propulsion force required at different missions and perform
shown in Fig. 14(b). The benefits of shape optimization at differ- massively multi-point wing design, which is not affordable using
ent missions are not constant, which leads to drag reductions of CFD-based optimization.
0.2 ∼ 6.3 KN. The reduction is higher at missions with high Mach One limitation of this work may be the coarse mesh used in
numbers and flight altitudes. We compare the C P of the CRM and CFD analyses. The high-fidelity aerodynamic functions solved by
the optimized wings, which are generated by using the data-based refined CFD meshes might be more complex. In addition, this work
C P models and merely cost 20 seconds. As shown in Fig. 15, we only introduces aerodynamic models, the influences of other dis-
find that more significant shock waves are produced on the CRM ciplines have not been considered. It is more meaningful to thor-
wing at these missions. This leaves more room for improvement oughly consider practical multidisciplinary constraints in wing de-
and thus shape design optimization reduces more drag at these sign. This work may play an instructive role in the construction of
missions. The optimization achieves the least drag reductions at data-based models for other disciplines.
missions with M = 0.85, which implies that the CRM wing is well
designed to be operated at this speed. Declaration of competing interest
Although wing shape design optimization could significantly re-
duce the required propulsion force, the reduction shown in Fig. 15 The authors declare that they have no known competing finan-
cannot be achieved by a particular wing without using morph- cial interests or personal relationships that could have appeared to
ing techniques. In order to obtain an optimized wing that reduces
influence the work reported in this paper.
the total drag most, we perform a massively multi-point optimiza-
tion [44] considering the 441 missions using the data-based model.
Acknowledgements
The optimized wing is compared with the CRM wing in Fig. 16,
where the C P is marked with a counter of drag reduction. The
average drag reduction is 0.8 KN, which is smaller than that in We acknowledge the Tier 2 grant from the Ministry of Ed-
Fig. 15. Nevertheless, the compromise reduces the drag force at ucation, Singapore (R-265-000-661-112). The computational re-
most missions. sources of the National Supercomputing Centre, Singapore (https://
www.nscc.sg) are acknowledged.
6. Conclusions
Appendix A
This work presents a generic data-based approach to fast and
high-fidelity wing shape design optimization. Global wing modes The deep-learning-based optimal sampling method combines
are used to parameterize wing shapes. A multi-round sampling an optimization process with deep-learning geometric models to
strategy is used to generate the training data. About 140,000 CFD find realistic wing shapes that satisfy the specified geometric con-
analyses of different wings at various operating conditions are per- straints. The XDSM diagram of this method is shown in Fig. 17,
formed. The optimal DNN hyperparameters for each aerodynamic and the details on the deep convolutional generative adversarial
model are selected by trial and error. The DNN models are cou- network (DCGAN) and convolutional neural network (CNN) of air-
pled in a gradient-based optimization framework to solve different foils can be found in [11]. First, we use the DCGAN model trained
wing design problems. by UIUC airfoils to generate sectional slices for the wings as the
The DNN models are verified by comparing with CFD analysis starting points of the optimization. Then, the optimization is per-
results of 47,967 testing wings, and the mean errors of C L , C D , formed by minimizing the displacement to ensure the sparsity of
and C M are all smaller than one count. The C P models have been wing samples. To obtain realistic wing samples, the optimization is
shown to accurately capture multiple shock waves on arbitrary subject to validity constraints. Validity constraints are computed
wing shapes. The data-based optimization results are compared by the CNN-based discriminator, which quickly detects geomet-
with those solved by CFD-based optimization in different single- ric abnormality without using any CFD analyses. Thus, these wing

11
J. Li and M. Zhang Aerospace Science and Technology 112 (2021) 106639

Fig. 15. C P distributions on the baseline wing (dashed lines) and optimized wings (solid lines) at different cruise altitudes and Mach numbers. The colors show the reduced
values of the wing drag via shape design optimization.

