Aerospace Science and Technology 115 (2021) 106845

Contents lists available at ScienceDirect

Aerospace Science and Technology

www.elsevier.com/locate/aescte

On the application of artificial neural network for the development of

a nonlinear aeroelastic model
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero, A. Cremades ∗
CMT-Motores Térmicos, Universitat Politècnica de València, Camino de Vera s/n, Valencia, 46022, Spain

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

Article history: Aeroelastic Computational Fluid Dynamics simulations have traditionally been associated to a high
Received 10 November 2020 computational cost, making them prohibitive in a initial phase of the design. Analytic models, which
Received in revised form 21 April 2021 may not be accurate for nonlinear aerodynamics, have traditionally been utilized in order to size those
Accepted 18 May 2021
structures. Recently, some authors have proposed the use of artificial neural networks to reduce the error
Available online 26 May 2021
Communicated by Damiano Casalino
in the prediction of aerodynamic coefficients of bluff bodies, which have separated flow over a substantial
part of its wetted surface. This article proposes a method based on neural networks for calculating
Keywords: the dynamic aerodynamic coefficients of a flat plate. The procedure, which is applied for different
Aeroelasticity network typologies (feed-forward and long-short term memory neural networks), is, then, coupled with
Artificial neural network a structural solver in order to create an aeroelastic reduced order model. The results are compared with
Fluid structure interaction CFD aeroelastic simulations, showing a high reduction of computational cost (99%) without penalties
Nonlinear aerodynamics in the accuracy. The instabilities are captured and the mean deformation, amplitude and frequency of
Stall flutter
the motion are predicted. In addition, the different neural network models are compared evidencing
Computational fluid dynamics
that for the aeroelastic calculation feed-forward networks are most efficient in terms of accuracy and
computational cost.
© 2021 The Author(s). Published by Elsevier Masson SAS. This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction The study of the aeroelastic problem may be carried out using
different tools and approaches. A research might be conducted an-
Modern structures are designed to reduce the amount of mate- alytically [13], numerically [14] or experimentally [15]. Among the
rial required to prevent failure. As the frame weight is decreased, latest, the studies may be performed directly on a manufactured
the structure becomes more efficient and environmentally sustain- product or using a wind tunnel model. Relative to the first option,
able. Moreover, saving resources supposes a reduction on its cost Evans et al. [16] carried out measurements of the deformation of a
of production, obtaining a more competitive product [1]. However, 5 kW wind turbine. Argentini et al. [17] used a wind tunnel model
the improvement on the weight is also related with a decrease of to analyze the buffeting response of long span bridges. Although
the torsion and bending stiffness of the frame, evidencing the ne- wind tunnels are useful to reproduce the wind flow under con-
cessity of finding a Pareto-optimal solution of structural integrity trolled conditions, these experiments have many restrictions as the
and cost. For applications where the structure is exposed to wind scale effects, wall and support interference and aerodynamic dis-
loads, the reduction of stiffness may increase the effect of the tortion [18]. These limitations evidence the importance of using
aeroelastic phenomena, producing instabilities at low free stream other methods, such as Computational Fluid Dynamics (CFD) sim-
velocities [2]. ulations, to verify the results.
Due to the previous reasons, aeroelastic analysis has been a hot Additionally, numerical simulations may also be used to calcu-
topic in aerospace [3], civil [4] or automotive [5] engineering, and late aeroelastic phenomena. In order to calculate the whole set
a wide amount of literature dealing with these phenomena can be of equations, a CFD / Computational Solid Dynamics (CSD) anal-
found. Some researchers analyzed divergence experimentally [6], ysis could be performed. For instance, Dong et al. [19] and Ren
analytically [7] or numerically [8]. Other authors focused on other et al. [20] predicted flutter in turbomachinery and Wang et al. [21]
effects as flutter [9], [10] or buffeting [11], [12]. calculated a 3D wind turbine blade. Nevertheless, the computa-
tional cost associated with the resolution of the complete set
of equations is unaffordable for most engineering problems and,
* Corresponding author. therefore, many authors as Peng and Jinglong [22], Tsushima and
E-mail address: [email protected] (A. Cremades). Su [23], Amooozgar et al. [24] and Kwon et al. [25] reduced the

https://doi.org/10.1016/j.ast.2021.106845
1270-9638/© 2021 The Author(s). Published by Elsevier Masson SAS. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero et al. Aerospace Science and Technology 115 (2021) 106845

Nomenclature

Symbols clα̈ Lift coefficient angle of attack acceleration derivative

dyn
α Angle of attack cl Dynamic lift coefficient
α Angle of attack amplitude of the motion clst Steady lift coefficient
θ Pitch amplitude of the motion cp Power coefficient
t Time step f Frequency of the motion
x Cell size f 1 , f 2 , f 3 , f 4 Intermediate steps in 4th order Runge-Kutta
w Plunge amplitude of the motion H Width of the channel
 E cm Energetic error of the oscillation cycle h Thickness of the plate
 Amplitude of the forced oscillations I∗ Nondimensional inertia of the 2D section
θ Pitch of the section I 2D Inertia of the 2D section
θ∗ Nondimensional pitch of the section k∗ Nondimensional stiffness
θ0 Initial pitch of the section kθ Torsion stiffness
θ̄ Mean pitch of the section kw Bending stiffness
μ∞ Free stream viscosity k∗w Nondimensional bending stiffness
ρ∞ Free stream density Lu Upstream length of the domain
A Matrix of the linear system Ld Downstream length of the domain
a∞ Free stream sound speed m∗ Nondimensional mass of the 2D section

b Vector of the linear system m2D Mass of the 2D section
c Chord of the plate Ma Mach number
cm Moment coefficient Re Reynolds number
dyn
cm Dynamic moment coefficient T Period of the oscillation
st t Time
cm Steady moment coefficient
cl Lift coefficient V∞ Free stream velocity
cl0 Zero angle of attack lift coefficient w Plunge of the section
cl α Lift coefficient angle of attack derivative w̄ Mean plunge of the section
clα̇ Lift coefficient angle of attack velocity derivative y+ Nondimensional wall distance

