Aerospace Science and Technology 30 (2013) 163–174

Contents lists available at ScienceDirect

Aerospace Science and Technology

www.elsevier.com/locate/aescte

Hopf bifurcation analysis of typical sections with structural

nonlinearities in transonic flow
Elizangela Camilo a,1 , Flávio D. Marques b,∗,2 , João Luiz F. Azevedo c,3
a
Universidade Federal de São Paulo, São José dos Campos, SP, 12231-280, Brazil
b
Universidade de São Paulo, São Carlos, SP, 13566-590, Brazil
c
Instituto de Aeronáutica e Espaço, São José dos Campos, SP, 12228-903, Brazil

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

Article history: The paper is concerned with direct aeroelastic bifurcation analyses of an airfoil system in which both
Received 26 February 2013 aerodynamic and structural nonlinearities are considered. Here, structural dynamics is treated in terms
Received in revised form 23 July 2013 of polynomial nonlinearities associated with the pitching stiffness. Two CFD tools are employed in the
Accepted 27 July 2013
present work and they are based on the Euler formulation. For Hopf bifurcation analysis, a structured
Available online 26 August 2013
grid CFD code is used and flutter boundaries are found with the inverse power method. Previous work
Keywords: has demonstrated the applicability of such approach for both airfoil and wing configurations with a linear
Hopf bifurcation structural model. The novelty of the present effort is the use of this procedure for the investigation of the
Nonlinear systems aeroelastic behavior with structural nonlinearities. Time-marching aeroelastic analysis is also performed
Aeroelasticity and compared with direct calculation of Hopf bifurcation points in order to verify the approach. In
Inverse power method the time-marching case, a CFD code solves the flowfield using an unstructured computational domain
Transonic aeroelasticity discretization. The results shown in the present paper are particularly concentrated in the investigation
of flutter boundaries and typical limit cycle oscillation nonlinear effects for high-subsonic and transonic
flows over a NACA 0012 airfoil-based typical section. The investigation reveals interesting nonlinear
dynamics when both aerodynamic and structural nonlinearities interact.
© 2013 Elsevier Masson SAS. All rights reserved.

1. Introduction cursions in unsteady transonic flows surrounding airfoils. For such

flow conditions, the unsteady forces generated by the motion of
In the last decade, nonlinear dynamic analysis developed the shock waves have been shown to destabilize the airfoil pitching
quickly, both from a theoretical and an experimental point of view, motion and affect the bending-torsional flutter condition by low-
in a vast diversity of fields in science and engineering with impor- ering the flutter speed in the so-called transonic dip phenomenon
tant impact in aeroelasticity. However, most aeroelastic analyses [16]. The structural nonlinearities are subdivided into distributed
of flight vehicles are performed under the assumption of linearity. and concentrated nonlinearities. The distributed nonlinearities are
Under this assumption, the characteristics of flutter and divergence spread over the entire structure, whereas the concentrated nonlin-
can be obtained using well-established tools. On the other hand, earities act locally. The concentrated nonlinearities are known to
the influence of nonlinearities on modern aircraft is becoming in- have significant effects on the aeroelastic responses and to yield
creasingly important and the need for more accurate predictive chaotic motion and limit cycle oscillation (LCO), even below the
tools grows stronger. flutter speed [10,1].
The nonlinearities in aeroelastic analyses are divided into aero- Computational aeroelasticity is the field in which aeroelastic
dynamic and structural nonlinearities. In this paper, the aerody- analyses are performed by means of numerical solution of the
namic nonlinearities arise from the presence of shock wave ex- structural motion equations together with the unsteady aerody-
namic loading equations. Computational Fluid Dynamics (CFD) cal-
culations are generally used when the transonic flow regime is
*Corresponding author.
considered (examples can be found in Refs. [14,12,21] and [11]).
E-mail addresses: [email protected] (E. Camilo), [email protected]
(F.D. Marques), [email protected] (J.L.F. Azevedo). In this context, a significant amount of effort has been devoted to-
1
Adjunct Professor, Institute of Science and Technology, Universidade Federal de wards the numerical solution of transonic aeroelastic phenomena,
São Paulo, UNIFESP.
2
not only in the prediction of transonic dip effects [7,8], but also
Associate Professor, Aeroelasticity Laboratory, Escola de Engenharia de São Car-
los, EESC/USP.
towards the calculation of LCO. Euler and Navier–Stokes schemes
3
Senior Research Engineer, Aerodynamics Division, Departamento de Ciência e have been coupled with Computational Structural Dynamics (CSD)
Tecnologia Aeroespacial, DCTA/IAE/ALA. models [2,15] for such work.

