Journal of Computational Science 62 (2022) 101698

Contents lists available at ScienceDirect

Journal of Computational Science

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

Extended computational capabilities for high-fidelity fluid–structure

simulations
Nicola Fonzi ∗, Vittorio Cavalieri, Alessandro De Gaspari, Sergio Ricci
Department of Aerospace Science and Technology, Politecnico di Milano, Via La Masa 34, 20156 Milano, Italy

ARTICLE INFO ABSTRACT

Keywords: Current research efforts in aeroelasticity aim at including higher fidelity aerodynamic results into the
Aeroelasticity multi-disciplinary simulation environments. In the present effort, an updated Python-based Fluid–Structure
Python Interaction framework has been included in SU2 code to allow for efficient and fully open-source simulations
Computational fluid dynamics
of detailed aeroelastic phenomena. The method provides a powerful and easily approachable environment.
Fluid–structure interaction
The developed software has been assessed against three test cases of increasing complexity, coming from
Flutter
the aeronautic community. These applications showed how the framework can capture strongly nonlinear
aerodynamic effects, such as shock waves, and their interaction with the structural dynamics of the model.

1. Introduction methods cease to be valid [12,13]. It is interesting to note that newly

presented concept planes, all have peculiar characteristics that will
Aeroelasticity has always been an important topic in aerospace hinder the application of classical methods for their analysis [14,15].
engineering. Since the dawn of aviation, aeroelasticity has affected the In all these cases, it is required to have a more complete description
design of aircraft, as well as developments to avoid flutter [1,2]. Recent of aerodynamics. The most general model that can be used is pro-
advances in multi-disciplinary optimisation enable to take into account vided by a Computational Fluid Dynamic (CFD) model. Modern efforts
aeroelastic effects in the design process [3]. This would allow to reduce in aeroelasticity are directed towards a better, and more efficient,
the need for later redesign and to take advantage of aeroelasticity use of this tool for flutter prediction, and aeroelastic phenomena in
to design more efficient aircraft [4]. For example, strategies can be general [16,17]. This can be clearly seen as the latest aeroelastic
implemented to design longer endurance aircraft and lighter structures, prediction workshops focused on the use of CFD in this field of re-
or to install active aeroelastic control systems on-board [5–7]. search [18]. In these workshops, the focus was on transonic application,
However, the common practice to address all these aspects mostly thus conditions where nonlinearities in the flow are present due to the
relies on the use of potential aerodynamic methods, namely Doublet appearance of shock waves. However, nonlinearities can come from
Lattice Method (DLM) and Vortex Lattice Method (VLM). Corrections
an infinite range of sources. There is an excellent review of mod-
for close to transonic effects must be provided to these methods if that
ern nonlinear aeroelasticity in Dowell and Tang [19]. Some possible
regime is of interest, and this requires more accurate computations [8].
sources have been already mentioned above, others, less obvious, may
Further, future trends towards highly flexible wings, deep-transonic
include concentrated nonlinearities in the structure. Indeed, concen-
speeds, and less conventional lifting surface designs, will hinder the
trated nonlinearities in control surface attachments are often of interest,
application of potential methods. Indeed, Aerodynamic Influence Coef-
and they are usually studied using a quasi-linearisation, obtained via
ficients (AIC) Matrices are usually computed in a certain configuration
a describing function approach. Usual linear aerodynamic methods are
and small perturbations are assumed for aeroelastic analysis. This
would not hold true for the large displacements exhibited by new then applied to the resulting model to obtain stability boundaries. How-
designs like the Boeing-787 [9] or High Altitude Long Endurance ever, it has been shown that flutter predictions, obtained via nonlinear
aircraft [10]. Further, deeper in the transonic range, where the well time marching simulations, may differ if CFD or potential methods are
known transonic dip occurs [11], strong nonlinear aerodynamic effects used [20,21]. Thus, it may be of interest to introduce CFD instead of
can only be captured with more complex models. DLM, and assess the stability of these equivalent linear systems.
Finally, where non conventional configurations are utilised, like T- Therefore, there is the need for always better and easier to use
Tails, it is well known that the basic assumptions behind potential methods for high-fidelity aeroelasticity. An efficient procedure able to

∗ Corresponding author.
E-mail address: nicola.fonzi@polimi.it (N. Fonzi).

https://doi.org/10.1016/j.jocs.2022.101698
Received 19 November 2021; Received in revised form 8 April 2022; Accepted 3 May 2022
Available online 13 May 2022
1877-7503/© 2022 Elsevier B.V. All rights reserved.
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

work presented here that led to the development of a new framework

Nomenclature embedding high-fidelity aerodynamics in different types of aeroelastic
𝛼 Geometric angle of attack analyses.
𝜔̄ Uncoupled natural pulsation ratio As of December 2020, the code has been embedded in SU2 package
itself, allowing anyone to exploit these capabilities, compiling only one
𝜒 Normalised static unbalance
software [26]. The Nastran-like format also allows to use another open-
𝐂 Damping matrix
source code, called NeoCASS [27,28], for the automatic generation of
𝐌 Mass matrix
the structural finite element model.
𝐔 Matrix mapping the degrees of freedom to The developed framework is efficient enough for a more routinely
the physical structural nodes use of high-fidelity aeroelasticity. Moreover, it has the great advantage,
𝐕 Matrix of the eigenvectors of the mass compared to other software available, that can be completely based
matrix on open source projects. Other programmes are already available, and
 Pitching moment already proved their effectiveness, but most of them are restricted in
Ma Mach number use. FUN3D, developed by NASA, is a powerful high-fidelity aeroelastic
Re Reynolds number software, based on a fully unstructured formulation, that is however
𝜇𝑚 Mass ratio export restricted and only available to US residents [29]. EZNSS, which
𝜔ℎ Uncoupled plunging natural pulsation is based on a Chimera structured grid formulation, is also an effective
𝜔𝑖 Pulsation of the 𝑖th mode mean of simulating aeroelastic system, but it is restricted to Israeli
Air Force [30]. ZEUS, from ZONA Technology, is a high-fidelity Euler
𝜔𝛼 Uncoupled pitching natural pulsation
solver, which is for commercial use only [31]. The most common
𝜌 Fluid density
open source choice is OpenFOAM [32]. However, due to the very new
𝜃 Pitch angle community related to SU2 development, some helpful features, espe-
𝐹̃𝑖 Generalised force on the 𝑖th mode cially related to mesh deformation and adjoint studies, are continuously
𝜉𝑖 Modal damping being improved and prove to be extremely helpful in the context of
𝑏 Half-chord length aeroelasticity. Adjoint optimisations have been previously performed
𝑐 Chord length using the Python interface [33], with a beam model, and the work
𝐶(𝑘) Theodorsen function presented here may be exploited to ease this process. Thus, the authors
𝐶𝛼 Damping for the pitching mode believe that the present contribution may be of great help for both
𝐶ℎ Damping for the plunging mode researchers and practitioners.
The plan of the paper is as follows. In Section 2 of this paper,
𝐶𝑙 Lift coefficient
the used methods are outlined. First, a general explanation about the
𝐶𝑙 𝛼 Slope of the lift curve
framework is provided. Next, details are given for the fluid solver, the
ℎ Plunge displacement interface between fluid and structure, and the structural solver. Space
𝐼 Inertia will be given to both a general external solid solver and the new native
𝐾𝛼 Stiffness for the pitching mode solver. In Section 3, results obtained with the code are presented. We
𝐾ℎ Stiffness for the plunging mode considered three test cases; a NACA 0012 airfoil, a wing in transonic
𝐿 Lift flow, and a numerical simulation of an entire model in the wind tunnel.
𝑚 Mass The NACA 0012 results are quite common and here reported as a
𝑞𝑖 Generalised modal coordinate reference to verify the good implementations of the methods. Similar
𝑟2𝛼 Normalised inertia results have already been reported, for SU2, in a previous paper [34].
However, in the context of the work in [34], a specialised structural
𝑆 Reference surface
solver, only able to solve for a typical section airfoil, has been used.
𝑆𝑚 Static moment of inertia
Here, we reproduce the same results with the general native structural
𝑈∞ Free-stream velocity
solver, that can later be used for an arbitrary application. Thus, while
the results will be similar in nature, they come from a completely
different source. As far as the transonic wing is concerned, this is the
Benchmark SuperCritical Wing (BSCW). Again, forced oscillations have
couple even complex aerodynamic computations, without being too already been studied for this case. However, in this paper, we also
time-consuming, would help the transition from potential-based meth- present flutter analysis. The full power of the implementation will be
ods. Further, it would be desirable to have an open source framework proved in the last set of results where a complex structural model, fully
so that custom modifications can be easily embedded. flexible (the BSCW only has pitch and plunge degrees of freedom), and
A concern specific to industrial applications, is related to the in- immersed in a strongly nonlinear flow field, will be studied in terms of
clusion of new methods in established workflows. The best practises dynamics and stability. Finally, in Section 4, conclusions will be drawn
usually involve the use of MSC/Nastran as the aeroelastic solver [22], and future work directions outlined.
thus it is common to model the structure using this format.
For these reasons, in the current effort, the authors updated and 2. Framework for aeroelastic computation
extended the Python-based Fluid–Structure Interaction (FSI) frame-
work available in the open-source code SU2 [23]. This framework The overall goal of this research is the development of a highly per-
was originally derived from CUPyDO [24,25], to exploit SU2 for the forming and easily approachable code for high-fidelity aeroelasticity.
aerodynamic solution, and an external solver for the structural solution. For this purpose, the work focused on a Python-based Fluid–Structure
A native structural solver, designed to work on finite element mod- Interaction (FSI) framework, coupled with CFD simulations. This frame-
els (FEM) supplied in MSC/Nastran format, has been developed and work was originally proposed in [25] and it is embedded in the well
introduced in the framework. However, the code has been developed known SU2 code, which is an open–source collection of tools written
to work independently of the finite element model format and the in C++ and Python for the numerical solution of partial differential
FSI interface is as general as possible, allowing for other structural equations. Efforts are directed to the introduction of a new native solver
solvers to be coupled. These two aspects are the key points of the for the integration of structural equations, based on finite element

2
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

Fig. 1. Different code levels of the Python-based FSI framework.