complexity of the problem by using Euler equations, 2D unsteady papers, as Abbas et al. [39] and Li et al. [40], the artificial neu-
potential model and modified small disturbance theory, respec- ral network is fitted using CFD results. In Abbas et al. [39] the
tively, for the calculation of the aerodynamic loads. Other re- network is joined with a structural solver in order to obtain the
searchers, as You et al. [26] have simplified the structure to a 1D aeroelastic response. However, the aerodynamic model only uses
beam reducing the cost of the computation of the CSD solver. feed forward networks (FFN) and, thus, the capabilities of deep
In order to reduce the computational cost associated with the learning tools such as Long Short Term Memory (LSTM) networks
3D aeroelastic problem, many researchers calculate a characteris- are not evaluated. Furthermore, in Li et al. [40], an LSTM net-
tic aeroelastic section [27], which has been traditionally used in work is used to calculate the unsteady aerodynamic behavior of a
literature [28]. For instance, in the work of Camilo et al. [29], the bridge deck. Nevertheless, the author does not compare the LSTM
aeroelastic response of a 2D section using CFD is analyzed to ob- solutions with FFN, evaluating the convenience, precision and lim-
tain the response of nonlinear structures. Even though, the cost of itations of each typology.
computing a wide range of unsteady 2D simulations is still high. This paper aims at completing the research developed in the
In order to reduce computing time, different Reduced Order literature, proposing a reduced order model (ROM) for decreasing
Models (ROM) have been proposed [30]. A great amount of the
computing time in the initial design of beams under wind loads.
aerodynamic ROMs are based on proper orthogonal decomposition
The ROM will be applied to a 2D flat plate geometry immersed
(POD). Although these methods proved to be an accurate model in
in a wind channel. The flat plate is chosen because it experiences
many problems, as the optimization of a compressor [31] or the
nonlinear aerodynamic phenomena even for low values of the an-
aeroelastic calculation of a wing [32], POD “ROMs are only suitable
gle of attack. In addition, the channel width is taken in a way such
for a frozen aeroelastic model configuration” [30]. The associated
that the aerodynamics of the system will be influenced by the ver-
frequencies of the ROM are dependent upon the mass and stiffness
tical position of the plate. Two different surrogate models, using
distribution and, therefore, a change on the structure will lead to
FFN and LSTM networks are developed to estimate the aerody-
a different motion and eigenvalues of the flow field and, thus, to a
different POD. namic loads. These models are fed by means of CFD calculations,
On the other hand, artificial neural networks (ANN) have and they are able to provide an accurate estimation of the nonlin-
demonstrated to be capable of predicting a wide range of physical ear unsteady aerodynamic forces acting over the plate. The ANNs
phenomena, as compressor performance [33], volumetric efficiency are coupled with a structural model, allowing the estimation of the
of ICE [34] or diesel engine emissions [35]. Moreover, some authors nonlinear aeroelastic features with a relatively low computational
have applied ANN to solve aerodynamic [36] and aeroelastic prob- cost.
lems. For instance, Wu and Kareem [37] modeled the hysteretic The paper is structured as follows. First, in Section 2, the main
nonlinear behavior of bridge decks under aerodynamic load. In geometrical features of the test case will be described. Later, in
their paper, the aerodynamic coefficient is splitted into static and Section 3, a full description of the theoretical bases of the problem
dynamic coefficients in order to obtain the nonlinearities in a tur- are provided. After that, in Section 4, the ANN and CFD methodolo-
bulent flow. Another example is the research of Chen et al. [38], gies will be explained in order to obtain the main results, which
in which the flutter derivatives of a flat plate are calculated using will be collected later, at Section 5. Finally, the main conclusions
an artificial neural network fed with experimental data. In other of the work will be provided, at Section 6.

2
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero et al. Aerospace Science and Technology 115 (2021) 106845

Fig. 1. Domain of the simulation (not to scale), 3D plate simulation (left) and 2D simplification of the problem (right).
 
1 dyn
2. Description of the test case m2D ẅ + k w w = 2
2
ρ∞ cV ∞ clst (α , w ) + cl (ᾱ , α , α̇ , α̈ ) →
m∗ w¨∗ + k∗w w ∗ = clst (α , w ) + cl
dyn
(ᾱ , α , α̇ , α̈ )
The test case consists of a flat plate immersed in a wind tun-
nel, taken from Gil et al. [41], as shown in Fig. 1. The structure is (2)
reduced to a combination of linear and torsional springs, Fig. 1b.
where the angle of torsion of the plate is θ , the aerodynamic angle
The flat plate section is selected due to its aerodynamic nonlineari-
of attack is α , w is the plunge and θ ∗ and w ∗ are the nondimen-
ties for low angles of attack, as well as its similarities to important
sional torsion and plunge. The nondimensional numbers governing
engineering elements such as solar panels. The channel is also in-
cluded in the analysis to take into account the effect of the vertical the behavior of the plate (the nondimensional torsional stiffness,
position in the aerodynamic steady coefficients. The plate has a Equation (3), the nondimensional inertia, Equation (4), the nondi-
chord c = 0.1 m, and a thickness h = 0.04c. The channel presents mensional flexural stiffness, Equation (5) and the nondimensional
the following dimensions: upstream length L u = 5c, downstream mass, Equation (6)) may be calculated in order to reproduce a
length L d = 15c and a width of H = 4c, Fig. 1a. Upstream and three-dimensional structure, as demonstrated by Gil et al. [41]:
downstream dimensions are taken in a manner that the inlet and kθ
outlet boundary conditions do not affect the simulation. k∗ = 1
(3)
The walls of the channel are supposed sufficiently far from the 2
ρ∞ v 2∞ c 2
plate for assuming slip boundary conditions (Euler walls) with- I 2D
out affecting substantially the aerodynamic response of the flat
I∗ = 1
(4)
2
ρ∞ c 4
plate, similarly as in [42]. The inlet flow is perpendicular to the
inlet section. The characteristics of the flow are: density ρ∞ = kw
k∗w = (5)
1.18 kg/m3 , viscosity μ∞ = 1.85 · 10−5 Pa s and a sound speed 1
2
ρ∞ v 2∞ c
of a∞ = 340 m/s. The velocity V ∞ is varied between 7 m/s and m2D
ρ V c
42 m/s. Thus, Reynolds number (Re = ∞μ ∞ ) changes between m∗ = 1
(6)
∞
V∞ 2
ρ∞ c 2
4.5 · 10 and 2.7 · 10 ; and Mach number (Ma =
4 5
between a∞
)
0.02 and 0.12. The holding structure of the plate is reproduced Following as Wu and Kareem [37], the aerodynamic coefficients
by means of two springs of torsional and linear stiffness kθ = (cl and cm ) of equations (1) and (2) are expressed as the sum of
dyn dyn
30.46 Nm/m and k w = 558.46 N/m2 respectively, Fig. 1b. Finally, the stationary, clst and cm
st
, and the dynamic terms, cl and cm .
the mass and inertia of the section are m2D = 0.36 kg/m and In addition, ᾱ is the average value of the angle of attack of the
I 2D = 3.9 · 10−4 kgm2 /m. oscillation and α is the increment of angle of attack with respect
to ᾱ . The angle of attack might be defined as a function of the
3. Theoretical background motion variables θ and w:
 
