Original Article

Proc IMechE Part G:

J Aerospace Engineering
A review of the artificial neural 0(0) 1–10
! IMechE 2019

network surrogate modeling Article reuse guidelines:

sagepub.com/journals-permissions

in aerodynamic design DOI: 10.1177/0954410019864485

journals.sagepub.com/home/pig

Gang Sun1 and Shuyue Wang2

Abstract
Artificial neural network surrogate modeling with its economic computational consumption and accurate generalization
capabilities offers a feasible approach to aerodynamic design in the field of rapid investigation of design space and optimal
solution searching. This paper reviews the basic principle of artificial neural network surrogate modeling in terms of data
treatment and configuration setup. A discussion of artificial neural network surrogate modeling is held on different
objectives in aerodynamic design applications, various patterns of realization via cutting-edge data technique in numerous
optimizations, selection of network topology and types, and other measures for improving modeling. Then, new frontiers
of modern artificial neural network surrogate modeling are reviewed with regard to exploiting the hidden information
for bringing new perspectives to optimization by exploring new data form and patterns, e.g. quick provision of candidates
of better aerodynamic performance via accumulated database instead of random seeding, and envisions of more physical
understanding being injected to the data manipulation.

Keywords
Artificial neural network, surrogate modeling, aerodynamic design, machine learning, optimization

Date received: 2 April 2019; accepted: 21 June 2019

One major obstacle is the right optimization pattern/

Introduction approach for the project. The traditional solution to
The sector of commercial aeronautics is meeting the aerodynamic design problems tends to be a top-down
challenges of public expectation of cheaper fares and approach that relies on physics modeling. For exam-
reduced environmental impact upon community noise ple, the inverse design of pressure distribution desig-
around airports and global warming.1 The European nating, as one of the many methods proposed in wing
Union Vision For 2025 requires progress in low- drag reduction problems, modifies the geometry via
emission energy to secure a sustainable development. iterations until the designated pressure distribution
The Advisory Council for Aviation Research over surface is obtained.5 This method is later chal-
(ACARE) has set a 2050 goal for civil aeronautical lenged by new alternatives in terms of computational
industry to allow a 75% reduction in CO2 emissions resources consumption, e.g. heuristic algorithms com-
per passenger kilometer, a 90% reduction in NOx bined with the parametric geometry description
emission, and a 65% reduction in the perceived method.6 Also, nonlinearity features the aerodynamic
noise emission of flying aircraft.2 Therefore, many optimization problems not only in the physics behind
institutions are dedicated to research and technical flow phenomena, but also in the influence of geometry
exploration in aerodynamic design and optimization.3 description approach on corresponding aerodynamic
For example, flight tests of the BLADE laminar wing performance. Furthermore, many aerodynamic design
boast a 50% wing friction reduction and up to 5% problems require time-consuming evaluations, com-
less CO2 emissions; Clean Sky and SESAR projects pared with the duration of operation of the
introduce a new generation of aircraft reducing emis-
sions by 15–20%.4
1
Aerodynamic design is related to airfoil, aircraft Department of Aeronautics & Astronautics, Fudan University, China
2
AECC Commercial Aircraft Engine Co., Ltd, China
wing, turbine engine blades, unmanned aerial vehicle
(UAV), etc. Generally, typical aerodynamic designs Corresponding author:
and optimizations formulate objectives, optimization Shuyue Wang, AECC Commercial Aircraft Engine Co., Ltd, China.
algorithms, constraint functions, and design variables. Email: [email protected]
2 Proc IMechE Part G: J Aerospace Engineering 0(0)