On the other hand, if the structural solution is to be obtained from

the structural solver after an actual integration of the equations of
motion, the aeroelastic system will be studied. In this case, the coupling
between the fluid solver and the structural solver is provided in a
tightly coupled framework.
In particular, for each time step, or just once in case of a steady
computation, the solution process is the following:

1. The initial positions of the structural nodes are obtained and

interpolated via Radial Basis Functions (RBF) onto the aero-
dynamic mesh. In this phase, if required, grid velocities are
computed. In the very first time step, if there is an initial
deformation, fictitious velocities are avoided by imposing zero
grid velocity in all the domain.
2. The fluid solver is run, obtaining the distribution of aerodynamic
forces on the surface at the interface between fluid and structure.
Fig. 2. Available simulation settings for the current framework.
SU2 is a vertex centred finite volume solver, thus the solution is
already obtained at the aerodynamic nodes.
3. If the structural solution is prescribed, the displacement field is
models, coming from Nastran, while support for different external imposed.
structural solvers is allowed. 4. Else:
The current state of the framework is presented in 1. The Simplified
(i) The forces are interpolated via RBF onto the structural
Wrapper and Interface Generator (SWIG) [35] is used to obtain a
nodes.
Python Application Programming Interface (API) to the lower level
(ii) The structural solver is run to obtain the new positions.
functions. Obviously, not all the lower level functions have a counter-
part in Python, but the main ones, required to drive the simulation, are 5. Convergence is checked. This is done computing the root mean
mapped. The interface has been standardised so that, when adding a square of the incremental structural displacements, between the
new structural solver, no modifications are required. If all the required previous FSI iteration and the current one.
routines to extract nodes’ position and set the loads are present, regard- 6. If convergence is reached, the structural displacements for the
less of the process underlying in the lower level functions, the coupled next time step are predicted.
simulation will run smoothly. 7. Else, the structural solution is relaxed with an Aitken relaxation,
The solution process has also been generalised and standardised so and the steps of this list are repeated.
that four kinds of simulations are now allowed, as depicted in Fig. 2.
The process outlined above can also be graphically seen in Fig. 3.
In the case of prescribed structural motion, the aerodynamic re-
If the structural solver used does not allow for one of the outlined
sponse to specific inputs is computed. Computing the aerodynamic
solution processes, the interface will still work with the available
responses is really useful if, for example, the aerodynamic system must
options. This is useful as it increases the flexibility of the code.
be identified [36]. These responses can be either computed in a steady In the next subsections, more details will be provided about the
or unsteady simulation. In the former case, we may want, for example, different components of the framework.
to understand how, with a certain displacement field, the aerodynamics
changes around our body. Using the present code it will not be required 2.1. Aerodynamic solver - SU2
to prepare again the mesh, as the framework will automatically deform
it around the new structural interface. An ongoing research inside the The open source code SU2 was originally developed at Stanford
Department of Aerospace Science and Technology at Politecnico di Mi- University [23]. The solution method uses a finite volume formulation,
lano is, for example, exploiting this feature for the design optimisation vertex centred, with a dual mesh grid. This formulation may offer
of morphing wing devices [37,38] and their final verification through advantages in terms of accuracy and stencil size, depending on the
nonlinear structural analyses coupled to CFD simulations [39]. problem at hand [40]. The entire code has been developed from the

3
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

Fig. 3. Solution process flow chart.

very beginning with attention to maximum flexibility, allowing for easy mesh deformation step. Grid velocities for the Arbitrary Lagrangian–
inclusion of new routines and thus making it the perfect platform for Eulerian (ALE) method are then computed using finite differences in
multiphysics problems [34,41]. time.
Further, the C++ code of SU2 offers the possibility to be wrapped in a
Python API, so that high level functions can be called with ease. Indeed, 2.2. Aero-structure interface method
the fluid and structural solvers can be called as a normal Python mod-
ule, giving the opportunity to easily create scripts for multi-disciplinary The interpolation between fluid and structure is performed via a
design optimisations (MDO) or complex coupled simulations. RBF interpolation [44].
Several numerical schemes are available. In the context of this The displacements at the structure side are obtained starting from
work, the Jameson–Schmidt–Turkel scheme has been utilised for the the positions of the undeformed structural nodes and the obtained fluid
mesh displacements are also defined starting from the initial mesh.
convective part of the equations [42,43]. The choice is due to the great
The RBF method involves the choice of a kernel function and a poly-
range of problems that this method can tackle, including transonic
nomial order for the interpolation. In this case, the interpolation poly-
problems. Green–Gauss reconstruction schemes are used for the viscous
nomial is linear, as this recovers correctly rigid body translations [45],
part.
and the kernel function is the CP C2 described in [44].
Both steady and unsteady simulations will be presented. In all
The interface is completely parallelised, in order to maximise per-
cases, the pseudo time has been used to converge the solution with
formance. The communication between processors exploits the Message
a local Courant–Friedrichs–Lewy (CFL) number that will be specified
Passing Interface (MPI) for Python [46] as main library. Arrays built
in the relevant section of the article. At each pseudo time iteration the using the high-level mathematical library, named NumPy, are used
linear solver provides the change in solution. Again, different linear in combination with the MPI library to provide optimal speed. The
solvers are available and the chosen one will be specified on a case linear problems coming from the interpolation are solved using an
by case basis. If unsteady simulations are performed, finite difference interface to a specific library of data structures and routines for the par-
approximations, second order in time, are used for the time derivative allel solution of scientific applications modelled by partial differential
term. equations [47].
In the context of FSI problems, the positions of the grid nodes can be
communicated to the CFD solver via the Python wrapped functions. The 2.3. Structural solver
fluid solver wrapping has been modified so that the FEM solver internal
to SU2 itself is used as a mesh solver. This allows to use all the internal In the previous sections it was explained how the interface is
solver capabilities to improve the efficiency and effectiveness of the standardised to accept any generic structural solver, provided that it