ẇ ẇ
In this section, the main aeroelastic equations are introduced. In α = θ − arctan ≈θ − (7)
V∞ V∞
addition, the basis of artificial neural network is exposed, showing
the principal typologies used in this work. Finally, the CFD models 3.2. Artificial neural networks
are presented.
ANN are routinely used for regression problems, easily approx-
3.1. Aeroelastic equivalent section imating any continuous function. In this work, ANN are trained
with data from the aerodynamic simulations, leading to improve-
The motion of the equivalent 2D section is governed by Equa- ments in the estimations of the aerodynamic performance of the
tions (1) and (2), which states the equilibrium for torsional and plate over those of the stationary and linear aerodynamic mod-
vertical degrees of freedom respectively. els. The networks have been calculated by using TensorFlow with
  Keras engine. In this paper, two different models have been tested
1 dyn
I 2D θ̈ + kθ θ = 2
ρ∞ c 2 V ∞
2 st
cm (α , w ) + cm (ᾱ , α , α̇ , α̈ ) → and compared for predicting the nonlinear aerodynamic coeffi-
cients of a flat plate: Feed Forward Networks and a Long Short
I ∗ θ¨∗ + k∗ θ ∗ = cm
dyn
st
(α , w ) + cm (ᾱ , α , α̇ , α̈ ) Term Memory networks. FFN obtain the value of the coefficients
(1) using the input variables of the calculated time step. On the other

3
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero et al. Aerospace Science and Technology 115 (2021) 106845

Fig. 2. Schematic architecture of the ANN.

Fig. 3. Scheme of the work flow of the aeroelastic ROM. The procedure includes training, validation and application of the ANN.

hand, LSTM networks accumulate the information of all the pre- training and validation is presented. The optimum number of neu-
vious time steps [43]. Both types of networks have been tested rons is calculated, the data set error is analyzed and the results of
with the same architecture, Fig. 2. All the studied networks use a the training and test of the network are shown. Finally, the cou-
multi layer perceptron with three layers: input layer, hidden layer pled solver algorithm is described and the numerical models are
and output layer. The number of neurons of the hidden layer, N 1 , described. The complete methodology of the paper is summarized
is optimized in terms of computational cost and accuracy, as ex- in the diagram of Fig. 3.
plained in Section 4. In addition, as the aerodynamic coefficients
are continuous, the activation functions of the neurons must also 4.1. CFD: data base construction
be continuous.
In order to use ANN, the network must be trained using data, The training data set is constructed with a matrix of 324 CFD
which can be obtained from different sources. In this work, a vir- bidimensional forced oscillation simulations, which are selected to
tual CFD wind tunnel has been used to generate the information be representative of a wide range of physical conditions. In addi-
of the training data. However, the method admits data obtained tion, the initial angle of attack and the amplitude of the movement
from different sources. The training process has been performed is limited to the range in which vortex shedding is negligible,
for both types of network using a RMSprop algorithm, until train- which is similar to many engineering problems, as solar panels
ing and test loss function reach an asymptotic behavior or start to in stow position [48]. Strong vortex shedding leads to important
increase, to prevent from overfitting. random oscillations of the aerodynamic coefficients, decreasing the
accuracy of the network under interest conditions. Finally, the
3.3. Computational fluid dynamics fluid-structure interaction is calculated through a strongly coupled
algorithm, which shows stability under strong nonlinear interac-
The aerodynamic forces acting on the test case can be ob- tions [49].
tained using different tools and approaches. One of these tools is The training simulations represent a forced oscillatory pitching
CFD, which calculates the solution of the conservation equations motion with the formulation of equation (8):
of the fluid. Here, the simulations, which have been calculated us-
ing commercial software Simcenter STAR-CCM+® , make use of the
θ = θ0 +  sin (2π f t ) (8)
Unsteady Reynolds-averaged Navier-Stokes equations [44]. A k − ω
with shear stress transport (SST) turbulence model is utilized be- being θ0 the initial pitch angle,  the amplitude of the motion and
cause it has been proved to predict accurate results under adverse f the frequency of the motion. The matrix of simulations contains
pressure gradient as the ones appearing in a flat plate ([45] and the training cases shown in Fig. 4.
[46]). The k − ω SST was proposed by Wilcox [47] to solve the lim- The simulation is performed using the commercial software
itations of the k − ω far from the walls, where a k −  model is Simcenter STAR-CCM+® over the domain of Fig. 1a. The discretiza-
used. This model requires to maintain the value of the nondimen- tion of the domain is made with a polygonal mesh, as presented
sional wall distance to the first layer of computational cells, y + , in in Fig. 6. A coupled solver with second order upwind ROE FDS
the viscous sublayer ( y + < 5) for the major part of the wall. scheme is used for computing the advection terms ([50], [51]),
whereas the gradients are computed with a hybrid Gauss-Least
4. Methodology Squares Method with Venkatakrishnan limiter [52]. The rotation
is established using an overset region in the near field [53]. A
Along this section the methodology for the different submodels mesh of 5.1 · 104 elements has been chosen making a mesh in-
and calculation is presented. First, the CFD data base construction dependence process based on Richardson’s extrapolation (RE) [54].
is detailed: the oscillations are defined and CFD mesh and setup A spatial discretization error lower than the 2% has been obtained
preparation is explained. Then the Artificial Neural Network model for the major part of the aerodynamic domain. The value of the

4
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero et al. Aerospace Science and Technology 115 (2021) 106845

Fig. 4. Training simulation data set matrix.

Fig. 5. Computation of aerodynamic lift and moment coefficient for different meshes. Comparison with RE. N is the number of elements of the mesh.