1270-9638/$ – see front matter © 2013 Elsevier Masson SAS. All rights reserved.
http://dx.doi.org/10.1016/j.ast.2013.07.013
164 E. Camilo et al. / Aerospace Science and Technology 30 (2013) 163–174

Time-marching analysis of nonlinear aeroelasticity can give the by the approximate Riemann solver of Osher and Chakravarthy, us-
detailed motion characteristics. However, the cost of these cal- ing Monotone Upstream-centered Schemes for Conservation Laws
culations motivates attempts to find quicker ways of evaluating (MUSCL) interpolation for second-order accuracy and the van Al-
stability, while still retaining the detailed aerodynamic information bada limiter [19]. Eq. (5) is a nonlinear system of algebraic equa-
given by CFD predictions. Due to large computational requirements tions. These equations are solved by an implicit method [9], the
associated with time domain analysis in nonlinear aeroelasticity, main features of which are an approximate linearization, to reduce
research work has used bifurcation theory for computing flutter the size and condition number of the linear system, and the use of
points more efficiently [7,8,17]. Hopf bifurcation defines a singu- a preconditioned Krylov subspace method to calculate the updates.
lar point in which an eigenvalue of the system Jacobian matrix The steady-state solver is applied to unsteady problems within a
crosses the imaginary axis at the flutter condition. In this fashion, pseudo-time stepping iteration [13].
the problem of locating a one-parameter stability boundary is re- For the time-marching aeroelastic solver, the 2-D Euler equa-
duced from multiple, unsteady, time-marching computations to a tions are discretized by a finite volume procedure in an unstruc-
single steady-state calculation of a modified system [22,4,5]. tured mesh. The equations are discretized in space by a centered
The current paper presents an investigation on the effects of scheme, together with added artificial dissipation terms. The arti-
concentrated structural nonlinearities on the Hopf bifurcation pre- ficial dissipation operator, D, can be written as
diction methodology for an airfoil moving in pitch and plunge. In
order to solve the aerodynamic problem using the Hopf bifurcation D = d2 (w f ) − d4 (w f ), (6)
methodology, the Euler equations are discretized using a struc-
tured grid [7]. Time-marching analyses are also performed and where d2 (w f ) represents the contribution of the undivided Lapla-
these results are compared with the Hopf bifurcation predictions. cian operator, and d4 (w f ) is the contribution of the biharmonic
In the CFD code employed for time-marching analyses, the gov- operator [14].
erning equations are integrated by a finite volume discretization The biharmonic operator is responsible for providing the back-
on unstructured grids [3]. Post-bifurcation behavior is analyzed ground dissipation to damp high frequency uncoupled error modes
to show influence of nonlinear structural terms on LCO with the and the undivided Laplacian artificial dissipation operator prevents
time-marching solver. The results are, then, presented to illustrate oscillations near shock waves. The Euler solver is integrated in time
the performance of the respective schemes in resolving the non- by a second-order accurate, 5-stage, explicit, Runge–Kutta time-
linear dynamics in an efficient manner. stepping scheme [18].
Time integration of the coupled fluid-structure equations of
2. Aerodynamic simulations motion is obtained as follows:

In the present study, the flow is assumed to be governed by 1. Starting from a converged steady-state solution of the flow
the two-dimensional, time-dependent Euler equations, which may over the rigid airfoil, perform a time step of the Euler equa-
be written in conservative form and Cartesian coordinates as, tions in the initial position of the airfoil and calculate values
for C l and C m ;
∂w f ∂ Fi ∂ Gi 2. The new position and velocity of the airfoil are obtained by
+ + = 0, (1)
∂t ∂x ∂y solving the dynamic equations of motion, Eq. (7), using a
fourth-order Runge–Kutta time-stepping scheme;
where w f = (ρ , ρ u , ρ v , ρ E ) T denotes the vector of conserved
3. The aerodynamic mesh is moved, and the mesh nodal veloci-
variables, Fi and Gi are the inviscid flux vectors, given by, ties are calculated, in order to accommodate the new position
⎛ ⎞
ρU ∗ and velocity of the airfoil;
⎜ ρ uU ∗ + p ⎟ 4. The unsteady Euler equations are resolved in order to obtain
Fi = ⎝ ⎠, (2)
ρ vU ∗ new aerodynamic coefficients;
∗ 5. Then, one can return to step 2 to calculate new position and
U (ρ E + p ) + ẋp
⎛ ⎞ velocity of the airfoil, and the process is repeated.
ρV ∗
∗
⎜ ρ uV ⎟
Gi = ⎝ ⎠, (3) 3. Equations of motion
ρvV ∗ + p
∗
V (ρ E + p ) + ẏ p The equations of motion of the aeroelastic system are based on
where, ρ , u, v, p and E denote the density, the two Cartesian the typical section. With a linear structural behavior, the dynamics
components of velocity, the pressure and the specific total energy, are represented by
respectively, U ∗ and V ∗ are the two Cartesian components of the
dws
velocity relative to the moving coordinate system, which has local = Rs , (7)
velocity components ẋ, and ẏ, i.e., dt
where
U ∗ = u − ẋ V ∗ = v − ẏ . (4) ⎡ ⎤ ⎛⎧ ⎫
−1
1 0
0 0 ⎪ 0 ⎪
The flow solution in the Hopf bifurcation methodology is ob- ⎨ ⎬
⎢0 0 12 xα ⎥ ⎜ (2/μs π )C l
1
tained using a parallel multiblock code. A fully implicit steady Rs = ⎣ ⎦ ⎝
0 0
1 0 ⎪
⎩ 0 ⎪
⎭
solution of the Euler equations is obtained by advancing the so-
0 0 12 rα
xα 2
(4/μs π )C m
lution forward in time by solving the discrete nonlinear system of ⎡ ⎤ ⎞
equations 0 −1 0 0
⎢ (2ω R /Ū )2
0 0 0 ⎥ ⎟
−⎣ ⎦ ws ⎠ (8)
wnf +1 − wnf  0 0 0 −1
= R f wnf +1 . (5) 0 0 2rα2
/(Ū )2 0
t
The term on the right-hand side in Eq. (5), that is the residual, and ws = [h, ḣ, α , α̇ ] T . In these expressions, h is the plunge linear
represents the discretization of the convective terms, given here displacement; α is the incidence, or pitch angular displacement;
 E. Camilo et al. / Aerospace Science and Technology 30 (2013) 163–174 165

√
rα = ( I α /m) is the radius of gyration defined in terms of the stability limit of the system at values of μ when A(w0 , μ) has one
pitch moment of inertia, I α , and the airfoil mass per unit span, m; eigenvalue which crosses the imaginary axis.
xα is the offset between the center of mass and the elastic axis; The power method [20] is an algorithm for calculating the
μs = m/πρ∞ b2 is the airfoil-to-fluid mass ratio defined in terms dominant eigenvalue and eigenvector pair of any given diagonal-
of the fluid freestream density, ρ∞ , and the semi-chord, b; ω R = izable matrix, A. Its extension to the shifted inverse power method
ωh
ω is the ratio of the uncoupled natural frequencies of plunging, is practical for finding any eigenvalue, provided that a good ini-
α
ωh , and pitching, ωα ; Ū = U∞
is the reduced velocity defined in tial approximation is known. Therefore, the present approach as-
bωα
sumes that the n × n matrix, A, has distinct eigenvalues, labeled
terms of the fluid freestream velocity, U ∞ ; and C l and C m are the
lift coefficient and the moment coefficient about the elastic axis,
λ1 , λ2 , . . . , λm , and consider the λ j eigenvalue. Then, a constant σ
can be chosen such that 1/(λ j − σ ) is the dominant eigenvalue of
respectively.
Several classes of nonlinear stiffness contributions have been (A − σ I)−1 .
studied in papers treating the open-loop dynamics of the aeroe- Hence, for the A = ∂ R/∂ w matrix, the initial guess, x0P , and
lastic system [16,1]. In this work, torsional polynomials have been constant, σ , the Inverse Power Method (IPM) for a kth-iteration
included in the model to represent nonlinear smooth behavior, that is given by [6]
is
Pk+1 = (A − σ I)−1 xkP , (15)
 