4
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

has the required wrapped functions. As of today, efforts are directed in

the creation of a module for the commercial code Abaqus, and others
are planned for the near future.
The native structural solver offers an efficient and straightforward
means to obtain an aeroelastic solution using SU2 as aerodynamic
solver. The standard input for the structural solver consists of two
sets of information. First, the set of grid nodes’ locations, which are
required for the interpolation between fluid and structure. Second, the
modal shapes normalised for unit mass and the modal frequencies. The
structural solver will use the latter to build the system of equations to
be solved, and the former to define the displacements, velocities, and
accelerations of the physical structural nodes, as a function of the modal
solution.
It must be noted that the structural solver is usually used with
the above preparation scheme, but not limited to this. Indeed, a non-
diagonal system of modal equations can be provided, which are, for
example, obtained with the use of fictitious masses [48,49] or adding
static modes. Further, scalings different from the unit mass can also be
used.
Finally, beside the solution of the dynamics of a set of modes,
an arbitrary system of equations can also be integrated, if properly
formulated. So, for example, we may directly solve for all the degrees
of freedom of a structural system, using the identity matrix instead of
the modal shape matrix. Overall, the native structural solver was born
as a fast mean to couple structural modal models with SU2, but can be
used with much more flexibility to integrate any system of structural
equations.
In case the native solver is used, the input–output relations of the
framework are as represented in Fig. 4. In the figure, thin lines are
used to identify inputs or outputs, while heavier lines are used to
Fig. 4. Framework layout in case the native solver is used.
mark the modules. The 𝐔 matrix contains the relation between the
amplitude of the degrees of freedom and the displacements of all the
structural nodes, which may be the matrix of mode shapes, in case
modal coordinates are used. At each time step, the modal equations are advanced in time with
The equations, in case of decoupled modal equations normalised for a generalised-𝛼 algorithm [50]. This avoids the modification of the
unit mass, are stated as: system into a set of first order Ordinary Differential Equations (ODEs)
⎧ and can also provide excellent numerical properties, like unconditional
𝑞̈1 + 𝜔21 𝑞 = 𝐹̃1
⎪ stability and no numerical damping of dynamics with frequencies
⎪ 𝑞̈2 + 𝜔22 𝑞 = 𝐹̃2 below half of the sampling frequency.
⎨
⎪ … In case of a steady solution, the inertia forces are anyway taken
⎪ 𝑞̈𝑛 + 𝜔2𝑛 𝑞 = 𝐹̃𝑛 into account. In this way, also in the static computation the Newmark
⎩
method is used, instead of a normal linear solver, adding numerical
Where 𝑞𝑖 is the generalised modal coordinate, 𝜔𝑖 is the related pul-
sation, 𝐹̃𝑖 its generalised force, and 𝑛 is the total number of modes stability to the coupled simulation.
considered. The generalised force, in turn, is obtained using the mode
shapes matrix 𝐔 as: 𝐹̃𝑖 = 𝑈𝑖𝑇 𝐅, with 𝐅 the vector of structural nodal 3. Validation test cases
forces.
The structural solver offers the opportunity to include a propor- Three test cases are here presented, with the purpose of validating
tional damping. This is specified as a percentage of the critical damping the framework for different aerodynamic conditions and applying it
and adds to each equation a term 2𝜉𝑖 𝜔𝑖 𝑞̇ 𝑖 . Where 𝜉𝑖 is the percentage of
to structural models of increasing complexity. All the additions and
structural modal damping.
modifications will be tested, thus the native solver will be used with
In case of non-decoupled equations, the introduction of damping
all the possible simulation types presented in 2.
is not so straightforward. In typical applications, the same modal
damping is prescribed in a range of frequencies. Thus, it is assumed
that a certain 𝜉 is provided to the solver and this must be applied 3.1. NACA 0012 airfoil
to all the modes resulting from the eigenanalysis of the nondiagonal
matrices. The modal damping is thus applied in the following way. The first test case here introduced concerns a NACA 0012 airfoil.
First, an eigenanalysis is performed obtaining the eigenvalues 𝑑𝑖 and The ultimate goal of this section is to demonstrate the capability of
eigenvectors 𝐕𝑖 . The mass matrix is then diagonalised as: the proposed FSI tool in solving a typical basic aeroelastic problem,
̃ = 𝐕𝑇 𝐌𝐕
𝐌 namely the flutter of a two-dimensional pitching–plunging airfoil. This
application is handled step by step, following the schematic presented
The diagonal damping matrix elements can then be obtained as: in Fig. 2. In this way, the features of the code and the available solution
𝐶̃𝑖𝑖 = 2𝜉𝑖 𝑑𝑖 𝑀
̃ 𝑖𝑖 methods are gradually shown and validated.
The reference model is a pitching–plunging NACA 0012 airfoil in
Finally, the obtained matrix is brought back to the initial mathematical
a free-stream flow. The structural model is made by a single point,
space as:
positioned at the rotation axis, with two degrees of freedom, pitch and
𝐂 = 𝐕−𝑇 𝐂𝐕
̃ −1 plunge. Inertia and mass of the airfoil are concentrated at the centre

5
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

Fig. 5. FEM mesh for the NACA 0012 airfoil.

of mass of the airfoil, at a certain distance from the rotation axis. The 3.1.1. Imposed deformation aerodynamics
equations of motion are available analytically and read: At the beginning, a simple steady aerodynamic analysis is examined
{ to validate the capability of the interface to impose mesh deformation.
𝑚ℎ̈ + 𝑆𝑚 𝜃̈ + 𝐶ℎ ℎ̇ + 𝐾ℎ ℎ = −𝐿 The analysis at an angle of attack of 3 degrees is repeated twice. First,
𝑆𝑚 ℎ̈ + 𝐼 𝜃̈ + 𝐶𝛼 𝜃̇ + 𝐾𝛼 (𝛼 + 𝜃) =  an angle of attack of 3 degrees is directly set in SU2 in the free-
stream definition. Later, an angle of attack of 0 degrees is assumed,
Where 𝑚 is the mass of the airfoil, 𝐼 the inertia around the rotation
but a pitch rotation of 3 degrees is imposed through the fluid–structure
axis, 𝑆𝑚 the static moment of inertia at the rotation axis, 𝐶 and 𝐾 interface. The aerodynamic conditions for this case are representative
the damping and stiffness respectively. 𝐿 and  are the lift and the of a high Reynolds (Re) number and low Mach number flow. The Re
pitching up moment. The degrees of freedom are 𝜃, pitch, and ℎ, number is equal to 6 millions, while the Mach number is equal to
plunge. On the other hand, 𝛼 is the geometric angle of attack. 0.1. The temperature is set to 293.15 K. Given these conditions, the
However, as we want to develop a general framework, we will not flow will not behave much differently from what can be predicted via
use these equations, but we will rather build a really simple FEM model common linear incompressible models. The simulation is considered
to represent the airfoil. The mesh is extremely simple; it is realised converged when the root mean square of the residuals of the density
using a set of rigid elements that connects several slave nodes to a single equation is decreased of 5 orders of magnitude. The comparison of the
master node, positioned on the rotation axis. The master node only has pressure coefficient distribution obtained in the two cases, reported in
Fig. 8, does not exhibit any difference, demonstrating the ability of the
two degrees of freedom: pitch and plunge. One of the slave nodes, at
interpolation.
the position of the centre of mass, houses the mass and inertia of the
airfoil.
3.1.2. Static aeroelasticity
The interpolation between the structural and fluid meshes uses RBF. In the second problem here considered, the two structural degrees
The limitation of this interpolation is due to the fact that, if a 2D of freedom of pitch and plunge are free to change, but the steady
problem is concerned, the structural points cannot all lie on the same equations for the flow are solved. It should be noted that, equivalently
line. Equivalently, if a 3D problem is tackled, the points cannot lie all to the solution of the aerodynamics which exploits the pseudo time
in the same plane. For this reason, the thickness is represented in the integration, the structural equations are also solved integrating in time.
FEM structural mesh as shown in Fig. 5. Indeed, including damping and inertia in the structural solution helps
It can be noted that the exact geometry of the airfoil is not required. convergence as it avoids over/under shoots of the solution. In this
case, the time step chosen is usually relatively large, as there are no
The interpolation will take care of displacing correctly the fluid mesh.
constraints given by the Nyquist criterion, or the numerical damping.
However, thickness must somehow be represented.
The inertial and damping characteristics can also be chosen arbitrarily,
In what follows, the results of four different analyses of increasing as, in the limit of convergence, both velocity and acceleration are zero.
complexity are reported, validating the proposed methods on the test The stiffness in pitch and plunge are respectively 𝐾ℎ = 205 N∕m and
case at hand. 𝐾𝛼 = 2025 N m∕rad. The aerodynamic conditions are coincident with
For the first three subsections, the aerodynamic model will be those used for the previous test case, except for the Mach number,
based on Reynolds Averaged Navier–Stokes (RANS) equations, with which is reduced in order to maintain small deformations. The Mach
Menter’s Shear Stress Transport (SST) turbulence model, and a pseudo- number is halved. The initial angle of attack of the airfoil is also the
time CFL number of 20. In the last subsection, related to the dynamic same, thus it is 𝛼 = 3◦ . The convergence criterion still requires the
aeroelasticity, also a Euler model will be used, again with a local CFL of density residuals to drop of 5 orders of magnitude. As far as the FSI
loop is concerned, convergence is based on the difference of the norm
20. In both cases, the linear solver is a Flexible Generalised Minimum
of the displacements vector between two FSI iterations, as shown in
Residual (FGMRES). This solver is relatively memory intense, but, for
Fig. 3, which has to be smaller than 1 micrometer. The result of the SU2
this small case, provides the best performance in terms of accuracy.
analysis is compared with the theoretical solution, and with the X-FOIL
Depending on the aerodynamic model, two different meshes are software. The former is obtained from simple linear incompressible
used. In the first mesh, for the RANS simulations, the fluid domain theories. As the rotation axis is placed exactly on the quarter-chord, and
is discretised with 133,000 nodes, with refining close to the airfoil the airfoil is symmetric, no pitch rotation is predicted by the theory. As
surface, in order to correctly represent the turbulent boundary layer. far as the plunge is concerned, the equation will simply read:
The first cell is placed at a height of 𝑦+ ≈ 1. In the second mesh, there 1 2
𝐾ℎ ℎ = 𝜌𝑈 𝑆𝐶 𝛼 (1)
is no need to resolve the boundary layer. The distribution of points, 2 ∞ 𝑙𝛼
away from the airfoil surface, is the same as before. A total of 19,500 Solving for ℎ, and using the values for the present example, we obtain:
nodes are used in this case. The base mesh for both models is pictured
in Fig. 6, while Fig. 7 compares the different treatments of the layers 1 2 𝐶𝑙𝛼
close to the surface. ℎ= 𝜌𝑈 𝑆 𝛼 = 0.29 m (2)
2 ∞ 𝐾ℎ

6
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

Fig. 6. Aerodynamic mesh for the RANS simulations of the NACA 0012.

Fig. 7. Different treatment of the near wall cells in RANS and Euler simulations, for the NACA 0012 airfoil.