Fig. 6. Sketch of the computational mesh (not to scale) showing the different zones of refinement.

aerodynamic coefficients obtained using these methods can be ob- cillations and between 7 · 10−4 and 0.0042 in the aeroelastic cases.
served in Fig. 5. The final mesh has a size of xglobal = 0.02 m, In Fig. 7b, the distribution of the nondimensional wall distance y +
xwake = 0.004 m in the wake, xoverset = 0.002 m in the overset over the plate can be observed. Note that its value is lower than 5
and xwall = 0.0004 m. Some details of the computational mesh in the whole wall, ensuring that the boundary layer remains in the
can be seen in Fig. 6. viscous sublayer.
Relative to the time discretization, a second order scheme is
4.2. Artificial neural network: training and validation
utilized. The time step is selected to keep Courant-Friedrichs-Levy
number (C F L = t xV ) below 1, as observed in Fig. 7, in order to In this section, the procedure to fit the phenomena of Sec-
obtain enough time resolution, similarly to Torregrosa et al. [42] tion 4.1 with a neural network is presented. The influence of the
and Adeniyi et al. [55]. To reach the previous C F L, the simulation number of neurons is analyzed, evaluating the energetic global er-
uses a nondimesional time step of t Vc∞ = 0.002 for the forced os- ror of the cycles, in Section 4.2.1. In addition, the training and val-

5
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero et al. Aerospace Science and Technology 115 (2021) 106845

Fig. 7. CFL distribution and nondimensional wall distance over the plate.

Fig. 8. Power coefficient comparison between the ANNs and the CFD simulations for two representative number of neurons on the hidden layer. The cycle is shown for an
fc
initial incidence of 2.5 deg, an amplitude of 5 deg and a nondimensional frequency of V = 0.1250.
∞

idation mean squared errors are evaluated and analyzed in order To perform the neuron independence analysis, the energetic er-
to define the number of epochs required for the training process. ror of the aerodynamic cycle,  E cm , is calculated as the difference
All the studied networks have been trained using a RMSprop al- of power coefficients (c p = cm Vθ̇ c ) integrated in time for CFD and
∞
gorithm on 90% of the forced oscillation data. The remaining 10% ANN, as stated by Equation (9).
which corresponds to arbitrary movements is used for the valida-  t +T  t +T
tion. t (c p AN N )dt − t (c p C F D )dt
As stated in Section 3, the dynamic coefficient calculated by  E cm =  t +T (9)
means of the neural networks is added to a stationary aero- t
(c p C F D )dt
dynamic coefficient, called quasi-steady model. This quasi-steady In Fig. 8, the power coefficient comparison between CFD and
model interpolates, linearly, the coefficients from the steady CFD ANN is presented for a network of 2 and 50 neurons (Figs. 8a and
results, as shown in Section 5. Therefore, the models shown in 8b respectively). In the figure, an improvement of accuracy com-
the section are: CFD (2D forced oscillations), quasi-steady (aerody- pared with the steady solution is observed.
namic coefficients interpolated from steady polar), FNN and LSTM In Fig. 9, the cycle energetic error density function is shown as
(dynamic coefficient calculated by means of FNN and LSTM respec- a function of the number of neurons of the first layer for a FNN,
tively, added to the quasi-steady coefficient). Fig. 9a, and a LSTM network, Fig. 9b. Observe how for N 1 = 50, the
energetic error becomes approximately constant, being 50 neurons
4.2.1. Neuron independence analysis the optimal choice in terms of accuracy and computational cost.
The number of neurons used to calculate the aerodynamic In addition, the error of the network also depends on the num-
model states the number of weights and variables to adjust the ber of epochs of the training process [56]. The number of training
ANN. Therefore, it determines the accuracy and computing cost of epochs is selected to ensure that the logarithmic validation er-
training the network [43], but also the probability of network over- ror has either reached an asymptotic limit or started to increase.
fitting. A neuron independence analysis is performed to reduce the Fig. 10 shows how the validation mean squared error (MSE) flat-
number of neurons without affecting the results. This analysis is tens for the previous number of epochs. In addition, a k-fold cross
applied over the different configurations: FFN and LSTM. All the validation analysis [57] is performed, obtaining a mean prediction
studied ANN consist of an input layer, a hidden layer with N 1 error of 5.34 · 10−4 for 10-folds using the FNN model.
neurons and an output layer with N 2 = 2 neurons, as previously The distribution of the error in the whole data set is shown in
presented in Fig. 2. Fig. 11. This figure shows the distribution of the input variables,

6
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero et al. Aerospace Science and Technology 115 (2021) 106845

Fig. 9. Distribution of the energy error for the different number of neurons. (For interpretation of the colors in the figure(s), the reader is referred to the web version of this
article.)

Fig. 10. Evolution of the MSE as a function of the epoch of training for FNN and LSTM networks.

Fig. 11. Squared error distribution for the input variables. The plots on the diagonal show the distribution of the training variables, while the subfigures out of the diagonal
show the squared error distribution for each pair of input variables.

7
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero et al. Aerospace Science and Technology 115 (2021) 106845

Fig. 12. Training and validation of the FNN neural network.