optimization algorithm. The computational resource experiments to aid the optimization. Meanwhile,
consumption grows rapidly with the utilization of reduced-ordered modeling is a similar method that
high-fidelity tools for simulations.7 In addition, most focuses on abstract regeneration of complex flowfield
aerodynamic design problems require an appropriate by modeling.17 For example, the curse of the dimen-
amount of sample points generated via distributing sionality can be relieved with reduced-order modeling
approaches of design of experiment (DoE) methods, in unsteady aerodynamics at varying flow condi-
e.g. D-optimality, Monte Carlo, and Latin hypercube tions.18 Comparatively, surrogate modeling is more
in multivariable design space, before the set of experi- applicable than reduced-order modeling, because the
mental points is sent to simulator for responses. former is more related to the input and output data of
For example, large numbers of analyses have to be optimization problem and is less involved in detailed
carried out in constructing Hessian at the design aerodynamics.19 Therefore, this paper focuses on
point by sampling the design space when there are surrogate model thereafter.
many variables, particularly when using finite differ- Artificial neural network (ANN) is proposed as a
ence methods to evaluate gradients.8 In fact, many data-driven method to transform engineering analysis
optimizations depend on some forms of internal and design.20 It emulates biological information pro-
model for design space exploration.9 For example, cessing detouring the need of any objective functions;
simulations of flow field around new airfoils for dif- thus it offers a feasible solution to the technological
ferent set of shape variables10 or newly morphed needs in aerodynamic optimization, which will be dis-
geometry11 have to be conducted to populate the cussed in the following section. It can be considered as
design space in order to find the optimal candidate. an interpolation tool for obtaining data that are not
That is to say, obtaining enough information to pre- originally present in the training data. For example,
dict a design landscape in a hypercube of increasing data-fit models are generated using regression of high-
dimensions is a barrier in many optimizations.12 fidelity simulation data from the input to the output.7
Approaches including abstraction of problem dimen- ANN has been widely applied in modern surrogate
sions and simplification of geometry topology have modeling with its advantage of consuming trivial com-
been adopted in many studies to relieve this difficulty, putational effort; thus, it can be used as the assistant for
which yet leads to the inevitable deviation from real computational fluid dynamics (CFD) calculation for a
design.13 A strategy of optimization is demanded to large number of designed geometry in a short time,
be efficient, intelligent, and credible for balancing the which greatly increases the optimization efficiency.21
modeling efforts and design space scope.14 Its accurate generalization and parallel computation
The rapid advance in data science has brought new capabilities in complex engineering design problems
insights into aerodynamic design and optimization by are helpful in the rapid investigation of design space
constructing surrogate models (also termed as meta- and searching for optimalities.12 For example, ANN
model or response surfaces). The ‘‘aerodynamic has been used to expedite decision-making process in
design’’ refers to the design of airfoil, wing, engine early stages of aircraft design process and to select
nacelles, etc. that are components of an aircraft. proper combination of engine thrust, wing area, and
A surrogate model is proposed as a data-driven and the aircraft weight without going through elaborate
bottom-up approach used when an outcome of inter- details of other direct approaches.22 The applications
est cannot be easily obtained or the inner mechanism and prospects in some new frontiers of ANN surrogate
of simulation is not assumed to be known. It can be modeling will be discussed in a later section in detail.
viewed as the response of simulator to the data points
in design space comprehensively harnessing high-
fidelity simulations and experiments to aid the
Artificial neural network
optimization, sometimes dominating the whole opti- The advanced intellectual capability and processing
mization process, and sometimes functioning just as a power of human brains come from the biological
supplementary aid.15 The rise of surrogate model neural network that are composed of numerous chem-
comes with the application of inverse design. ically connected neurons.12 A neuron is connected to
Traditionally, airfoil design is important for aircraft one another with axons and dendrites; the connecting
wings, helicopter rotor blades, etc.16 Given the bound- regions between them are synapses.23 A neuron
ary conditions of coming flow, the airfoil shape deter- receives input from many sources, and then generates
mines its aerodynamic performance including the a unique output that can be passed on to other
pressure distribution over the surface. Navier–Stokes neurons in turn.24 The complexity of biological
equations are applied for obtaining the pressure dis- neural network determines the level of intelligence.24
tribution of specified airfoil shape. Then, the pressure The architecture and strengths of synaptic connec-
distribution is specified and the airfoil shape is tions often adapt to external stimuli, which is how
obtained as an output of complex aerodynamic learning takes place in organisms.25 Similarly, ANN
shape optimization procedure.11 In this circumstance, propagates the computed values from the input neu-
a surrogate model can be introduced to comprehen- rons to the output neurons, where learning takes place
sively harness data from high-fidelity simulations and by changing the weights representing the connections
Sun and Wang 3