Table 1
Pitch rotations and plunge displacements, as predicted by
various methods.
Model ℎ (m) 𝜃 (rad)
Theory 0.289 0
X-Foil 0.281 4.0e−04
SU2 0.306 3.4e−04

is imposed as a sinusoidal wave with an amplitude of 1 degree, no bias,

and a frequency of 8 Hz. This prescribed motion is imposed through the
interface. The equations of motion for the structure are not integrated.
The equations of fluid dynamics are solved considering at each time
step the shape associated to the assigned evolution of the pitch angle.
The simulation is run in SU2 with a Mach number of 0.1, and Re
number equal to 6 millions. Each time step, with size of 1 millisecond,
the aerodynamic convergence is reached with a drop of around 5 orders
of magnitudes in the density residuals. After a small initial transient,
Fig. 8. Coefficient of pressure distribution around the NACA 0012 airfoil, at 3◦ of the solution reaches a periodic oscillation. The comparison of the lift
angle of attack, as computed imposing the flow direction in SU2 or the pitch rotation coefficient from the CFD simulation with the one from Theodorsen
in the FSI interface.
theory is presented in Fig. 9.
The theory predicts a lift coefficient with the following expression:
[ ] [ ( ) ̇ ]
In order to also capture the effect of the pitching moment, results from 𝑐 𝜃̇ 𝜃̈ 3𝑐 𝜃
𝐶𝑙 = 2𝜋 − (𝑥𝑓 − 𝑐∕2) + 2𝜋𝐶(𝑘) 𝛼 + − 𝑥𝑓 (3)
X-FOIL, in terms of coefficient of lift and coefficient of moment, have 4 𝑈 𝑈2 4 𝑈
been used [51], with similar Re. Where the position of the rotation axis is identified with 𝑥𝑓 , and the 𝑥
In Table 1 there is a comparison of the results obtained with the vector is positive in the flow direction. As, in our case, the rotation axis
three different methods, showing good match in all cases. The values is placed at the quarter chord, a positive acceleration and velocity of
predicted by SU2 show almost no pitch rotation and a similar plunge the pitch degree of freedom will increase the coefficient of lift, thanks
to the other methods. to the added mass effect and the dynamically induced angle of attack.
In the figure, it can be seen a good match between the incompress-
3.1.3. Forced motion aerodynamics ible theory and the CFD computation. The same phase is predicted and
The first time-variant analysis here considered is the response to a only a small difference in the amplitude can be seen. This difference
forced motion. The plunge displacement is fixed while the pitch angle in amplitude is relatively common when comparing a prediction made

7
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

Table 2
Density values used for the RANS simulations of the NACA
0012 dynamic FSI.
Mach Density (kg∕m3 )
0.1 2.072
0.2 1.036
0.3 0.691
0.357 0.580
0.364 0.569

are present. Thus, comparing the Theodorsen theory and SU2 is not
fully meaningful.
The time domain histories match nicely until the flutter point is
approached. Here, both because the Mach is higher and nonlinear
effects start to appear, and because the thick CFD model predicts flutter
sooner, the solutions start to diverge. In general, for the first three Mach
numbers, the match is good. As often happens, the Euler equations
Fig. 9. Unsteady coefficient of lift of the NACA 0012 airfoil, computed by SU2 with overestimate the lift peaks, while including viscosity the results are
the FSI interface and by the Theodorsen function. Imposed sinusoidal pitching motion
closer to the incompressible theory. The fourth Mach number is clearly
with zero mean angle of attack, 1◦ amplitude, and 8 Hz frequency.
at the edge of flutter, with pitch and plunge moving at a very similar
frequency, and decaying slowly. It can also be seen that the Euler
simulation seems to tend more to the instability. The last Mach number
with the thin foil assumption and a computation that takes into account is past the flutter point. Very large oscillations are present, but they
the entire thickness or real experiments [52]. This difference can are not physical. Indeed, we should recall that the structural model
also be seen in the previous results, where, in Table 1, the plunge is based on a linear FEM, which fails when large displacements or
displacement coming from SU2 is larger than the one predicted with rotations are involved. Confirming our predictions, Euler equations
the coefficients coming from X-FOIL and from the theory. predict a faster increase in the Limit Cycle Oscillation (LCO) amplitude.
The comparison with the thin airfoil, small displacement, theory is not
3.1.4. Dynamic aeroelasticity meaningful at this point.
For this final example, two possible aerodynamic models have been
considered. The first is based on the RANS equations used before. The 3.2. Benchmark SuperCritical Wing
other uses Euler equations for the aerodynamics. This has been done to
completely verify the code for all the possible uses. In this section, a more complex case will be tackled. From the
The Mach number will be gradually increased, with a fixed Re structural point of view, no much complicacy is added. However, from
number of 4 millions and a fixed temperature of 273.15 K, until the the aerodynamic point of view, the flow is significantly more nonlinear,
instability point of the aeroelastic system is reached. The classical in the transonic range, exhibiting high unsteadiness and continuous
pitch–plunge flutter will then be visible. shocks formation. The case is commonly known as the BSCW [53]. The
wing is a semirigid, semispan, rectangular wing based on the SC(2)0414
It must be noted that, to keep the Re number fixed in the RANS
supersonic airfoil, with two degrees of freedom in pitch and plunge. The
calculation, so that the same flow of energy from the momentum to
chord is 0.4064 m (16 in.) and the span is 0.8128 m (32 in.).
the energy equations is present in all the calculations, the density of
The analyses considered in this section reproduce two of the ref-
the flow must vary for the different Mach numbers.
erence test cases used for the Aeroelastic Prediction Workshop [54],
The equations presented at the beginning of this section, governing
whose results can be compared with wind-tunnel experiments of the
the airfoil dynamics, are usually adimensionalised to obtain results
BSCW model that were conducted in the NASA Langley Transonic
independent from the free-stream density of the flow.
Dynamics Tunnel. A picture of the wing mounted on the suspending
Indeed, we can define the following parameters: apparatus is reported in Fig. 12, taken from Heeg et al. [18].
𝑆𝑚 2 𝐼𝑓 𝜔 𝑚 The first case here reported is a forced excitation analysis, in which
𝜒= , 𝑟 = , 𝜔̄ = ℎ , 𝜇𝑚 = (4)
𝑚𝑏 𝛼 𝑚𝑏2 𝜔𝛼 𝜋𝜌∞ 𝑏2 an oscillating turntable varies the pitch of the wing about an axis at
√ √ the 30% of the chord [55]. The second case is the flutter analysis
𝐾ℎ 𝐾𝛼 of the pitching and plunging wing [56]; unlike the previous case,
Where 𝑏 is the semi chord of the airfoil, 𝜔ℎ = 𝑚
, 𝜔𝛼 = 𝐼𝑓
. If we
the pitch motion is about the centre chord. Although the BSCW was
fix them, the structure will behave always the same regardless of 𝜌∞ .
semirigid [56], the simulations are performed on a rigid wing, with a
The selected parameters for this example are taken from Sanchez
single degree of freedom in the forced case and two degrees of freedom
et al. [34]: 𝜒 = 0.25, 𝑟𝛼 = 0.5, 𝜔𝛼 = 45 rad∕s, 𝜔̄ = 0.3185 and 𝜇𝑚 = 100.
in the flutter case.
On the other hand, as the Re number does not appear in the Euler
From the structural modelling point of view, the system is made of
equations, in that case we can directly fix both the temperature and one master node only, with two degrees of freedom. These correspond
pressure, thus the density. to the pitching and the plunging of the wing. Several slave nodes are
The different values of Mach numbers and density values, for RANS connected rigidly to the master node, and are positioned so to represent
simulations, are reported in Table 2. the thickness and the span of the three-dimensional wing, as previously
No structural damping is included and a time step of 1 millisecond done for the test case of the NACA airfoil. In order to retain stability,
is used. At each time step, the inner FSI convergence is based on a they cannot all lie in one single plane. Lumped masses, positioned at the
maximum norm of the structural displacement vector of 1 micrometer. slave nodes, are used to represent the mass distribution of the model,
The comparison of the time histories is shown in Fig. 10. The results and two spring elements are placed on the free degrees of freedom of
in terms of frequencies are reported in Fig. 11. the master node to represent the elastic properties.
It can be seen how the frequency merging is well captured; after the The aerodynamic model is significantly more complex, due to the
flutter point the two frequencies are coincident and nonlinear effects physical phenomena to be represented. RANS equations will be used,

8
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

Fig. 10. Time domain responses of the NACA 0012 airfoil to an initial condition in pitch of 5◦ , for different Mach numbers.