Fig. 16. C P distributions on the CRM wing (dashed lines) and the wing designed by massively multi-point optimization (solid lines) at different cruise altitudes and Mach
numbers. The colors show the reduced values of the wing drag via the optimization.

samples could define the feasible domain of the high-dimensional sectional airfoils, so wing samples merely differ in z coordinates.
design space. Thus, these samples can be represented by the z coordinates of the
We use eight sectional airfoils to represent each wing sample. n slice points. Assuming we have m samples, which are assembled
The sampling process only perturbs the z coordinates of the eight in a snapshot matrix as

12
J. Li and M. Zhang Aerospace Science and Technology 112 (2021) 106639

Fig. 17. Deep-learning-based optimal wing sampling method proposed by Li et al. [3].
⎡ ⎤
z11 − z1CRM z12 − z1CRM ... 1 − z1
zm CRM
[15] M.B. Giles, N.A. Pierce, An introduction to the adjoint approach to de-
⎢ 1 ⎥
⎢ z2 − z2CRM z22 − z2CRM . . . CRM ⎥
2 − z2
zm sign, Flow Turbul. Combust. 65 (2000) 393–415, https://doi.org/10.1023/A:
⎢ ⎥ 1011430410075.
A=⎢ .. .. .. ⎥, (4)
⎢ .. ⎥ [16] G.K.W. Kenway, C.A. Mader, P. He, J.R.R.A. Martins, Effective adjoint approaches
⎣ . . . . ⎦ for computational fluid dynamics, Prog. Aerosp. Sci. 110 (2019) 100542, https://
doi.org/10.1016/j.paerosci.2019.05.002.
zn1 − znCRM zn2 − znCRM ... znm − znCRM [17] J. Li, S. He, J.R.R.A. Martins, Data-driven constraint approach to ensure low-
speed performance in transonic aerodynamic shape optimization, Aerosp. Sci.
where zCRM
i
is ith z coordinate of the CRM baseline. Performing Technol. 92 (2019) 536–550, https://doi.org/10.1016/j.ast.2019.06.008.
SVD on A, we obtain [18] S. Chen, Z. Lyu, G.K.W. Kenway, J.R.R.A. Martins, Aerodynamic shape opti-
mization of the common research model wing-body-tail configuration, J. Aircr.
53 (1) (2016) 276–293, https://doi.org/10.2514/1.C033328.
A = U V T, (5) [19] R. Lei, J. Bai, D. Xu, Aerodynamic optimization of civil aircraft with wing-
mounted engine jet based on adjoint method, Aerosp. Sci. Technol. 93 (2019)
where columns in U represent mode shapes of the wing. With 105285, https://doi.org/10.1016/j.ast.2019.07.018.
the diagonal entries of  assembled in descending order, and the [20] Z. Lyu, J.R.R.A. Martins, Aerodynamic design optimization studies of a blended-
modes in U are sorted in terms of their importance. wing-body aircraft, J. Aircr. 51 (5) (2014) 1604–1617, https://doi.org/10.2514/1.
C032491.
[21] N.R. Secco, J.R.R.A. Martins, RANS-based aerodynamic shape optimization of a
References strut-braced wing with overset meshes, J. Aircr. 56 (1) (2019) 217–227, https://
doi.org/10.2514/1.C034934.