between neurons.26 In many applications, ANN is reduce the error of prediction EðXÞ  ¼ o  o^ between
needed to adjust their internal structure to produce the observed value in the training sample and the pre-
correct outputs for sample inputs, thus approximating dicted outcome. Recurrent connections can also be uti-
the implicit relationship.27 lized where the predicted outputs become part of the
Various input–output functions and learning meth- next input vector. When the error value EðXÞ  is non-
ods can be implemented in realizing neural networks. zero, the weights in the neural network need to updat-
The configuration of neural network is usually one of ing in the negative direction of error gradient.
the biggest problems in optimization. Figure 1 shows The configuration of the value set to each neuron
the structure of perceptron as fundamental component and weights is the result of neural network training.
of ANN architecture, which contains one input node, The operation of a perceptron lays foundation of
one hidden node, and an output node. Generally, the ANN; its interpretation as a computational unit is
paradigm of ANN is composed of a multi-layer per- useful because it allows to put together multiple
ceptron (MLP) based on feed-forward, supervised units in order to create powerful models in neural
learning, and an error back-propagation training algo- network training.25 Thereby, the basic ANN architec-
rithm. Consider a situation where each training ture comprises an input layer, hidden layer(s), and an
instance is of the form ðX,  oÞ: each X ¼ ½x1 , . . . , xd  output layer. Information processing, proceeding
contains d feature variables, and o 2 f1, þ 1g con- from left to right within each layer of the ANN,
tains the observed value (which is given to the designer occurs at many simple processing units or elements.
as a part of the training data). The input layer contains ANN topology is established/modified according to
d nodes that transmit d features X ¼ ½x1 , . . . , xd  with the detailed situation of the optimization problem: In
edges of weight W  ¼ ½w1 , . . . wd  to an output node. single-layer neural networks (Figure 1), the training pro-
The input layer does not perform any computation cess is straightforward because the error can be com-
in  
Pd its own right, and the linear function W  X ¼ puted as a direct function of the weights, which allows
w x
i¼1 i i is computed at the output nodes. easy gradient computation. In multilayer networks
Subsequently, an activation function sign of this (Figure 2), the error is a complicated composition func-
value in form of real number predicts the dependent tion of the weights in earlier layers. The gradient of a
variable of X. The activation function as the source of composition function is computed using back-propaga-
nonlinearities determines the input–output relationships tion algorithm, which leverages the chain rule of differ-
of each processing unit as well as the form of the final ential calculus, computing the error gradients with
solution. Therefore, the P prediction
d  o^ is computed as: respect of weights in terms of summations of local-gra-
o^ ¼ signðW  XÞ
 ¼ sign w x where the sig- dient products over the various paths from nodes to
i¼1 i i
moidal function sign maps a real value to either –1 outputs. The principle of weight adjusting is as follows
or þ 1, which is a common approach in the ANN con- " #
figuration. Appropriate activation function can provide @L @L X @o Y k1
@hiþ1
the desired nonlinear relationship between the input ¼ 
@wðhr1 , hr Þ @o ½hr , hrþ1 ,..., hk , o2P
@h k i¼r @hi
and output vectors. The circumflex on top of the vari-
able o indicates a predicted value instead of an observed @hr
 , 8r 2 1 . . . k
one. The primary objective for ANN training is to @wðhr1 , hr Þ

Figure 1. Basic architecture of a perceptron.

4 Proc IMechE Part G: J Aerospace Engineering 0(0)

immediate feedback within design iterations. With sur-

rogate modeling, CFD can be saved during optimiza-
tion except for final design validation. ANN is used in
many applications of surrogate models due to its huge
convenience available for problems with large amount
of data. The principle behind surrogate modeling is
that data at input and output is related through the
pattern of the trained neural network.

Direct applications of ANN surrogate modeling in

aerodynamic design
Many ANN surrogate modeling have been applied in
Figure 2. Basic architecture of a multilayer perceptron.
optimization. ANN has been implemented efficiently to
interpolate the aerodynamic pressure loads for one-
where L is the error function; h is the hidden layer, way UAV fluid structure interaction.30 The result
with subscript indicating layer sequence; wðhr1 , hr Þ is shows good agreement with the actual pressure profile
the weight value connecting layers hr and hr1 .25 on aircraft compared against two-dimensional curve
The initialization of weight and biases for ANN is fitting with higher order polynomials. With data train-
investigated to select the effective starting points for ing, ANN is able to learn active control strategy from
training network efficiently and accelerating the conver- experiments of mass flow rates of two jets on sides of a
gence.12 The optimal number of nodes in the hidden cylinder.31 Its predictive capability is shown insensitive
layer and the optimal number of hidden layers is pro- to numerical instabilities and convergence difficulties
blem-dependent. Networks with different numbers of typically associated with computational processes.24
neurons were evaluated to minimize the regression Turbomachinery uncertainty analysis requires per-
error; however, these numbers should be kept low for forming a large number of simulations, the computa-
the computational efficiency. The number of nodes and tional cost of which can be greatly alleviated with
layers should be increased if convergence difficulties are ANN surrogate modeling.32 ANN has successfully
encountered, but should not exceed the total number of helped to estimate the separation point and stall
input and output variables. A simpler network with no speed of cascaded fins from the relationship with the
hidden layers may be computationally efficient, but it number of fins in the cascade.33 Surrogate models with
represents only linear mapping between input and ANN have been shown a good alternative to conven-
output quantities, known as flat networks and can be tional solution with regard to the prediction of aero-
inadequate to model nonlinear relationships. dynamic coefficients of airplanes of high accuracy.28
ANN modeling of learning from the accumulation ANN has been used for space mapping for transonic
of expertise have found their way into practical appli- airfoil aerodynamic shape optimization.34 ANN has
cations in many areas.25 The developed techniques been trained by the data of fuselage drag coefficient
assist in addressing a wide range of complex problems obtained by accumulated experimental results con-
in aerodynamic design, where an ANN is fed with CFD ducted in wind tunnel to be capable of fuselage drag
simulation during training. For example, an aero- coefficient estimation for each parameter values of
dynamic database consisting of approximately 100,000 fuselage shape with respect to inputs without rigorous
cases calculated with a full-potential code with compu- computations.20 It has also helped to achieve optimal
tation of viscous effects was used for the neural network profiles for minimizing time-averaged drag and buffet
training, with the aid of backpropagation algorithm, magnitude in the supercritical airfoil design.10
scaled gradient algorithm, and Nguyen–Wridow The numerical search for the optimum shape is of
weight initialization.28 Among the techniques, a surro- great interest for aircraft and turbomachinery
gate model established/aided by neural network method designers. The authors of this paper have both been
attracts many scientists due to its potential to automat- dedicated to an applicable airfoil/wing inverse design
ically give the reference geometry according to the method with the help of ANN and database (Figure 3)
design target.29 The following section will discuss in in a design for a transonic swept wing of a passenger
detail the implementation of ANN in surrogate model- jet.16 It can directly generate profiles fitting the
ing in the field of aerodynamic design. requested aerodynamic performance with trained
neural network, avoiding the repetitive cut-and-try.35
Surrogate model in aerodynamic design
with ANN Variation, improvement, and development of ANN
A surrogate model is aimed at reducing computational
surrogate modeling
resource consumption in aerodynamic design and opti- Selection of type of neural network. The type of neural
mization. Efforts have been put so that designers get network is an important option that lays influence
Sun and Wang 5