9
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

gas at Mach 0.7, Re number equal to 4.5 millions and a 3 degrees mean
angle of attack.
The available experimental data is the unsteady pressure distri-
bution at the 60% span station. Therefore, the results reported here
are limited to that station. Fig. 14 depicts the variation in time of
the pressure coefficient at the 60% span station. It can be noted the
formation of a shock wave at the suction side, during the pitching
motion. Indeed, it can clearly be seen in Fig. 14 how, when the angle of
attack approaches the value of 4◦ , a shock forms at around 10 percent
of the chord.
The transfer function between the pressure coefficient and the pitch
angle is also computed. It is then compared with that obtained from
the experimental data. Fig. 15 represents the amplitude of the transfer
function for both the upper and lower surfaces. The phase is reported
in Fig. 16.
The obtained results are really close to those in the available liter-
ature. An excellent review paper, prepared after the second aeroelastic
prediction workshop [54], compares results obtained by all the par-
Fig. 11. Modal frequencies as a function of the Mach number for the NACA 0012 ticipants at the workshop and provides the required information to
airfoil. understand the present example. In all the simulations, it can be seen
how most of the amplitude response is at the leading edge, and this
is correctly predicted by the numerical methods, including the present
one. Two peaks are visible on the upper side. The first, is due to the
acceleration of the flow as the wing pitches up, and the deceleration
when the wing pitches down. This is due to the geometry of the airfoil
only. The second, higher, is due to the formation of the shock wave.
The flow is attached, as it can be seen in the phase plot. Indeed, on the
upper surface, with a positive pitch up movement, we have a decrease
in pressure coefficient (phase is −180◦ ), while on the lower side the
opposite is true. A peak in the phase can be observed in the upper
surface, after the shock. This may be due to a small separation bubble
formed by the shock itself that creates delay in the response of the
pressure coefficient to pitching.
On the lower side, only one peak in the magnitude plot is visible,
as there is no shock wave formation.
Discrepancies are found between simulations and experimental val-
ues close to the shock wave. One sensor on the upper surface is not
matching simulation values. The same discrepancies have been found
by all the other participants to the workshop, and have been attributed
Fig. 12. Physical apparatus used in the experiments for the BSCW.
to a fault sensor.
Source: Taken from Heeg et al. [18].
In the phase plot, differences arise also close to the trailing edge
due to the transfer function going to zero. Under these conditions, the
ratio between imaginary and real part of the transfer function is ill
with a standard Spalart–Allmaras turbulence model, as large regions conditioned and the computation prone to error.
of separation are not expected to occur, up to the instability point. As stated before, the results well match others in the literature and
The physical thermal conductivity is modelled with a constant Prandtl confirm the applicability of the developed framework for complex, 3D,
number of 0.755, while the turbulent conductivity is not modelled. aerodynamically nonlinear cases.
The aerodynamic boundary conditions do not take into account the
presence of the wind tunnel itself, as this is slotted and there should 3.2.2. Dynamic aeroelasticity
not be any shock reflection. Only the splitter plane is modelled, using a The final test for the current framework was the search of the flutter
no-penetration boundary condition. The other external boundaries are point for the BSCW wing. This has also been called the case 2, in
placed far from the wing, at approximately 100 chords from it, and are the second aeroelastic prediction workshop. Experimental results are
modelled with a far-field Riemann boundary condition. At the wing, no- available and show a flutter dynamic pressure of 8082 Pa at Mach
slip is imposed, and cells close to the surface are placed at 𝑦+ ≈ 1. The number 0.74, Re number equal to 4.4 millions, and no angle of attack.
mesh is unstructured, with 3 million nodes and prison layers close to The model is characterised by a plunge mode with a frequency of
the surface, and it is reported in Fig. 13 together with the identification 3.3 Hz and a pitch mode with a frequency of 5.2 Hz.
of the boundary conditions. First, a coupled simulation is run using the experimental values for
As the computation is now significantly more expensive, an adaptive the flutter point. Results, in terms of displacements of the wing at the
strategy for the CFL number is used. The minimum is set to 1, while Leading Edge (LE) and Trailing Edge (TE), and twist angle, are reported
the maximum is set to 100. Due to the possibility of having a relatively in Fig. 17. It can be seen that the system is only lightly damped,
large local pseudo time, the Bi-Conjugate Gradient method is used for confirming the proximity of the flutter point. However, it is stable.
the linear solver. A physical time step size of 1 ms is used. In order to test whether the code could actually predict an unstable
behaviour for the BSCW, the dynamic pressure was increased by 5%.
3.2.1. Forced motion aerodynamics The simulation is then run again, with the same parameters except
In this forced response case, the wing is oscillated in pitch, with for the new dynamic pressure, and the time histories are reported
an amplitude of 1 degree at 10 Hz. The analysis is performed in R-134a in Fig. 18. It can be clearly seen an unstable behaviour with the

10
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

Fig. 13. Aerodynamic mesh for the RANS simulation of the BSCW. Euler boundary condition in orange, far-field boundary conditions in green, no-slip boundary condition in blue.

600 mm which limits wall effects while maximising the local thickness.
The aerodynamic reference condition is characterised by a freestream
Mach number of 0.82.
The aeroelastic model is equipped with a classical aileron able to
dynamically move with a bandwidth which is compatible with the
frequency content of the gust generator. The hinge axis is located at
75% of the local chord, between 65% and 85% of the wing span [58].
The model can be connected to the wind tunnel in three ways.
First, with a rigid connection, used to impose a precise angle of attack.
Second, with an elastic connection designed so that the pitch mode of
the demonstrator is at 20 Hz. Finally, with a motorised connection, able
to force the pitch with varying frequency and amplitude [59].
A linear unsteady aerodynamic method cannot take into account
the nonlinearities at the considered reference Mach number. For this
reason, the FSI approach based on CFD is adopted as higher fidelity
Fig. 14. Coefficient of pressure distribution at the 60% span station of the BSCW. method for flutter verification. It must be pointed out that the CFD
Imposed sinusoidal pitching motion with 3◦ mean angle of attack, 1◦ amplitude, and computation is used to include the nonlinearities effects of the aerody-
10 Hz frequency. namics, but the finite element model is still linear. After flutter studies,
the stability margin with respect to the position of the rotation axis is
investigated, considering three different positions in wind direction.
displacements growing in amplitude, especially for the wing twist. This
confirms that the numerical flutter point is in good agreement with the 3.3.1. Structural modelling
experimental one, and it also confirms the capabilities of the developed The wing of the complete wind tunnel model is made of a single
code. piece, milled from solid, providing a common steel platform. This
platform is directly connected to pitch shaft on one side of the fuselage.
3.3. Transonic wind tunnel test simulation of a flexible demonstrator Multiple cover parts are connected at the lower surface of the wing
where instrumentation and aileron actuation have to be installed. The
The Clean Sky 2 European research project GUDGET aims at the aileron linkage also consists of a complete single piece mechanism,
design and manufacturing of an innovative experimental set-up for the equipped with compliant hinges, which excludes any free play, but
investigation of gust loads in transonic flow conditions. The experimen- adds stiffness to the actuation kinematics. All cover parts restore the
tal set-up is composed of a gust generator and an aeroelastic half-model structural stiffness of the wing and make it locally lighter.
to be installed in the transonic ONERA S3Ch wind tunnel facility. One The finite element model is based on a continuous plate discreti-
of the main goals of the wind tunnel tests is to study nonlinear effects sation able to reproduce the stiffness distribution of the actual wing,
induced by high amplitude gust loads on the aeroelastic behaviour of considering all local details described above. In this plate model, the
the model. Since the structural behaviour is expected to be linear, the camber distribution represents the mean surface of the wing, while
numerical verifications should be performed considering the interaction the thickness distribution is used to assign thickness properties to each
between a linear finite element model and an unsteady aerodynamic plate element by a mapping able to extract the information from the
model able to capture nonlinearities effects. Thus, it is the perfect solid Computer Aided Design (CAD) model. This plate model is then
application of the present code. tuned in terms of mass distribution matching the inertial properties
The wind tunnel is a transonic facility with a 0.8 m × 0.8 m obtained by the CAD model. The fuselage is considered as rigid, and
square test section of 2.2 m length. It covers a Mach number range its inertial properties are introduced into the model by lumped masses
from 0.3 to 1.2 and operates at atmospheric stagnation pressure and placed in the centre of gravity of each mechanical component included
temperature. The aeroelastic half-model consists of a wall-mounted in the CAD model. The complete model is balanced by two additional
fuselage connected to a swept wing. The aerodynamic wing shape is masses placed in the front and the back of the pitch axis position.
based on the supercritical OAT15 A airfoil [57]. The swept angle is All these lumped masses are rigidly connected to the rotational axis
equal to 30 deg, while the wing twist ensures a shock wave parallel together with additional nodes that are used for the interpolation of
to the leading edge, as well as a constant pressure along the span in the fuselage displacements on the aerodynamic mesh.
cruise conditions, without separation at the wing root. The chord varies The complete finite element model is validated using the vibration
from 204.55 mm at the root to 170.45 mm at the tip, over a span of modes computed on a full-solid model consisting of about half million

11
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

Fig. 15. Transfer function magnitude between the coefficient of pressure of the BSCW at the 60% span station and the pitch rotation. Imposed sinusoidal pitching motion with
3◦ mean angle of attack, 1◦ amplitude, and 10 Hz frequency.

Fig. 16. Transfer function phase between the coefficient of pressure of the BSCW at the 60% span station and the pitch rotation. Imposed sinusoidal pitching motion with 3◦
mean angle of attack, 1◦ amplitude, and 10 Hz frequency.

Fig. 17. Time histories of the wing displacements and wing twist for the experimental flutter point.

12
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

Fig. 18. Time histories of the wing displacements and wing twist for the numerical flutter point.