the points where the neural network has been evaluated and the utilized to integrate the differential equations. For a deeper expla-
areas with the higher error. The squared error is demonstrated to nation of the numerical methods, the reader is referred to Sec-
be low for the analysis domain, presenting the higher errors for tions 4.3.1, 4.3.2 and 4.3.3. The aerodynamic forces are updated
the mean points of the oscillation near the stall point of the plate: each time step, the steady coefficients are calculated by a 2D in-
incidence of 4.5 deg and null increment of the angle of attack. terpolation function and the dynamic coefficients are obtained by
In Figs. 12 and 13 the performance of the ANN is shown for using an artificial neural network.
training and validation data. Both typologies of networks show
an accurate behavior in the simulation of nonlinear aerodynam- 4.3.1. Aeroelastic equations: state system
ics compared with the quasi-steady approach. Real and predicted The aeroelastic equations are linearly solved updating the aero-
values overlap in the major part of the cycles. Although the large dynamic forces and moments every time step. The aeroelastic lin-
amount of data makes it difficult to visualize the training curves, ear system can be expressed in terms of the matrix A and the
Fig. 14 evidences similar behavior in training and testing.  The vectors xt and x˙ t represent the states of the system
vector b.
In order to clarify the results shown in Figs. 12 and 13, Fig. 14 and its derivatives in the time step t respectively.
shows the linear regression between the real and the predicted
value of the coefficient, comparing it with the ideal non-error so- x˙ t = Axt + b →
lution. A fair agreement between ANN predictions and CFD calcu- ⎧ ⎫ ⎡ ⎤⎧ ⎫ ⎧ ⎫
0 0 1 0 ⎪ θt ⎪ 0 ⎪
⎪
⎪ θ̇ ⎪ ⎪ ⎪ ⎪
lations can be observed in this Figure. ⎨ t ⎪⎬ ⎢0 0
⎪
0 1 ⎥ ⎨ wt
⎪
⎬ ⎪
⎨0 ⎪
⎬
ẇ t
=⎢
⎣ − Ikθ
⎥
0 0 ⎦ ⎪ θ̇t
+ ρ∞ v ∞ c cm
2 2
4.3. Aeroelastic ROM algorithm ⎪
⎪ θ̈ ⎪ 0 ⎪ ⎪ 2I 2D ⎪
⎩ t ⎪ ⎭ 2D ⎪
⎩ ⎪
⎭ ⎪ ⎩ ρ∞ v 2 c c ⎪
⎪ ⎪
⎭
ẅ t 0 − mk2D
w
0 0 ẇ t ∞ l
2m2D
The artificial neural network is coupled with a structural solver,
Equations (1) and (2). The solution of the equations is integrated (10)
from the initial conditions using a four-step predictor-corrector al-
gorithm. The ROM algorithm is illustrated in Fig. 15. 4.3.2. 4th order Runge-Kutta
The model starts setting the initial conditions. A fourth or- In the initial time steps of the calculation, the aeroelastic sys-
der Runge-Kutta scheme is applied for the initialization due to tem is linearly solved using Equations (11) and (12). In these equa-
its accurate results [58]. Finally, a four-step predictor-corrector is tions, A is the coefficient matrix and b  the independent terms

8
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero et al. Aerospace Science and Technology 115 (2021) 106845

Fig. 13. Training and validation of the LSTM neural network.

Fig. 14. Regression plots for the moment coefficient.

vector. The vectors xt are the solution (pitch and plunge) of the 4.3.3. 4 steps predictor-corrector
problem. Each time step, the aeroelastic system is linearly solved using a
  4 steps predictor-corrector method. This solver calculates the solu-
f 1 = Axt + b f 2 = A xt + 1 f 1 t + b tion in a certain time step from the solution in the previous 4 time
2
    (11) steps. Thus, it requires three steps of initialization using another
f 3 = A xt + 1 f 2 t + b f 4 = A xt + f 3 t + b
2 algorithm, as 4th order Runge-Kutta. Using the same nomenclature
1   than in 4.3.2, the numerical solver is presented in equations (13),
xt +1 = xt + f 1 + 2 f 2 + 2 f 3 + f 4 t (12) (14) and (15).
6
9
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero et al. Aerospace Science and Technology 115 (2021) 106845

Fig. 15. Aeroelastic Reduced Order Model flow diagram.

Fig. 16. Stationary aerodynamic coefficients as a function of the angle of attack and the plunge of the plate.

f = Axt + b (13) the stationary coefficient interpolation curves and the forced oscil-
    
lation results are shown.
x pred = xt + 24 55 f xt , A, b − 59 f xt −1 , A, b
t
    (14) On one hand, in Fig. 16, the stationary coefficients are shown as
+37 f xt −2 , A, b) − 9 f xt −3 , A, b a function of the angle of attack and the plunge. The coefficients
     exhibit a nonlinear behavior for angles of attack higher than 4 deg.
xt +1 = xt + 720
1  + 646 f xt , A, b
251 f x pred , A, b In addition, the moment is significantly affected by the vertical
     
−264 f xt −1 , A, b + 106 f xt −2 , A, b − 19 f xt −3 , A, b position of the plate.
On the other hand, Fig. 17 shows the cyclic behavior of the
(15) aerodynamic coefficients due to the forced oscillation motion.
Aerodynamic loops vary as a function of the initial angle of attack
5. Results
of the plate, the amplitude and the frequency of the movement.
For low angles of attack, in the linear aerodynamic range (Figs. 17a
5.1. CFD
and 17b), coefficient clα decreases with the frequency. In addition,
The oscillating motion of the plate produces cycles which mod- the second derivative becomes higher, increasing the curvature of
ify the value of the linear aerodynamic derivatives (Equation (16)). the loop. For post-stall conditions (Figs. 17c and 17d) the behav-
As mentioned in Section 3, the aerodynamic coefficients might be ior of the system is highly nonlinear. In these cases, the effects
expressed as the sum of the stationary and the dynamic terms. of nonlinearities tend to increase with frequency. Thus, the coeffi-
cient presents an oscillating cycle with the frequency of the vortex
c α̇ c 2 α̈
cl = cl0 + clα α + clα̇ + clα̈ 2
+ ... (16) shedding. In addition, for high initial incidences of the plate, an in-
V∞ V∞ crease of the lift slope is noticed as an effect of the nonlinearities.
As the section aerodynamics is nonlinear, CFD has been used The coefficient clα grows for the highest angles of attack becoming
for the calculation of the aerodynamic coefficients. In this section, higher than the potential lift slope of 2π , as shown in Fig. 17c.

10
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero et al. Aerospace Science and Technology 115 (2021) 106845

Fig. 17. Aerodynamic cycles for a characteristic point of the dataset.

5.2. Aeroelastic ROM In addition, in Fig. 18 the instability zone may be identified
as the characteristic stiffness in which there is an abrupt change
in slope. The instability zone is represented by a shaded area. In-
The ROM is then tested on the equivalent rigid cross section side this area the motion of the plate is amplified with time. The
of Fig. 1a. The nondimensional stiffness, k∗ , is varied in order to prediction of LCO and instabilities are also accurate with the use
represent different working conditions: stability range, limit cycle artificial neural networks. To evaluate the LCO prediction and the
oscillation (LCO) and postflutter. limitations of the ROM transient results, the amplitude and fre-
The ROM shows good accuracy to predict mean values of the quency of the predicted motion are analyzed.
deformation, as can be observed in Fig. 18. As stated in the previ- Fig. 19 shows the amplitude of the simulated motion. Both FNN
ous sections, the utilization of artificial neural networks improves and LSTM show a fair similitude with the CFD simulation, while
in a significant way the prediction of oscillatory nonlinear aerody- quasi-steady aerodynamics tend to overpredict the value of defor-
namics with respect to the steady polar, as it captures the dynamic mation for higher values of k∗ . The stall flutter may be identified
effects on the flow which can damp the oscillation or prevent stall. as the point where amplitude abruptly increases. The instability

11
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero et al. Aerospace Science and Technology 115 (2021) 106845

Fig. 18. Comparison of the mean deformation of the ROM and CFD simulation for the different aerodynamic models.