method represents an accurate UAV aerofoil with 10

geometry design variables.12 Parametric section
(PARSEC) is compared with other kinds of paramet-
rization method and is evaluated as appropriate for
airfoil description due to its accuracy and
intuitiveness.16
ANN designers are inspired by the idea that the
outcome of data training may largely be dependent
on the form the input data takes. On one hand, the
form of input data should be excluded with nonmea-
ningful information so that the training can be guided
with some kind of direction. On the other hand, the
form itself may hinder the ANN to automatically
relate the crucial physical meaning that effectively
decides the output data (e.g. aerodynamic perform-
ance) during training. For example, the geometric rep-
resentations are sometimes not effective for neural
network training since the hidden semantic meaning
of the vectors of input data varies.37 Data tend to lose
its apparent physical meaning in the process of par-
ametrization and normalization. Obviously, the data
form has influence on the information where the data
are transmitted to the neural network during training.
Figure 3. An airfoil database for inverse design.
Some knowledge of aerodynamic optimization that is
not expressed in data will never be ‘‘understood’’ by
on the effect of the ANN surrogate modeling. There the neural network.
was an intensive comparative study on the approxi- The aerodynamic performance of an airfoil is func-
mation performance of three prospective surrogate tion of its geometrical shape, which implies in many
models: ANN, radial basis function (RBF), and sup- situations (especially when the target aerodynamic
port vector regression.32 The result shows that ANN performance is not multi-objective) where there may
outperforms others, but it may alter if the problem is be more than one set of geometrical shape that fits the
changed because ANN needs a certain data set to be aerodynamic feature given by designers. Under this
effectively trained. A data set is an assembly of data. circumstance, classification via self-organizing map-
Different surrogate modeling approaches, e.g. RBF, ping (SOM) can relieve this challenge, applications
Kriging, and support vector regression, etc., have of which can be seen in Figures 4 and 7, where data-
been compared to bring the efficient global optimiza- base of airfoils are classified into several groups
tion closer to reality.8 It is certain that the best selec- according to the similarity of feature data.
tion of ANN architecture always accomplishes with Therefore, new forms of geometric representation
different data sets of problem.20 The outcomes can be method are proposed specifically to the network train-
influenced by many factors, and therefore there is no ing. Geometry data instead of abstract parameters can
simple deduction to which type of surrogate model be directly input into deep learning network in the
outperforms others. The main approaches in the fron- form of coordinate of wing profile points, or even in
tiers of ANN surrogate modeling proposed to the form of tensors that records the connection of the
enhance the estimating abilities by recent studies are concerned point with other neighboring points, with-
discussed as follows. out the need for complex parametrization.38 Even at
the output, the form can be extended to two-dimen-
Treatment to the input and output data for neural network sional and three-dimensional pressure distribution
training. Conventionally, the input data to the ANN is over a solid body of aircraft, which can be beneficial
simply geometric parameters in geometry-aerody- for aerodynamic design.11 For example, a signed dis-
namic performance surrogate modeling, e.g. wing tance function (SDF) sampled on a Cartesian grid was
planform, airfoil geometry and flight condition, proposed as a universal representation for different
etc.28 For example, wing planform parameters of geometries, which is shown to be effective for convo-
UAV design were determined through an aero- lutional neural networks.37 SDF not only provides
dynamic optimization process using both genetic local geometry details, but also contains additional
algorithms (GAs) and ANNs.36 The number of par- information of the global geometry structure.
ameters should be kept as little as possible; otherwise Another solution is data filtering via principal com-
the sufficiency of sample data would be difficult to be ponent analysis (PCA), a statistical procedure that
satisfied in the network training. A class/shape func- uses an orthogonal transformation to convert a set
tion transformation (CST) geometry parametrization of observations of possibly correlated variables into
6 Proc IMechE Part G: J Aerospace Engineering 0(0)

function, the generic form of relation between the

input and output variables can be expressed as
Ivakhnenko polynomial.