[1] A. Jameson, Aerodynamic design via control theory, J. Sci. Comput. 3 (3) (1988) [22] Z.-H. Han, J. Chen, K.-S. Zhang, Z.-M. Xu, Z. Zhu, W.-P. Song, Aerodynamic
233–260, https://doi.org/10.1007/BF01061285. shape optimization of natural-laminar-flow wing using surrogate-based ap-
[2] Z. Lyu, G.K.W. Kenway, J.R.R.A. Martins, Aerodynamic shape optimization in- proach, AIAA J. 56 (7) (2018) 2579–2593, https://doi.org/10.2514/1.J056661.
vestigations of the common research model wing benchmark, AIAA J. 53 (4) [23] R. Yondo, E. Andrés, E. Valero, A review on design of experiments and surro-
(2015) 968–985, https://doi.org/10.2514/1.J053318. gate models in aircraft real-time and many-query aerodynamic analyses, Prog.
Aerosp. Sci. 96 (2018) 23–61, https://doi.org/10.1016/j.paerosci.2017.11.003.
[3] J. Li, M. Zhang, Adjoint-free aerodynamic shape optimization of the common
[24] J. Li, J. Cai, K. Qu, Drag reduction of transonic wings with surrogate-based opti-
research model wing, AIAA J. (2021) 1–11, https://doi.org/10.2514/1.j059921.
mization, in: Lecture Notes in Electrical Engineering, Springer, Singapore, 2019,
[4] Y. Shi, C.A. Mader, S. He, G.L.O. Halila, J.R.R.A. Martins, Natural laminar flow
pp. 1065–1080.
airfoil design using a discrete adjoint approach, AIAA J. 58 (11) (2020), https://
[25] D.R. Jones, M. Schonlau, W.J. Welch, Efficient global optimization of expen-
doi.org/10.2514/1.J058944.
sive black-box functions, J. Glob. Optim. 13 (1998) 455–492, https://doi.org/
[5] J. Xu, L. Qiao, J. Bai, Improved local amplification factor transport equation for
10.1023/A:1008306431147.
stationary crossflow instability in subsonic and transonic flows, Chin. J. Aero-
[26] Z. Han, C. Xu, L. Zhang, Y. Zhang, K. Zhang, W. Song, Efficient aerodynamic
naut. 33 (12) (2020) 3073–3081, https://doi.org/10.1016/j.cja.2020.05.012.
shape optimization using variable-fidelity surrogate models and multilevel
[6] Y. Yu, Z. Lyu, Z. Xu, J.R.R.A. Martins, On the influence of optimization algo-
computational grids, Chin. J. Aeronaut. 33 (1) (2020) 31–47, https://doi.org/10.
rithm and starting design on wing aerodynamic shape optimization, Aerosp.
1016/j.cja.2019.05.001.
Sci. Technol. 75 (2018) 183–199, https://doi.org/10.1016/j.ast.2018.01.016.
[27] J. Li, M.A. Bouhlel, J.R.R.A. Martins, Data-based approach for fast airfoil analy-
[7] S. Skinner, H. Zare-Behtash, State-of-the-art in aerodynamic shape optimisation
sis and optimization, AIAA J. 57 (2) (2019) 581–596, https://doi.org/10.2514/1.
methods, Appl. Soft Comput. 62 (2018) 933–962, https://doi.org/10.1016/j.asoc.
J057129.
2017.09.030. [28] J. Li, M. Zhang, On deep-learning-based geometric filtering in aerodynamic
[8] J. Li, J. Cai, K. Qu, Adjoint-based two-step optimization method using proper shape optimization, Aerosp. Sci. Technol. 112 (2021) 106603, https://doi.org/