Fig. 19. Comparison of the amplitude of the motion of the ROM and CFD simulation using different aerodynamic models.

Fig. 20. Time cycles of torsion and plunge for a nondimesional stiffness of k∗ = 8.5.

appears when the nondimensional stiffness decreases to the flut- intermediate situation. The motion starts to increase the ampli-
ter value k∗f , in other words, when the free velocity reaches the tude in the same conditions than the FNN. Nevertheless, it is con-
flutter velocity V f . strained to low values, near to the damped aerodynamics, similarly
However, although both typologies of neural networks show to the LSTM (Fig. 20). In addition, when the stall flutter is reached,
similar behaviors, slightly different results might be observed in the CFD simulations increase the amplitude faster than the ANNs,
the amplitude of the motion. FNN is more sensitive to the changes Fig. 21.
of stiffness. Therefore, before the stall flutter, the amplitude starts Figs. 20 and 21 show the time evolution of the previous model
to increase. However, LSTM increases the aerodynamic damping. for two different working conditions. In these figures, small differ-
The motion is totally damped and only grows when it reaches ences in the frequency of the oscillation may be observed. The fre-
the stall flutter condition. CFD aeroelastic simulation presents an quency of the motion is also analyzed in Fig. 22, observing an error

12
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero et al. Aerospace Science and Technology 115 (2021) 106845

Fig. 21. Time cycles of torsion and plunge for a nondimesional stiffness of k∗ = 6.5.

Fig. 22. Comparison of the frequency of the motion of the ROM and CFD simulation for the different aerodynamic models.

lower than 5% for the twist frequency. Differences in frequency are are prepared to be capable of using wind tunnel data. The ANN
slightly more noticeable when analyzing plunge non-dimensional have been trained to reduce both training and validation errors and
frequency, where they are of approximately 10% for nondimen- have been proved to be independent of the number of neurons.
sional stiffness far from the aeroelastic instability. However, as it Two different typologies of network (FNN and LSTM), which are
was observed at Figs. 18 and 19, the influence of this frequency is utilized for similar problems in the literature, are tested to com-
of second order when estimating nonlinear aeroelastic instabilities. pare the deformation obtained and the accuracy of the ROM.
Finally, it is important to compare computational cost of the The networks have been tested for different conditions, show-
different models performed during this work. Note that all the ing high accuracy to reproduce the aerodynamic coefficients of the
simulations were performed on an Intel® Xeon® CPU ES-2630 v2, cross section given its motion as input. Although there is a gain
using 5 parallel processes. As an example, the average time for of accuracy comparing with the nonlinear steady polar, some ef-
computing each simulation was of 10 hours for the CFD compu- fects are lost with the use of the surrogate model. For instance,
tations, and between 3 and 5 minutes for FNN and LSTM, respec- the variations of the aerodynamic coefficients due to the fluctua-
tively. tions of the wake are not reproduced by the model. In addition, the
ANNs tend to reproduce a sinusoidal oscillation of the coefficients,
6. Conclusions
losing other effects as vortex shedding, which limits the range of
application of the artificial neural networks. Thus, the procedure
In this article, a Reduce Order Model (ROM), with a surrogate
works for nonlinear aerodynamics, in a certain range of application
aerodynamic model based on Artificial Neural Networks (ANN), is
in which the oscillations due to turbulence are small or negligible.
developed for calculating an aeroelastic equivalent cross section.
The main objective pursued is to test the accuracy of the ROM in For an oscillating plate, it may be applied for low angles of attack
the prediction of dynamic instabilities, as stall flutter, in nonlinear (approximately between −6 and 6 deg), low amplitudes of the os-
aerodynamic problems. cillation (between 0 and 5 deg) and low or moderate frequencies
fc
A flat plate is chosen as the studied geometry due to its nonlin- (approximately a maximum of V = 0.3). For higher frequencies
∞
ear behavior. The aerodynamic coefficients of the plate are proved the vortex shedding generated becomes more important and can
to vary with respect to the initial angle of attack, the amplitude reduce in a significant way the accuracy of the network.
and the frequency of the motion. In order to reproduce the differ- The validation of the ANN shows that there is not a significant
ent aerodynamic conditions that may appear around the plate, a difference between the feed-forward neural network and the long
set of CFD simulations have been calculated as the training dataset short term memory network. The previous fact, added to the eas-
of the network. Nevertheless, the methodology and the test case ier definition and training of the feed forward neurons, make this

13
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero et al. Aerospace Science and Technology 115 (2021) 106845