Fig. 19. Aerodynamic mesh for the GUDGET model in the wind tunnel. Red surface identifies the inlet, orange identifies the outlet, and green the no-penetration boundary
conditions.

tetrahedral elements and not suitable to perform aeroelastic analyses shock thanks to appropriate twisting of the wing. Due to the separation
coupled with CFD computations. Before using the plate model to verify present, SST was chosen as turbulence model.
the flutter behaviour in the transonic reference condition, several para- A representation of the mesh is reported in Figs. 19 and 20. The
metric analyses were performed for different values of the stiffness to be boundary patches are coloured based on the boundary condition ap-
assigned to the 2 aileron actuators and to the pitch mechanism: values plied. In red, the inlet is pictured, with orange we represent the outlet,
close to what is expected from the model design are 3600 Nm/rad for while green identifies no penetration. Finally, the no-slip boundary
the pitch mode and a total value of 31.8 Nm/rad for the aileron. condition is identified in blue. Please note that some surfaces are not
solid coloured, but only wireframe, for visualisation purpose. In the
3.3.2. Aeroelastic modelling figures, it can clearly be seen the refinement close to the model surface
The aeroelastic model consists of a modal base, reduced to the due to the prison layer cells mentioned above.
first six vibration modes computed on the finite element plate model As the mesh is relatively coarse, FGMRES was used as a linear solver
described in Section 3.3.1, coupled with a CFD model prepared starting for the local solution. Indeed, we are not interested in detailed aerody-
from the external surfaces of the CAD model. namic performances, but rather on macroscopic aeroelastic properties.
The CFD mesh, obtained after a convergence study, is made of an The latter usually require coarser mesh and this was exploited to speed
unstructured block, refined at the wing surface to include 50 prison up the computations. Further, a local CFL number of 40 was used for
layers, required for the boundary layer resolution, the first of which is the same purpose.
placed at 𝑦+ ≈ 1. The mesh consists of 2 million nodes and represents The physical time step size is 0.5 ms and 600 time steps are
the wind tunnel test section, which is a rectangular channel, and the performed, in order to assess the stability. Of these 600 time steps, the
half-plane model. In order to avoid spurious effects due to the boundary first 120 are actually purely aerodynamic. Indeed, we first initialise the
conditions, the test section is increased in length to 9 m. No-penetration correct aerodynamics, and only later we let the coupled system evolve.
boundary conditions were imposed on all the walls of the channel, This is done in order to avoid spurious overshoots in the structural
no-slip boundary condition was imposed on the model surface, and solution, at the very beginning of the simulation.
the total thermodynamic quantities of pressure and temperature were In Figs. 21 through 26, the modes used for the simulation are
imposed at the inlet. The outlet static pressure was then imposed so reported. On the left, the modes are plotted on the structural mesh.
that in the middle of the channel the Mach number was equal to On the right, the same modes are also reported after the RBF mapping
0.82. Results of a steady simulation at the prescribed Mach number from structural mesh to aerodynamic one. It can be seen how the shapes
are reported on the left of Fig. 28. Here, the coefficient of pressure is match nicely, proving the quality of our methodology.
plotted over the surface. It can be seen that, as already mentioned in the In the following results, different rotational axis positions are con-
introduction, the wing has been designed to have a smooth variation of sidered. The rotational axis is originally placed at the quarter Mean
the shock location along the span. Further, the root section is free from Aerodynamic Chord (MAC). The stability at the nominal operating

13
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

Fig. 20. Close-up view of the aerodynamic mesh for the GUDGET model in the wind tunnel. Light blue surfaces identify the no-slip boundary conditions.

Fig. 21. Comparison between pitch modes at the structural and aerodynamic side (20.1 Hz).

Fig. 22. Comparison between first bending modes at the structural and aerodynamic side (63.1 Hz).

Fig. 23. Comparison between aileron modes at the structural and aerodynamic side (120.7 Hz).

Fig. 24. Comparison between second bending modes at the structural and aerodynamic side (252.4 Hz).

14
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

Fig. 25. Comparison between bending/torsional modes at the structural and aerodynamic side (352 Hz).

Fig. 26. Comparison between torsional modes at the structural and aerodynamic side (459.5 Hz).

Fig. 27. Time histories of the GUDGET model pitch degree of freedom, for different rotational axis position, considering rigid (left) and fully flexible (right) model.

condition of Mach 0.82 is first assessed. Then, in order to verify the sufficient to excite oscillations around the pitch axis. Depending on
sensitivity of this stability to changes in the rotational axis position, the the relative position between the centre of rotation and the centre of
latter is moved of ±10% MAC from the original position and the stability pressure, the equilibrium point may be characterised by a positive or
analyses are repeated. The centre of mass of the shaft sustaining the negative angle of attack. It can clearly be seen that the system, in these
model is also moved to correctly represent the new mass distribution. aerodynamic conditions, is stable. Few cycles are required to damp out
the oscillations and reach the asymptotic value.
3.3.3. Rigid pitch stability analysis When we shift the rotational axis towards the leading edge, a larger
As stated above, the model has different elastic modes, but can also negative pitch moment is created, meaning larger oscillations at the
pitch with respect to the wind tunnel. The pitch is restrained with beginning of the simulation, and a larger negative equilibrium value.
an elastic system, introducing in the system a characteristic frequency Further, it can be noted that, when the rotational axis is placed at 35%
related to this body-wide movement. This frequency is the lowest one of the MAC, the equilibrium point shows a positive angle of attack.
in the system itself, thus it is of paramount importance to understand It is then clear that, for an intermediate position between 35% and
whether it will be stable or not. Indeed, the gust excitation provided to 25% there must be a point where the equilibrium is exactly at zero
the model will mostly influence this trajectory, and the model must be angle of attack. This is important to note as, in the proximity of this
dynamically stable in all cases. position for the rotational axis, we may find the minimum required
For this reason, the first set of analyses concentrated on this move- torque for the pitch actuation system. It is also to be expected that,
ment only. The model was kept rigid, except for the first mode which if the rotational centre is moved further in the positive 𝑥 direction,
represents the pitch itself. In Fig. 27 left, time histories of the pitch, for an unstable condition would be found. However, in the range of our
the different positions, are reported. interest, this was not the case.
At the beginning of the simulation the system is in the undeformed The stability of the system is probably due to the shock. This can
configuration, which has no angle of attack with respect to the wind. be seen in Fig. 28. Here, the evolution of the pressure coefficient
However, this position is not the equilibrium position and this is is reported, for two time moments, for the original position of the

15
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

Fig. 28. Pressure coefficients at the time instants corresponding to the maximum (left) and minimum (right) angle of attack.

rotational axis. The first time instant corresponds to the maximum peak The open-source coding is another important feature of the present
in angle of attack, which is also the initial condition, the last one to effort, as this increases significantly the spectrum of the possible users.
the minimum peak. It can be seen the movement of the shock towards The hope of the authors is that more and more practitioners will tend
the leading edge and its increase in strength, as the model pitches up. towards high-fidelity aeroelasticity thanks this work.
However, this mostly occurs at the tip, due to the twisted design of the In the future, extensions are planned to allow for an automatic
wing. identification of aerodynamics thanks to the possibility of imposing
Due to this effect, together with the sweeping of the wing, the over- the structural motion. Different techniques can then be implemented
all centre of pressure shifts back due to a positive pitching, stabilising that will eventually provide with a state space approximation of the
the model. aerodynamics. This will fill the gap between the efficiency of the
low-fidelity method DLM and the computational intensity of the high-
3.3.4. Flexible body and free pitch stability analysis fidelity method used in this work. Indeed, the user will be provided
After the confirmation that the pitch rotation was stable in the with the opportunity to run only few full simulations, and reuse the
operating condition, the full flexibility of the system was taken into linearised solutions afterwards.
account. In Fig. 27 right, the time histories of the pitch degree of
freedom are again reported, but this time contributions at different
CRediT authorship contribution statement
frequencies can be noted. Indeed, together with the pitch rotation,
which has a frequency of 20 Hz, higher frequencies appear in the
Nicola Fonzi: Conceptualization, Methodology, Software, Valida-
solution. Mainly, the bending frequency at approximately 60 Hz. Again,
tion, Writing – original draft, Visualization. Vittorio Cavalieri: Con-
in all cases, the system is stable. However, it can be seen that the
ceptualization, Methodology, Software, Validation, Writing – original
contributions at higher frequencies are less damped than the pitch
draft, Visualization. Alessandro De Gaspari: Writing – review & edit-
mode. Especially when the rotational axis is displaced of +10%, the
bending mode is only slightly reduced. It is then expected, at higher ing, Supervision, Funding acquisition. Sergio Ricci: Writing – review
speeds, to encounter possible dynamic instabilities. & editing, Supervision, Funding acquisition.