orthogonal decomposition and domain decomposition, AIAA J. 56 (3) (2018) 10.1016/j.ast.2021.106603.
1133–1145, https://doi.org/10.2514/1.j055773. [29] G.K. Kenway, G.J. Kennedy, J.R.R.A. Martins, A CAD-free approach to high-
[9] J.R.R.A. Martins, Perspectives on aerodynamic design optimization, in: AIAA fidelity aerostructural optimization, in: Proceedings of the 13th AIAA/ISSMO
SciTech Forum, AIAA, Orlando, FL, 2020. Multidisciplinary Analysis Optimization Conference, Fort Worth, TX, 2010.
[10] X. He, J. Li, C.A. Mader, A. Yildirim, J.R.R.A. Martins, Robust aerodynamic shape [30] D.A. Masters, N.J. Taylor, T.C.S. Rendall, C.B. Allen, D.J. Poole, Geometric com-
optimization—from a circle to an airfoil, Aerosp. Sci. Technol. 87 (2019) 48–61, parison of aerofoil shape parameterization methods, AIAA J. 55 (5) (2017)
https://doi.org/10.1016/j.ast.2019.01.051. 1575–1589, https://doi.org/10.2514/1.j054943.
[11] J. Li, M. Zhang, J.R.R.A. Martins, C. Shu, Efficient aerodynamic shape opti- [31] D.J. Poole, C.B. Allen, T.C.S. Rendall, Metric-based mathematical derivation of
mization with deep-learning-based geometric filtering, AIAA J. 58 (10) (2020) efficient airfoil design variables, AIAA J. 53 (5) (2015) 1349–1361, https://doi.
4243–4259, https://doi.org/10.2514/1.j059254. org/10.2514/1.j053427.
[12] N. Bartoli, T. Lefebvre, S. Dubreuil, R. Olivanti, R. Priem, N. Bons, J.R.R.A. Mar- [32] C.B. Allen, D.J. Poole, T.C.S. Rendall, Wing aerodynamic optimization using ef-
tins, J. Morlier, Adaptive modeling strategy for constrained global optimization ficient mathematically-extracted modal design variables, Optim. Eng. 19 (2)
with application to aerodynamic wing design, Aerosp. Sci. Technol. 90 (2019) (2018) 453–477, https://doi.org/10.1007/s11081-018-9376-7.
85–102, https://doi.org/10.1016/j.ast.2019.03.041. [33] J.T. Hwang, J. Jasa, J.R.R.A. Martins, High-fidelity design-allocation optimiza-
[13] J. Li, J. Cai, K. Qu, Surrogate-based aerodynamic shape optimization with the tion of a commercial aircraft maximizing airline profit, J. Aircr. 56 (3) (2019)
active subspace method, Struct. Multidiscip. Optim. 59 (2) (2019) 403–419, 1165–1178, https://doi.org/10.2514/1.C035082.
https://doi.org/10.1007/s00158-018-2073-5. [34] C.A. Mader, G.K.W. Kenway, A. Yildirim, J.R.R.A. Martins, ADflow: an open-
[14] A. Benaouali, S. Kachel, Multidisciplinary design optimization of aircraft source computational fluid dynamics solver for aerodynamic and multidisci-
wing using commercial software integration, Aerosp. Sci. Technol. 92 (2019) plinary optimization, J. Aerosp. Inform. Syst. 17 (9) (2020) 508–527, https://
766–776, https://doi.org/10.1016/j.ast.2019.06.040. doi.org/10.2514/1.i010796.