typology more efficient to obtain the aerodynamic coefficients of a [7] L. Librescu, S. Thangjitham, Analytical studies on static aeroelastic behavior of
plate under low or moderate angles of attack. forward-swept composite wing structures, J. Aircr. 28 (1991) 151–157.
[8] J.P. Thomas, E.H. Dowell, K.C. Hall, Nonlinear inviscid aerodynamic effects on
Finally the ANNs have been coupled to a simplified structural
transonic divergence, flutter, and limit-cycle oscillations, AIAA J. 40 (2002)
solver. The results show an accurate prediction of the mean defor- 638–646.
mation in both torsion and plunge. The instability is also predicted [9] L. Sanches, T.A. Guimaraes, F.D. Marques, Aeroelastic tailoring of nonlinear typ-
accurately from the mean deformation curve as the point in which ical section using the method of multiple scales to predict post-flutter stable
lcos, Aerosp. Sci. Technol. 90 (2019) 157–168.
the slope of the curve increases abruptly. The time solutions have
[10] A.S. Thelen, L.T. Leifsson, P.S. Beran, Multifidelity flutter prediction using local
also been compared observing differences of behavior between the corrections to the generalized aic, Aerosp. Sci. Technol. 106 (2020) 106032.
types of aerodynamic models. Quasi-steady aerodynamics shows [11] T. Zhou, E. Dowell, S.S. Feng, Computational investigation of wind tunnel wall
a lack of accuracy when predicting a dynamic phenomenon. Both effects on buffeting flow and lock-in for an airfoil at high angle of attack,
ANNs present similar results. However, the LSTM solution tends Aerosp. Sci. Technol. 95 (2019) 105492.
[12] R. Camussi, A. Di Marco, C. Stoica, M. Bernardini, F. Stella, F. De Gregorio, D.
to damp the oscillation more than the CFD while FNN increases Barbagallo, Wind tunnel measurements of the surface pressure fluctuations on
the amplitude and predicts the instability for a slightly lower free the new vega-c space launcher, Aerosp. Sci. Technol. 99 (2020) 105772.
stream velocities. [13] M. Nikbay, P. Acar, Integrating analytical aeroelastic instability analysis into
As a summary, the previous facts lead to the following main design optimization of aircraft wing structures, J. Appl. Eng. Math. 1 (2011)
237–253.
conclusions:
[14] F. Roizner, M. Karpel, Parametric flutter margin method for aeroservoelastic sta-
bility analysis, AIAA J. 56 (3) (2018) 1011–1022.
• The aerodynamic loads in a dynamic problem are strongly con- [15] J. Sodja, F. Roizner, R. De Breuker, M. Karpel, Experimental characterisation of
ditioned by the mean angle of attack, the amplitude and the flutter and divergence of 2d wing section with stabilised response, Aerosp. Sci.
Technol. 78 (2018) 542–552.
frequency of the oscillation. Thus, the stationary polar cannot
[16] S. Evans, D. Bradney, P. Clausen, Development and experimental verification of
be used to obtain accurate solutions. a 5 kw small wind turbine aeroelastic model, J. Wind Eng. Ind. Aerodyn. 181
• The utilization of artificial neural networks (ANN) may reduce (2018) 104–111.
both the time of simulation (compared with CFD simulations) [17] T. Argentini, D. Rocchi, C. Somaschini, Effect of the low-frequency turbulence
and the error (compared with linear or steady models) in on the aeroelastic response of a long-span bridge in wind tunnel, J. Wind Eng.
Ind. Aerodyn. 197 (2020) 104072.
aeroelastic nonlinear problems. [18] K. Fujii, Progress and future prospects of cfd in aerospace - wind tunnel and
• Different typologies of neural networks can be applied with beyond, Prog. Aerosp. Sci. 41 (2005) 455–470.
similar results. However, recursive structures, for instance [19] X. Dong, Y. Zhang, Y. Zhang, Z. Zhang, X. Lu, Numerical simulations of flutter
LSTM networks with the previous dataset and activation func- mechanism for high-speed wide-chord transonic fan, Aerosp. Sci. Technol. 105
(2020) 106009.
tions tend to damp the oscillation. On the other hand, FNN [20] J. Ren, H. Huang, D. Wang, X. Dong, B. Cao, An efficient coupled-mode flutter
are faster and reduce the calculation cost, but they can be too analysis method for turbomachinery, Aerosp. Sci. Technol. 106 (2020) 106215.
conservative in the prediction of the aeroelastic instability. [21] L. Wang, R. Quant, A. Kolios, Fluid structure interaction modelling of
horizontal-axis wind turbine blades based on cfd and fea, J. Wind Eng. Ind.
Aerodyn. 158 (2016) 11–25.
Finally, this work completes the research developed in the lit-
[22] C. Peng, H. Jinglong, Prediction of flutter characteristics for a transport wing
erature, showing the applicability of the ANN to the aeroelastic with wingtip device, Aerosp. Sci. Technol. 23 (2012) 461–468.
problem, comparing different types of ANN and coupling the aero- [23] N. Tsushima, W. Su, Flutter suppression for highly flexible wings using passive
dynamics with the structure of the section. This procedure has and active piezoelectric effects, Aerosp. Sci. Technol. 65 (2017) 78–89.
[24] M. Amooozgar, S. Fazelzadh, H.H. Khodaparast, M. Friswell, J. Cooper, Aeroe-
been demonstrated to reduce in a significant way the computa-
lastic stability analysis of aircraft wings with initial curvature, Aerosp. Sci.
tional cost and to preserve the accuracy of the solution for arbi- Technol. 107 (2020) 106241.
trary sections and nonlinear aerodynamics. [25] H.J. Kwon, D.H. Kim, I. Lee, Frequency and time domain flutter computations
of a wing with oscillating control surface including shock interference effects,
Aerosp. Sci. Technol. 8 (2004) 519–532.
Declaration of competing interest
[26] Y. You, D. Na, S.N. Jung, Improved rotor aeromechanics predictions using a fluid
structure interaction approach, Aerosp. Sci. Technol. 73 (2018) 118–128.
The authors declare that they have no known competing finan- [27] C. Baxevanou, P. Chaviaropoulos, S. Voutsinas, N. Vlachos, Evaluation study of a
cial interests or personal relationships that could have appeared to Navier-Stokes cfd aeroelastic model of wind turbine airfoils in classical flutter,
J. Wind Eng. Ind. Aerodyn. 96 (2008) 1425–1443.
influence the work reported in this paper.
[28] E. Dowell, A Modern Course in Aeroelasticity, Springer, 2015.
[29] E. Camilo, F.D. Marques, J.L.F. Azevedo, Hopf bifurcation analysis of typical sec-
Acknowledgements tions with structural nonlinearities in transonic flow, Aerosp. Sci. Technol. 30
(2013) 163–174.
[30] D. Li, A.D. Ronch, G. chen, Y. Li, Aeroelastic global structural optimization us-
This project has been partially funded by Spanish Ministry of
ing an efficient cfd-based reduced order model, Aerosp. Sci. Technol. 94 (2019)
Science, Innovation and Universities through the University Faculty 105354.
Training (FPU) program with reference FPU19/02201. [31] J. Luo, Design optimization of the last stage of a 4.5-stage compressor using a
pod-based hybrid model, Aerosp. Sci. Technol. 76 (2018) 303–314.
[32] G. Chen, D. Li, Q. Zhou, A.D. Ronch, Y. Li, Efficient aeroelastic reduced or-
References
der model with global structural modifications, Aerosp. Sci. Technol. 76 (2018)
1–13.
[1] G. Rogers, J. Kristof, Reducing operational and product costs through environ- [33] K. Ghorbanian, M. Gholamrezaei, An artificial neural network approach to com-
mental accounting, Environ. Qual. Manag. 12 (2003) 17–42. pressor performance prediction, Appl. Energy 86 (2009) 1210–1221.
[2] S. Guo, Aeroelastic optimization of an aerobatic aircraft wing structure, Aerosp. [34] J.M. Luján, H. Climent, L.M. García-Cuevas, A. Moratal, Volumetric efficiency
Sci. Technol. 11 (2007) 396–404. modelling of internal combustion engines based on a novel adaptive learning
[3] C. Ledermann, P. Ermanni, R. Kelm, Dynamic cad objects for structural op- algorithm of artificial neural networks, Appl. Therm. Eng. 123 (2017) 625–634.
timization in preliminary aircraft design, Aerosp. Sci. Technol. 10 (2006) [35] J.M. Alonso, F. Alvarruiz, J.M. Desantes, L. Hernández, V. Hernández, G. Moltó,
601–610. Combining neural networks and genetic algorithms to predict and reduce
[4] B. Briseghella, L. Fenu, C. Lan, E. Mazzarolo, T. Zordan, Application of topologi- diesel engine emissions, IEEE Trans. Evol. Comput. 11 (2007).
cal optimization to bridge design, J. Bridge Eng. 18 (2013) 790–800. [36] J. Peng, C. Luo, Z. Han, Z. Hu, G. Han, Z. Jiang, Parameter-correlation study on
[5] D.K. Koli, G. Agnihotri, R. Purohit, Advanced aluminium matrix composites: the shock-shock interaction using a machine learning method, Aerosp. Sci. Technol.
critical need of automotive and aerospace engineering fields, Mater. Today 2 107 (2020) 106247.
(2015) 3032–3041. [37] T. Wu, A. Kareem, Modeling hysteretic nonlinear behavior of bridge aerody-
[6] M. Blair, T.A. Weisshaar, Swept composite wing aeroelastic divergence experi- namics via cellular automata nested neural network, J. Wind Eng. Ind. Aerodyn.
ments, J. Aircr. 19 (1982) 1019–1024. 99 (2011) 378–388.