Utilization of various levels of information. As a matter of

fact, the data in the design space can be classified by
the information conveyed by data itself.41 Different
levels of information hidden in physics can be
extracted via different procedures. For example, mul-
tiple design alternatives have to be quickly iterated in
preliminary design to make initial decisions without
high-fidelity simulations. Similarly, the multilevel sur-
rogate modeling (also termed as variable-resolution
model) is proposed to obtain the optimal area in the
design space quickly, after which the search for the
optimal point location is held in the neighboring
Figure 4. Classification of airfoil database via SOM. region of a smaller area in design space thus saving
computational resource. Designers can even directly
apply the deep learning approximation model in
design space exploration algorithms without training
extra lower-dimensional surrogate models.37 One such
application is a multilevel surrogate-based aerofoil
shape optimization.42 As three fundamental parts of
ANN, i.e. the form that input data takes, the type,
topology and configuration of neural network, and
the form the output data takes, the data flow should
be complete as well as well-arranged. The progress
can also be made at the procedure of surrogate mod-
eling, which can be in new form other than simple
data bunching.

Unexploited hidden information in optimization

ANN is used to utilize the resource of accumulated
airfoil data so that it can be able to learn from large
amount of data, instead of using rule-based program-
ming.43 There is still room for improvement in making
ANN learn in a ‘‘smarter’’ way. The hidden informa-
Figure 5. Framework of the PCA–ANN-based inverse design tion has to be exploited so that the design space con-
model. tains new perspectives beneficial to optimization.
Take drag reduction for example: the flow phenom-
enon involves many factors that may have impact on
a set of values of linearly uncorrelated variables ‘‘prin- the aerodynamic performance, e.g. the position and
cipal components’’. PCA reduces the dimensions of vibrating frequency of separating in buffets in the
the problem with kernel function in interpolating non- boundary layer of transonic airfoil; the peak of pres-
linear problems, thus saving computational resources. sure value over the surface that indicates shock wave
In the inverse design model of stall lift robustness for position and the distribution of the pressure that
high-lift device, PCA is applied to operate on the determines the profile drag of an aircraft wing; the
input data to the network being trained and obtains location and emerging time of vortex cores; the distri-
satisfactory result (Figure 5).39 bution of boundary layer thickness. A good surrogate
Similarly, there are also studies about the treat- model should contain the above information during
ment of output data of neural network training. For neural network training, while does not mix all the
example, grouped method of data handling (GDMH) useful information into a mess.
neural network can be used in order to transform dis- It is known that different aerodynamic perform-
crete CFD data into continuous function.40 The para- ances of two airfoils are result of their different geom-
mount goal of GDMH modeling is to generate a etry.44 In this approach, ANN is expected to correlate
quadratic polynomial function in a feed-forward net- the relationship; thereby new airfoil geometry that
work whose coefficients are obtained with regression leads to better aerodynamic performance will be
technique. By means of a complex polynomial able to be produced. The authors of this paper have
Sun and Wang 7

done similar research in exploiting airfoil database of where the GA operates simultaneously with ANN.36
geometry and aerodynamic performance from accu- ANN is employed as one of the reliable and fast meth-
mulated experiment and CFD calculation results ods of predicting aerodynamic coefficients to
based on ANN.45 The proposed approach ‘‘database select optimized airfoils10 as well as a lift-drag ratio
self-expansion’’ is focused on quickly providing new optimization for airfoil46 with GA. Approximated
airfoil candidate that has better aerodynamic per- pre-evaluations based on ANN are used in a hybrid
formance (Figure. 6), which is different from other optimization procedure with GA to determine airfoil
algorithms e.g. particle swarm optimization, ant shape in three-dimensional wing design to benefit
colony optimization, etc. The database is classified from the accumulated knowledge thus reducing
for data concentration via SOM, as is shown in
Figure 7. Conclusively, the comparisons of geometry
before and after optimization are shown in Figure 8.
The corresponding comparison of aerodynamic per-
formance is shown in Figure 9.

Collaboration with other optimization algorithms. The pos-

sibilities of combining the advantages of different
optimization algorithms are studied by many
researchers. The heuristic algorithm of GA is usually
chosen due to its simplicity to implement, their
robustness, and flexibility in different situation of
various problems. One big disadvantage associated
with GA is that they are computationally time-con-
suming especially in aerodynamic optimization where
a CFD solver is used for the fitness function evalu-
ation. This makes the use of ANN an efficient way to
reduce the computational time, since ANN decouples
the aerodynamic solver from the optimization process Figure 7. SOM geometry classification situation.

Figure 6. Illustration of database self-expansion flow chart.

8 Proc IMechE Part G: J Aerospace Engineering 0(0)

implementing enhanced neural networks (ENNs)

have been conducted in establishing a hybrid opti-
mizer, which is executed to search for the first tenta-
tive optimal point.12 The analysis code is performed
on the tentative optimal point to check the difference
between the surrogate model ENN and the analysis
code.