13
J. Li and M. Zhang Aerospace Science and Technology 112 (2021) 106639

[35] A. Yildirim, G.K.W. Kenway, C.A. Mader, J.R.R.A. Martins, A Jacobian-free approx- [46] T. Braconnier, M. Ferrier, J.-C. Jouhaud, M. Montagnac, P. Sagaut, Towards an
imate Newton–Krylov startup strategy for RANS simulations, J. Comput. Phys. adaptive POD/SVD surrogate model for aeronautic design, Comput. Fluids 40 (1)
397 (2019) 108741, https://doi.org/10.1016/j.jcp.2019.06.018. (2011) 195–209, https://doi.org/10.1016/j.compfluid.2010.09.002.
[36] P. Spalart, S. Allmaras, A one-equation turbulence model for aerodynamic flows, [47] A. Chiplunkar, E. Bosco, J. Morlier, Gaussian process for aerodynamic pres-
in: 30th Aerospace Sciences Meeting and Exhibit, 1992. sures prediction in fast fluid structure interaction simulations, in: Advances in
[37] J. Liu, W.-P. Song, Z.-H. Han, Y. Zhang, Efficient aerodynamic shape optimiza- Structural and Multidisciplinary Optimization, Springer International Publish-
tion of transonic wings using a parallel infilling strategy and surrogate mod- ing, 2017, pp. 221–233.
els, Struct. Multidiscip. Optim. 55 (3) (2016) 925–943, https://doi.org/10.1007/ [48] R.E. Perez, P.W. Jansen, J.R.R.A. Martins, pyOpt: a python-based object-oriented
s00158-016-1546-7. framework for nonlinear constrained optimization, Struct. Multidiscip. Optim.
[38] M.A. Bouhlel, S. He, J.R.R.A. Martins, Scalable gradient-enhanced artificial neu- 45 (1) (2012) 101–118, https://doi.org/10.1007/s00158-011-0666-3.
ral networks for airfoil shape design in the subsonic and transonic regimes, [49] A.B. Lambe, J.R.R.A. Martins, Extensions to the design structure matrix for the
Struct. Multidiscip. Optim. 61 (4) (2020) 1363–1376. description of multidisciplinary design, analysis, and optimization processes,
[39] D.G. Krige, A statistical approach to some basic mine valuation problems on Struct. Multidiscip. Optim. 46 (2012) 273–284, https://doi.org/10.1007/s00158-
the Witwatersrand, J. Chem. Metall. Min. Soc. 52 (1951) 119–139. 012-0763-y.
[40] G. Sun, S. Wang, A review of the artificial neural network surrogate modeling [50] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multiobjective
in aerodynamic design, Proc. Inst. Mech. Eng., Part G, J. Aerosp. Eng. 233 (16) genetic algorithm: NSGA-II, IEEE Trans. Evol. Comput. 6 (2) (2002) 182–197,
(2019) 5863–5872, https://doi.org/10.1177/0954410019864485. https://doi.org/10.1109/4235.996017.
[41] S.A. Renganathan, R. Maulik, J. Ahuja, Enhanced data efficiency using deep [51] S. Gass, T. Saaty, The computational algorithm for the parametric objective
neural networks and Gaussian processes for aerodynamic design optimization, function, Nav. Res. Logist. Q. 2 (1–2) (1955) 39–45, https://doi.org/10.1002/nav.
arXiv preprint arXiv:2008.06731, 2020. 3800020106.
[42] Y. Zhang, W.J. Sung, D.N. Mavris, Application of convolutional neural network [52] L. Zadeh, Optimality and non-scalar-valued performance criteria, IEEE Trans.
to predict airfoil lift coefficient, in: 2018 AIAA/ASCE/AHS/ASC Structures, Struc- Autom. Control 8 (1) (1963) 59–60, https://doi.org/10.1109/tac.1963.1105511.
tural Dynamics, and Materials Conference, American Institute of Aeronautics [53] G.K.W. Kenway, J.R.R.A. Martins, Multipoint aerodynamic shape optimization
and Astronautics, 2018. investigations of the common research model wing, AIAA J. 54 (1) (2016)
[43] D.P. Kingma, J. Ba, Adam: a method for stochastic optimization, arXiv preprint 113–128, https://doi.org/10.2514/1.J054154.
arXiv:1412.6980, 2014. [54] G. Chen, K.J. Fidkowski, Variable-fidelity multipoint aerodynamic shape opti-
[44] J. Li, J. Cai, Massively multipoint aerodynamic shape design via surrogate- mization with output-based adapted meshes, Aerosp. Sci. Technol. 105 (2020)
assisted gradient-based optimization, AIAA J. (2019) 1–15, https://doi.org/10. 106004, https://doi.org/10.1016/j.ast.2020.106004.
2514/1.j058491. [55] G. Chiandussi, M. Codegone, S. Ferrero, F. Varesio, Comparison of multi-
[45] M. Xiao, P. Breitkopf, R.F. Coelho, C. Knopf-Lenoir, M. Sidorkiewicz, P. Villon, objective optimization methodologies for engineering applications, Comput.
Model reduction by CPOD and kriging, Struct. Multidiscip. Optim. 41 (4) (2009) Math. Appl. 63 (5) (2012) 912–942, https://doi.org/10.1016/j.camwa.2011.11.
555–574, https://doi.org/10.1007/s00158-009-0434-9. 057.

14

You might also like