4. Conclusions Declaration of competing interest

In the present paper, an update and extension of the Python-based The authors declare that they have no known competing finan-
FSI framework embedded in SU2, for general aeroelastic studies, has cial interests or personal relationships that could have appeared to
been presented. The software, the first version of which derived from influence the work reported in this paper.
CUPyDO, couples a well known, actively developed, finite volume code,
SU2, with different structural solvers, via a Python interface. Thanks Acknowledgements
to the improved version, higher performances can be obtained, and
an easier inclusion of new structural solvers is possible. Further, the This project has received funding from the Clean Sky 2 Joint Un-
new interface allows for four possible solutions; steady and unsteady dertaking under the European Union’s Horizon 2020 research and
simulations, but also imposed motion responses, covering a wide spec- innovation programme under grant agreement No 831802.
trum of applications. A new native structural solver allows for the direct
solution of FSI problems without the need for other external structural
References
solvers. Thanks to the very high level of the programming language,
extensions and modifications are easily included.
[1] I.E. Garrick, W.H. Reed, Historical development of aircraft flutter, J. Aircr. 18
The native solver is designed to solve a set of structural equa- (11) (1981) 897–912, http://dx.doi.org/10.2514/3.57579.
tions of motion coming from a Nastran-like model. This allows for [2] E. Livne, Future of airplane aeroelasticity, J. Aircr. 40 (6) (2003) 1066–1092,
an integration of the present tool also in industrial or standardised http://dx.doi.org/10.2514/2.7218.
workflows. [3] M.M.J. Opgenoord, M. Drela, K.E. Willcox, Influence of transonic flutter on the
conceptual design of next-generation transport aircraft, AIAA J. 57 (5) (2019)
The implementation has been extensively tested on various test
1973–1987, http://dx.doi.org/10.2514/1.J057302.
cases, of increasing complexity. The first test was a standard pitching– [4] T.R. Brooks, J.R. Martins, G.J. Kennedy, High-fidelity aerostructural optimization
plunging airfoil, operating in subsonic conditions, for which analytical of tow-steered composite wings, J. Fluids Struct. 88 (2019) 122–147, http:
results are available. In the second case, a transonic application was //dx.doi.org/10.1016/j.jfluidstructs.2019.04.005.
considered, with a three-dimensional wing, for which experimental [5] B. Stanford, Optimal aircraft control surface layouts for maneuver and gust load
alleviation, in: AIAA Scitech 2020 Forum, Orlando, FL, USA, 2020, pp. 1–10,
values are present. The last test was performed on a fully flexible half-
http://dx.doi.org/10.2514/6.2020-0448.
plane wind tunnel model, operating in the deep transonic regime. This [6] A. De Gaspari, S. Ricci, L. Riccobene, A. Scotti, Active aeroelastic control over
involved highly nonlinear aerodynamic phenomena and demonstrated a multi-surface wing: Modelling and wind tunnel testing, AIAA J. 47 (9) (2009)
the great capabilities of the approach. 1995–2010, http://dx.doi.org/10.2514/1.34649.

16
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

[7] L. Marchetti, A. De Gaspari, L. Riccobene, F. Toffol, F. Fonte, S. Ricci, P. [32] The OpenFOAM Foundation, OpenFOAM v8 User Guide, CFD Direct Ltd,
Mantegazza, E. Livne, K. Hinson, Active flutter suppression analysis and wind 2011–2021, URL https://cfd.direct/openfoam/user-guide.
tunnel studies of an uncertain commercial transport configuration, in: AIAA [33] R. Bombardieri, R. Sanchez, R. Cavallaro, N.R. Gauger, Towards an open-source
Scitech 2020 Forum, Orlando, FL, USA, 2020, pp. 1–17, http://dx.doi.org/10. framework for aero-structural design and optimization within the su2 suite, in:
2514/6.2020-1677. Advances in Evolutionary and Deterministic Methods for Design, Optimization
[8] R. Thormann, D. Dimitrov, Correction of aerodynamic influence matrices for and Control in Engineering and Sciences, Springer, 2021, pp. 291–306.
transonic flow, CEAS Aeronaut. J. 5 (4) (2014) 435–446, http://dx.doi.org/10. [34] R. Sanchez, H.L. Kline, D. Thomas, A. Variyar, M. Righi, T.D. Economon, J.J.
1007/s13272-014-0114-3, Publisher: Springer. Alonso, R. Palacios, G. Dimitriadis, V. Terrapon, Assessment of the fluid-structure
[9] M. Wagner, G. Norris, Boeing 787 Dreamliner, Zenith Press, 2009. interaction capabilities for aeronautical applications of the open-source solver
[10] M.J. Patil, D.H. Hodges, C.E. Cesnik, Nonlinear aeroelasticity and flight dynamics SU2, in: Proceedings of the VII European Congress on Computational Methods
of high-altitude long-endurance aircraft, J. Aircr. 38 (1) (2001) 88–94. in Applied Sciences and Engineering, ECCOMAS Congress 2016, Institute of
[11] K. Isogai, On the transonic-dip mechanism of flutter of a sweptback wing, AIAA Structural Analysis and Antiseismic Research School of Civil Engineering National
J. 17 (7) (1979) 793–795, http://dx.doi.org/10.2514/3.61226. Technical University of Athens (NTUA) Greece, Crete Island, Greece, 2016, pp.
[12] V.D. Chuban, Influence of angle of attack and stabilizer deflection on T 1498–1529, http://dx.doi.org/10.7712/100016.1903.6597.
empennage flutter, J. Aircr. 42 (1) (2005) 264–268, http://dx.doi.org/10.2514/ [35] D.M. Beazley, et al., SWIG: An easy to use tool for integrating scripting languages
1.4213. with C and C++, in: Tcl/Tk Workshop, Vol. 43, 1996, p. 74.
[13] W.P. Jennings, M.A. Berry, Effect of stabilizer dihedral and static lift on T-tail [36] N. Fonzi, S.L. Brunton, U. Fasel, Data-driven nonlinear aeroelastic models of
flutter, J. Aircr. 14 (4) (1977) 364–367, http://dx.doi.org/10.2514/3.58785. morphing wings for control, Proc. R. Soc. Lond. Ser. A Math. Phys. Eng. Sci.
[14] Airbus S.A.S., Airbus reveals new zero-emission concept aircraft, 2021, Airbus, 476 (2239) (2020) 20200079.
URL https://www.airbus.com/newsroom/press-releases/en/2020/09/airbus-
[37] A. De Gaspari, L. Riccobene, S. Ricci, Design, manufacturing and wind tunnel
reveals-new-zeroemission-concept-aircraft.html.
validation of a morphing compliant wing, J. Aircr. 55 (6) (2018) 2313–2326,
[15] R.H. Liebeck, Design of the blended wing body subsonic transport, J. Aircr. 41
http://dx.doi.org/10.2514/1.C034860.
(1) (2004) 10–25.
[38] A. De Gaspari, F. Moens, Aerodynamic shape design and validation of an
[16] W. Li, X. Gao, H. Liu, Efficient prediction of transonic flutter boundaries for
advanced high-lift device for a regional aircraft with morphing droop nose, Int.
varying mach number and angle of attack via LSTM network, Aerosp. Sci.
J. Aerospace Eng. 2019, 7982168 (2019) 1–22, http://dx.doi.org/10.1155/2019/
Technol. 110 (2021) 106451, http://dx.doi.org/10.1016/j.ast.2020.106451.
7982168.
[17] L. Cavagna, G. Quaranta, P. Mantegazza, Application of Navier–Stokes simula-
[39] A. De Gaspari, Multiobjective optimization for the aero-structural design of
tions for aeroelastic stability assessment in transonic regime, Comput. Struct. 85
adaptive compliant wing devices, Appl. Sci. 10 (18) (2020) 30, http://dx.doi.
(11) (2007) http://dx.doi.org/10.1016/j.compstruc.2007.01.005.
org/10.3390/app10186380.
[18] J. Heeg, P. Chwalowski, D.E. Raveh, M.J. Dalenbring, A. Jirasek, Plans
and example results for the 2nd aiaa aeroelastic prediction workshop, in: [40] B. Diskin, J. Thomas, E. Nielsen, J. White, H. Nishikawa, Comparison of
56th AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials node-centered and cell-centered unstructured finite-volume discretizations part
Conference, 2015, p. 0437, http://dx.doi.org/10.2514/6.2015-0437. I: Viscous fluxes, in: 47th AIAA Aerospace Sciences Meeting Including the New
[19] E.H. Dowell, D. Tang, Nonlinear aeroelasticity and unsteady aerodynamics, AIAA Horizons Forum and Aerospace Exposition, in: Aerospace Sciences Meetings,
J. 40 (9) (2002) 1697–1707, http://dx.doi.org/10.2514/2.1853. American Institute of Aeronautics and Astronautics, Orlando, Florida, 2009, p.
[20] C. Huang, J. Huang, X. Song, G. Zheng, G. Yang, Three dimensional aeroelastic 20, http://dx.doi.org/10.2514/6.2009-597.
analyses considering free-play nonlinearity using computational fluid dynam- [41] T.D. Economon, F. Palacios, S.R. Copeland, T.W. Lukaczyk, J.J. Alonso, SU2: An
ics/computational structural dynamics coupling, J. Sound Vib. 494 (2021) open-source suite for multiphysics simulation and design, AIAA J. 54 (3) (2016)
115896, http://dx.doi.org/10.1016/j.jsv.2020.115896. 828–846, http://dx.doi.org/10.2514/1.J053813.
[21] K. Ni, P. Hu, H. Zhao, E. Dowell, Flutter and LCO of an all-movable horizontal [42] A. Jameson, Time dependent calculations using multigrid, with applications to
tail with freeplay, in: 53rd AIAA/ASME/ASCE/AHS/ASC Structures, Structural unsteady flows past airfoils and wings, in: 10th Computational Fluid Dynamics
Dynamics and Materials Conference, American Institute of Aeronautics and Conference, American Institute of Aeronautics and Astronautics, Honolulu, HI,
Astronautics, Honolulu, Hawaii, 2012, p. 9, http://dx.doi.org/10.2514/6.2012- U.S.A., 1991, p. 14, http://dx.doi.org/10.2514/6.1991-1596.
1979. [43] A. Jameson, Origins and further development of the Jameson–Schmidt–Turkel
[22] M.S.C. Software, MSC. Nastran Quick Reference Guide, MSC.Software scheme, AIAA J. 55 (5) (2017) 1487–1510, http://dx.doi.org/10.2514/1.
Corporation, 2004. J055493.
[23] F. Palacios, J. Alonso, K. Duraisamy, M. Colonno, J. Hicken, A. Aranake, [44] A. de Boer, M. van der Schoot, H. Bijl, Mesh deformation based on radial
A. Campos, S. Copeland, T. Economon, A. Lonkar, T. Lukaczyk, T. Taylor, basis function interpolation, Comput. Struct. 85 (11–14) (2007) 784–795, http:
Stanford university unstructured (SU2 ): An open-source integrated computational //dx.doi.org/10.1016/j.compstruc.2007.01.013.
environment for multi-physics simulation and design, in: 51st AIAA Aerospace [45] A. Beckert, H. Wendland, Multivariate interpolation for fluid-structure-interaction
Sciences Meeting Including the New Horizons Forum and Aerospace Exposi- problems using radial basis functions, Aerosp. Sci. Technol. 5 (2) (2001)
tion, in: Aerospace Sciences Meetings, American Institute of Aeronautics and 125–134, http://dx.doi.org/10.1016/S1270-9638(00)01087-7.
Astronautics, Grapevine (Dallas/Ft. Worth Region), Texas, 2013, p. 60, http: [46] L. Dalcín, R. Paz, M. Storti, J. D’Elía, MPI for Python: Performance improvements
//dx.doi.org/10.2514/6.2013-287. and MPI-2 extensions, J. Parallel Distrib. Comput. 68 (5) (2008) 655–662.
[24] D. Thomas, A. Variyar, R. Boman, T.D. Economon, J.J. Alonso, G. Dimitriadis, [47] L.D. Dalcin, R.R. Paz, P.A. Kler, A. Cosimo, Parallel distributed computing using
V.E. Terrapon, Staggered strong coupling between existing fluid and solid python, Adv. Water Resour. 34 (9) (2011) 1124–1139.
solvers through a python interface for fluid-structure interaction problems, in:
[48] M. Karpel, C.D. Wieseman, Modal coordinates for aeroelastic analysis with large
VII International Conference on Coupled Problems in Science and Engineering,
local structural variations, J. Aircr. 31 (2) (1994) 396–403, http://dx.doi.org/
CIMNE, 2017, p. 16.
10.2514/3.46499.
[25] D. Thomas, M.L. Cerquaglia, R. Boman, T.D. Economon, J.J. Alonso, G. Dimitri-
[49] M. Karpel, C.D. Wieseman, Time simulation of flutter with large stiffness changes,
adis, V.E. Terrapon, CUPyDO - An integrated Python environment for coupled
J. Aircr. 31 (2) (1994) 404–410, http://dx.doi.org/10.2514/3.46500.
fluid-structure simulations, Adv. Eng. Softw. 128 (2019) 69–85, http://dx.doi.
org/10.1016/j.advengsoft.2018.05.007. [50] J. Chung, G.M. Hulbert, A time integration algorithm for structural dynamics
[26] N. Fonzi, Dynamic Fluid–Structure Interaction (FSI) Using the Python Wrapper with improved numerical dissipation: The generalized–𝛼 method, J. Appl. Mech.
and a Nastran Structural Model, 2020, URL https://su2code.github.io/tutorials/ 60 (2) (1993) 371–375, http://dx.doi.org/10.1115/1.2900803.
Dynamic_FSI_Python. [51] M. Drela, XFOIL: An analysis and design system for low Reynolds number airfoils,
[27] L. Cavagna, S. Ricci, L. Travaglini, NeoCASS: An integrated tool for structural in: Low Reynolds Number Aerodynamics, Springer, 1989, pp. 1–12.
sizing, aeroelastic analysis and MDO at conceptual design level, Prog. Aerosp. [52] W.J. McCroskey, A cRitical Assessment of Wind Tunnel Results for the NACA
Sci. 47 (8) (2011) 621–635, http://dx.doi.org/10.1016/j.paerosci.2011.08.006. 0012 Airfoil, Technical Report, National Aeronautics And Space Administration
[28] F. Fonte, S. Ricci, Recent developments of NeoCASS the open source suite Moffett Field Ca Ames . . . , 1987.
for structural sizing and aeroelastic analysis, in: The International Forum on [53] J. Heeg, P. Chwalowski, Investigating the transonic flutter boundary of the
Aeroelasticity and Structural Dynamics 2019, IFASD 2019, Savannah, Georgia, benchmark supercritical wing, in: 58th AIAA/ASCE/AHS/ASC Structures, Struc-
USA, 2019, p. 22. tural Dynamics, and Materials Conference, American Institute of Aeronautics and
[29] R.T. Biedron, J.-R. Carlson, J.M. Derlaga, P.A. Gnoffo, D.P. Hammond, W.T. Astronautics, Grapevine, Texas, 2017, p. 24, http://dx.doi.org/10.2514/6.2017-
Jones, B. Kleb, E.M. Lee-Rausch, E.J. Nielsen, M.A. Park, et al., FUN3D Manual: 0191.
13.6, National Aeronautics and Space Administration, Langley Research Center, [54] J. Heeg, P. Chwalowski, D.E. Raveh, A. Jirasek, M. Dalenbring, Overview
2019. and data comparisons from the 2 nd aeroelastic prediction workshop, in: 34th
[30] Y. Levy, D. Raveh, The Eznss Cfd Code Theoretical and User’s Manual, Israeli CFD AIAA Applied Aerodynamics Conference, American Institute of Aeronautics and
Center Ltd, 2015, URL https://www.iscfdc.co.il/sites/default/files/eznss_0.pdf. Astronautics, Washington, D.C., 2016, p. 32, http://dx.doi.org/10.2514/6.2016-
[31] Z. Version, 9.2 User’s manual, ZONA technology, 2017. 3121.