Deep learning in surrogate modeling. In MLP architec-

ture, the learning capability can be increased by
Figure 8. Comparison of initial airfoil and the new airfoil’s adding hidden layers and/or units in hidden layer.
upper curve by database self-expansion. However, the trade-off space between the network
size and the learning capability is mainly determined
by stereotyped variability due to the underlying
assumption of ‘‘fully-connected’’ network structure.
In this context, ‘‘deep learning’’ has attracted many
studies in recent years, although it is just a sub-
concept under ANN. Compared with its conventional
MLP counterparts, deep learning is equipped with
more training layers and characterized layer (e.g. con-
volutional layer and pooling layer in the famous
LeNet-5).49 Conventional MLP are compared with
the convolutional neural network (CNN) results: the
deep learning surrogate modeling exhibits a competi-
tive prediction accuracy with minimal constraints in
geometric representation.50 Deep learning is enabled
to learn invariant high-level features when the data
have strong spatial and/or temporal correlations.
Despite the limitation in application (e.g. progress
are mostly seen in the frontier of image recognition)
Figure 9. The skin friction coefficient Cf and pressure coef- and the requirement of sample data of higher magni-
ficient Cp distribution at specified position of nacelle surface tude, deep learning is beneficial in the field of aero-
before and after database self-expansion. dynamic design surrogate model because it no longer
needs the a priori treatment to the input data (e.g. the
hand-crafting of features by experts), thus human
the number of CFD evaluations required at each experience can be less relied on and new perspective
generation.47 of design space can be created.
There are also other types of combination of opti- There are an increasing number of applications of
mization algorithms with ANN. A hybrid model that deep learning in aerodynamic design, most of which
combines state-space model (supported by wind- takes different forms than conventional ANN,
tunnel experimental data) and ANN is proposed to although the principle of surrogate model is not chan-
describe the aircraft unsteady aerodynamic character- ged. CNN is able to estimate the velocity field two
istics.48 Thereby, the separation point model in state- orders of magnitude faster than GPU-accelerated
space representation is reserved to describe the time CFD solver and four orders of magnitude faster
delay of the unsteady aerodynamic responses, while than a CPU-based CFD solver at a cost of a low
the conventional polynomial model is replaced by error rate.37 In a long endurance UAV aerofoil
ANN to improve accuracy and universality. An archi- design optimization, repetitively enhanced neural net-
tecture combining a variational autoencoder mapping works (RENN) method is developed and presented
shapes to latent representations and Gaussian process for complex and implicit engineering design problems,
regression is jointly trained to generate improved which constructs an accurate surrogate model and
shapes in the two-dimensional case.9 A novel algo- avoids over-fitting during neural networks training
rithm to estimate the optimum value of the fuselage from supervised learning data.12 The optimizer seeks
drag coefficient is designed by integrating ANN into a tentative optimum point, which is then repetitively
the algorithm of simultaneous perturbation stochastic added into the supervised learning data until tolerance
approximation (SPSA).20 is reached. CNN has been applied to map airfoil
There are also attempts of using different shapes to pressure distribution under the framework
approaches within the domain of ANN surrogate of classification problem using discretized pressure
modeling, although the basic pattern is quite similar coefficient.51 In the testing phase, a new pressure coef-
to one another. For example, the surrogate models by ficient distribution is given to the CNN model,
Sun and Wang 9