14
A.J. Torregrosa, L.M. García-Cuevas, P. Quintero et al. Aerospace Science and Technology 115 (2021) 106845

[38] C.-H. Chen, J.-C. Wu, J.-H. Chen, Prediction of flutter derivatives by artificial [49] P.M. Quintero, Characterization of fluid structure interaction mechanisms and
neural networks, J. Wind Eng. Ind. Aerodyn. 96 (2008) 1926–1937. its application to vibroacoustic phenomena, Ph.D. thesis, Universitat Politècnica
[39] T. Abbas, I. Kavrakov, G. Morgenthal, T. Lahmer, Prediction of aeroelastic re- de València, 2019.
sponse of bridge decks using artificial neural networks, Comput. Struct. 231 [50] J. Weiss, J. Maruszewski, W. Smith, Implicit solution of preconditioned Navier-
(2020) 106198. Stokes equations using algebraic multigrid, AIAA J. 37 (1999) 29–36.
[40] T. Li, T. Wu, Z. Liu, Nonlinear unsteady bridge aerodynamics: reduced-order [51] J. Weiss, W. Smith, Preconditioning applied to variable and constant density
modeling based on deep lstm networks, J. Wind Eng. Ind. Aerodyn. 198 (2020) flows, AIAA J. 33 (1995) 2050–2057.
104116. [52] V. Venkatakrishnan, On the convergence of limiters and convergence to steady
[41] A. Gil, A. Tiseira, P. Quintero, A. Cremades, Prediction of the non-linear aeroe- state solutions, AIAA-93-0880, 1995.
lastic behavior of a cantilever flat plate and equivalent 2d model, Aerosp. Sci. [53] H. Hadzic, Development and application of finite volume method for the com-
Technol. 113 (2021) 106685. putation of flows around moving bodies on unstructure, overlapping grids,
[42] A. Torregrosa, A. Gil, P. Quintero, A. Ammirati, H. Denayer, W. Desmet, Predic- 2005.
tion of flow induced vibration of a flat plate located after a bluff wall mounted [54] P. Roache, Verification and Validation in Computational Science and Engineer-
obstacle, J. Wind Eng. Ind. Aerodyn. 190 (2019) 23–39. ing, Hermosa Publishers, 1998.
[43] B. Kröse, P. van der Smagt, An Introduction to Neural Networks, University of [55] A.A. Adeniyi, H.P. Morvan, K.A. Simmons, A transient cfd simulation of the flow
Amsterdam, 1996. in a test rig of an aeroengine bearing chamber, Heat Transf. 5C (2014).
[44] S. Pope, Turbulent Flows, Cambridge University Press, 2009. [56] I.V. Tetko, D.J. Livingstone, A.I. Luik, Neural network studies 1. Comparison of
[45] R. Gavasane, P. Pai, V. Kumar, Numerical simulation of pitching and plunging overfitting and overtraining, J. Chem. Inf. Comput. Sci. 35 (1995) 826–833.
motion of flat plate using oversetmeshes, in: Symposium on Applied Aerody- [57] S. Yadav, S. Shukla, Analysis of k-fold cross-validation over hold-out validation
namics and Design of Aerospace Vehicle (SAROD 2013), 2013. on colossal datasets for quality classification, in: 6th International Advanced
[46] J. Morgado, R. Vizinho, M. Silvestre, J. Páscoa, Xfoil vs cfd performance predic- Computing Conference, 2016.
tions for high lift low Reynolds number airfoils, Aerosp. Sci. Technol. 75 (2016) [58] W. Zhang, Y. Jiang, Z. Ye, Two better loosely coupled solution algorithms of cfd
207–214. based aeroelastic simulation, Eng. Appl. Comput. Fluid Mech. 1 (2007).
[47] D. Wilcox, Multiscale model for turbulent flows, in: Proceedings of the 24th
AIAA Aerospace Science Meeting, Vol. 24, 1986, pp. 1311–1320.
[48] M.J. Emes, F. Ghanadi, M. Arjomandi, R.M. Kelso, Investigation of peak wind
loads on tandem heliostats in stow position, Renew. Energy 121 (2018)
548–558.

15