17
N. Fonzi et al. Journal of Computational Science 62 (2022) 101698

[55] J. Heeg, D.J. Piatak, Experimental data from the benchmark super- Vittorio Cavalieri received his M.Sc. degree in Aeronautical
critical wing wind tunnel test on an oscillating turntable, in: 54th Engineering from Politecnico di Milano, Italy. Currently, he
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials is a Ph.D. student at the Department of Aerospace Science
Conference, 2013, p. 1802, http://dx.doi.org/10.2514/6.2013-1802. and Technology, Politecnico di Milano, working on the op-
[56] B. Dansberry, M. Durham, R. Bennett, J. Rivera, W. Silva, C. Wieseman, timal design and experimental validation of morphing wing
D. Turnock, Experimental unsteady pressures at flutter on the supercritical devices based on compliant structures. His research inter-
wing benchmark model, in: 34th Structures, Structural Dynamics and Materials ests include multidisciplinary design optimisation, morphing
Conference, 1993, p. 1592, http://dx.doi.org/10.2514/6.1993-1592. structures, and aeroelasticity.
[57] A.M. Rodde, J.P. Archambaud, A Selection of Experimental Test Cases
for the Validation of CFD Codes, Technical Report AGARD-AR-303, NATO,
1994, URL https://www.sto.nato.int/publications/AGARD/AGARD-AR-303-VOL-
2/AGARD-AR-303-Vol-2.pdf.
Alessandro De Gaspari is Assistant Professor at Politecnico
[58] P. Molton, J. Dandois, A. Lepage, V. Brunet, R. Bur, Control of buffet phe-
di Milano where he was awarded of his Ph.D. in Aerospace
nomenon on a transonic swept wing, AIAA J. 51 (4) (2013) 761–772, http:
Engineering in 2010. His research topics are multidisci-
//dx.doi.org/10.2514/1.J051000, Publisher: American Institute of Aeronautics
plinary design optimisation, morphing aircraft, structural
and Astronautics.
dynamics, aeroelasticity, wind tunnel testing and active
[59] A. Lepage, Y. Amosse, D. Le Bihan, C. Poussot-Vassal, V. Brion, E. Rantet, A
aeroelastic control. He published many scientific papers,
complete experimental investigation of gust load: From generation to active
presented at conferences or published on international jour-
control, in: The International Forum on Aeroelasticity and Structural Dynamics,
nals. He joined several European projects and currently
IFASD 2015, Saint Petersburg, Russia, 2015, p. 19.
he is involved in the CleanSky 2-Airgreen 2 project, with
main focus on Morphing Structures and Load Control and
Alleviation WPs, and he is Co-PI inside GUDGET (GUst
Nicola Fonzi Graduated from Politecnico di Milano with a generators and model DesiGn for transonic wind tunnel
Bachelor Degree in Aerospace Engineering and from ETH tEsTs) project.
Zurich with a Master Degree in Mechanical Engineering,
is now a researcher in the AeroStructures Design Lab Sergio Ricci has a Ph.D. in Aeronautical Engineering from
(ASDL). His main research interests are related to nonlinear Politecnico di Milano where he is full Professor in Aerospace
aeroelasticity with low and high fidelity (CFD) models, both Structures at the department of Aerospace Science and
with numerical and experimental studies. Technology. His research interests include development of
numerical multidisciplinary optimisation procedures for the
conceptual and preliminary design of aerospace structures;
morphing aircraft; dynamic and active control of aerospace
structures; aeroelastic analysis, design, and test. He is
member of different scientific conferences and international
journals committees and author of about 140 publications
on international journals and conference Proceedings.

18