generating an airfoil shape that is close to the asso- ORCID iD

ciated airfoil with an average L2 error of less than 2%. Shuyue Wang https://orcid.org/0000-0001-8452-5602
Using conditional generative adversarial networks
(cGAN), new data-driven models are trained by References
deep learning for direct generation of solutions to 1. Abbas A, De Vicente J and Valero E. Aerodynamic
steady-state heat conduction and incompressible technologies to improve aircraft performance. Aerosp
fluid flow purely on observation without knowledge Sci Technol 2013; 28: 100–132.
of underlying physical models.52 2. Botez RM. Green aircraft minimum fuel consumption
methodologies. In: Oral or poster sessions, during
the GLOREP 2018 conference, Timisoara, Romania,
Conclusions 15–17 November 2018.
ANN surrogate modeling offers a feasible approach to 3. ACARE. Aeronautics and air transport beyond vision
aerodynamic design in the field of rapid investigation of 2020 (towards 2050), 2010.
4. Sky C. Clean sky. European Union Funding for Research
design space and optimal solution searching. The direct
and Innovation, 2018.
motive of a surrogate model is to establish a data-driven
5. Green BE, Whitesides JL, Campbell RL, et al. Method
and bottom-up approach used when an outcome of for the constrained design of natural laminar flow air-
interest cannot be easily obtained. The principle behind foils. J Aircraft 1997; 34: 706–712.
surrogate modeling is that data at input and output are 6. Wang S, Sun G and Li C. Natural laminar flow opti-
related through the pattern of the trained neural net- mization of transonic nacelle based on differential evo-
work. Introducing the basic principle of ANN as an lution algorithm. J Aerosp Eng 2019. DOI: 10.1061/
excellent surrogate modeling method, this paper focuses (ASCE)AS.1943-5525.0001028.
on its data treatment and configuration setup. 7. Kani JN and Elsheikh AH. Dr-rnn: a deep residual
ANN has successfully helped in direct applications recurrent neural network for model reduction. arXiv
of ANN surrogate modeling in aerodynamic design. preprint arXiv:170900939, 2017.
For example, the numerical search for the optimum 8. Forrester AI and Keane AJ. Recent advances in surro-
gate-based optimization. Prog Aerosp Sci 2009; 45:
shape is of great interest for aircraft and turboma-
50–79.
chinery designers. This paper discusses the selection 9. Eismann S, Bartzsch S and Ermon S. Shape optimiza-
of the type of neural network that may have influence tion in laminar flow with a label-guided variational
on the optimization effect. It is noted that ANN needs autoencoder. arXiv preprint arXiv:171203599, 2017.
a certain data set to be effectively trained. Treatment 10. Xu Z, Saleh JH and Yang V. Optimization of supercrit-
to the input and output data for neural network train- ical airfoil design with buffet effect. AIAA J 2018.
ing has also been discussed. In particular, new forms 11. Sekar V, Zhang M, Shu C, et al. Inverse design of airfoil
of geometric representation method are proposed spe- using a deep convolutional neural network. AIAA J
cifically to the network training. 2019.
ANN not only performs well in surrogate model- 12. Van Nguyen N, Lee J, Tyan M, et al. Repetitively
ing, but also produces potential advantages in opti- enhanced neural networks method for complex engin-
eering design optimisation problems. Aeronaut J 2015;
mization. New frontiers of modern ANN surrogate
119: 1253–1270.
modeling are reviewed in this paper. As for the util- 13. Samareh JA. A survey of shape parameterization tech-
ization of various levels of information, the multilevel niques. ntrsnasagov 1999.
surrogate modeling is proposed to obtain the optimal 14. Andradóttir S and Prudius AA. Balanced explorative
area in the design space quickly thus saving compu- and exploitative search with estimation for simulation
tational resource. There has also been attempts of optimization. Informs J Comput 2009; 21: 193–208.
sorting out optimization direction in the accumulated 15. Barton RR. Simulation optimization using metamodels.
database. Also, there are many types of combination In: Proceedings of the 2009 winter simulation conference
of optimization algorithms with ANN. Lastly, deep (WSC), 2009, pp.230–238. New York: IEEE.
learning is beneficial to the field of aerodynamic 16. Sun G, Sun Y and Wang S. Artificial neural network
design surrogate model because it no longer needs based inverse design airfoils and wings. Aerosp Sci
the a priori treatment to the input data and thus Technol 2015; 42: 415–428.
human experience can be relied less. Meanwhile, 17. Yu J and Hesthaven JS. Flowfield reconstruction
method using artificial neural network. AIAA J 2018;
new perspective of design space can be created.
57: 482–498.
18. Kou J and Zhang W. Multi-kernel neural networks for
Declaration of Conflicting Interests nonlinear unsteady aerodynamic reduced-order model-
The author(s) declared no potential conflicts of interest with ing. Aerosp Sci Technol 2017; 67: 309–326.
respect to the research, authorship, and/or publication of 19. Parish EJ and Duraisamy K. A paradigm for data-
this article. driven predictive modeling using field inversion and
machine learning. J Comput Phys 2016; 305: 758–774.
Funding 20. Oktay T, Çelik H and Uzun M. A novel learning algo-
The author(s) received no financial support for the research, rithm to estimate the optimum fuselage drag coefficient.
authorship, and/or publication of this article. Sakarya Univ J Sci 2017; 21: 63–68.
10 Proc IMechE Part G: J Aerospace Engineering 0(0)

21. Iuliano E and Pérez EA. Application of surrogate-based 38. Chen H, Deng K and Li R. Utilization of machine
global optimization to aerodynamic design. New York: learning technology in aerodynamic optimization.
Springer, 2016. Acta Aeronaut Astronaut Sin 2019; 40.
22. Oroumieh MAA, Malaek SMB, Ashrafizaadeh M, et al. 39. Wang X, Wang S, Tao J, et al. A pca–ann-based inverse
Aircraft design cycle time reduction using artificial design model of stall lift robustness for high-lift device.
intelligence. Aerosp Sci Technol 2013; 26: 244–258. Aerosp Sci Technol 2018; 81: 272–283.
23. Hopfield JJ. Neural networks and physical systems with 40. Safikhani H and Jamalinasab M. Pareto optimization
emergent collective computational abilities. Proc Natl of two-element wing models with morphing flap using
Acad Sci 1982; 79: 2554–2558. computational fluid dynamics, grouped method of data
24. Berke L, Patnaik S and Murthy P. Optimum design handling artificial neural networks and genetic algo-
of aerospace structural components using neural net- rithms. Int J Eng-Trans A 2018; 31: 666–672.
works. Comput Struct 1993; 48: 1001–1010. 41. Timnak N, Jahangirian A and Seyyedsalehib S. An
25. Aggarwal CC. Neural networks and deep learning. optimum neural network for evolutionary aerodynamic
New York: Springer, 2018. shape design. Scient Iran Trans B 2017; 24: 2490–2500.
26. Agarwal A and Biegler LT. A trust-region framework 42. Koziel S and Leifsson L. Multi-level surrogate-based
for constrained optimization using reduced-order mod- airfoil shape optimization. In: 51st AIAA aerospace
eling. Optim Eng 2013; 14: 3–35. sciences meeting including the new horizons forum and
27. Nilsson NJ. Introduction to machine learning. An early aerospace exposition, Dallas, Texas, 2013, p.778.
draft of a proposed textbook. 1998. http://robotics stan- 43. Ram RKG, Cooper YN, Bhatia V, et al. Design
ford edu/people/nilsson/mlbook html optimization and analysis of naca 0012 airfoil using
28. Secco NR and Mattos BSd. Artificial neural networks computational fluid dynamics and genetic algorithm.
to predict aerodynamic coefficients of transport air- Appl Mech Mater 2014; 664: 111–116.
planes. Aircraft Eng Aerosp Technol 2017; 89: 211–230. 44. Ganesh Ram R, Cooper YN, Bhatia V, et al. Design
29. Kocuk B, Altinel IK _ and Aras N. Approximating the optimization and analysis of naca 0012 airfoil using
objective function’s gradient using perceptrons for con- computational fluid dynamics and genetic algorithm.
strained minimization with application in drag reduc- Appl Mech Mater 2014; 664: 111–116.
tion. Comput Oper Res 2015; 64: 139–158. 45. Wang S, Sun G, Chen W, et al. Database self-expansion
30. Mazhar F, Khan AM, Chaudhry IA, et al. On using based on artificial neural network: an approach in air-
neural networks in uav structural design for CFD data craft design. Aerosp Sci Technol 2018; 72: 77–83.
fitting and classification. Aerosp Sci Technol 2013; 30: 46. Haryanto I, Utomo TS, Sinaga N, et al. Optimization
210–225. of maximum lift to drag ratio on airfoil design based on
31. Jean R, Miroslav K and Atle J. Artificial neural artificial neural network utilizing genetic algorithm.
networks trained through deep reinforcement learning Appl Mech Mater 2014; 493: 123–128.
discover control strategies for active flow control. arXiv 47. Duvigneau R and Visonneau M. Hybrid genetic algo-
preprint arXiv:180807664, 2018. rithms and artificial neural networks for complex design
32. Ju Y, Zhang C and Ma L. Artificial intelligence meta- optimization in CFD. Int J Numer Methods Fluids 2004;
model comparison and application to wind turbine 44: 1257–1278.
airfoil uncertainty analysis. Adv Mech Eng 2016; 8: 48. Guang O, Jun L and Ping Z. Aircraft unsteady
1687814016647317. aerodynamic hybrid modeling based on state-space
33. Sinha M, Kuttieri RA, Ghosh AK, et al. High angle of representation and neural network. In: International
attack parameter estimation of cascaded fins using conference on pattern recognition applications and meth-
neural network. J Aircraft 2013; 50: 272–291. ods, vol. 2, Porto, Portugal, 2017, pp.232–239.
34. Koziel S and Leifsson L. Knowledge-based airfoil shape 49. LeCun Y, Bottou L, Bengio Y, et al. Gradient-based
optimization using space mapping. In: 30th AIAA learning applied to document recognition. Proc IEEE
applied aerodynamics conference, New Orleans, 1998; 86: 2278–2324.
Louisiana, 2012, p.3016. 50. Zhang Y, Sung WJ and Mavris DN. Application of
35. Chen J, Sun G and Jin X. Intelligent aerodynamic convolutional neural network to predict airfoil lift
design for airfoil based on artificial neural network coefficient. In: 2018 AIAA/ASCE/AHS/ASC structures,
method. In: 2nd international conference on computer structural dynamics, and materials conference,
and automation engineering, ICCAE, Singapore, 2010, Kissimmee, Florida, 2018, p.1903.
pp.428–436. 51. Yilmaz E and German B. A convolutional neural
36. Boutemedjet A, Samardžić M, Rebhi L, et al. Uav aero- network approach to training predictors for airfoil
dynamic design involving genetic algorithm and artifi- performance. In: 18th AIAA/ISSMO multidisciplinary
cial neural network for wing preliminary computation. analysis and optimization conference, Denver,
Aerosp Sci Technol 2019; 84: 464–483. Colorado, 2017, p.3660.
37. Guo X, Li W and Iorio F. Convolutional neural net- 52. Farimani AB, Gomes J and Pande VS. Deep learning
works for steady flow approximation. In: Proceedings the physics of transport phenomena. arXiv preprint
of the 22nd ACM SIGKDD international conference on arXiv:170902432, 2017.
knowledge discovery and data mining, ACM, San
Francisco, California, 2016, pp.481–490.