M̄ (α ) = K α f (α ), (9) xkP = P /Pk ∞ .
k
(16)
where K α is considered as a global torsional stiffness (linear). The The shifted inverse power method can be used to calculate the
functional form of f (α ) can be expressed as, critical eigenvalue in the complex plane at a fixed freestream
Mach number. By computing the location for multiple values
f (α ) = f α0 + f α1 α + f α2 α 2 of freestream Mach number, the value at which the eigenvalue

m crosses the imaginary axis can be obtained.
+ f α3 α 3 + f α4 α 4 + · · · + f αm αm = f αi α i . (10)
i =0
5. Torsional polynomial nonlinearity in Jacobian matrix

If the pitching motion is modified to include the polynomial restor- The torsional polynomial nonlinearity is included in the Jaco-
ing moment behavior, then Eq. (8) becomes bian matrix. The calculation of A is most conveniently done by
⎡ ⎤−1 partitioning the matrix as
1 0 0 0  
1 ∂R f ∂R f  
⎢0 1 0 x
2 α ⎥ Aff Afs
Rs = ⎣ ⎦ ∂w f ∂ ws
= . (17)
0 0 1 0 ∂ Rs ∂ Rs Asf Ass
1 2 ∂w f ∂ ws
0 xα 0 r
2 α
⎧⎡ ⎤ The A f f block describes the influence of the fluid unknowns in
⎪ 0
⎨ the fluid residual. The dependence of the fluid residual on the
⎢ (2/μs π )C l ⎥
× ⎣ ⎦ structural unknowns is represented by A f s . The calculation of the
⎪
⎩ 0 
(4/μs π )C m − (2rα2 /(Ū )2 )( f α0 + m i Jacobian matrix terms for a linear structural model is described in
i =2 f α i α )
⎡ ⎤ ⎫ Ref. [7]. The addition of the structural nonlinearity does not change
0 −1 0 0 ⎪
⎬ the Jacobian matrix blocks A f f and A f s in the calculation for the
⎢ (2ω R /Ū )2 0 0 0 ⎥
−⎣ ⎦ ws . (11) linear model.
0 0 0 −1 ⎪
⎭ The Asf block involves calculating the dependence of the fluid
2
0 0 (2rα /(Ū )2 ) f α1 0 unknowns in the structural residual. For the pitch-plunge airfoil,
the fluid variables contribute to the structural equations through
4. Hopf bifurcation analysis
the lift and moment coefficients and, hence, again this term of
the Jacobian matrix remains unchanged. Finally, the only terms,
Consider the semi-discrete form of the coupled CFD–CSD sys- that are different, are those in the exact Ass Jacobian matrix, that
tem, describes the dependence of the structural equations on the struc-
dw tural unknowns leading to
= R(w, μ), (12)
dt ∂ Rs
where μ is a real-valued parameter affecting the residuals (for typ- ∂ ws
ical section, μ = Ū ). Moreover,
⎡ 0 1 0 0
⎤
−4(ω R )2 rα2 2 ∂C 2
xα −2rα
 T ⎢ 0
2rα l − det ( 2 4
+ μπ ∂ Cm
) 0⎥
=⎢ ⎥,
(Ū )2 det detμπ ∂ α (Ū ) ∂α
w = w Tf wsT , (13) ⎣ 0 0 0 1⎦
8xα (ω R )2 −4xα ∂ C l −4rα 2 m
0 − ( i =1 i f αi α i −1 )+ det8μπ ∂∂Cαm 0
is a vector containing the fluid unknowns, w f , and the structural (Ū )2 det detμπ ∂ α det(Ū )2
unknowns, ws , and (18)
 
R = R Tf T T
Rs , (14) where det = rα 2
− x2α .
The Jacobian matrix is obtained for the equilibrium position and
is a vector containing the fluid residual, R f , and the structural part of the eigenspace associated has been calculated to the critical
residual, Rs . For an equilibrium condition of this system, w0 (μ) value. In this case, the shifted IPM has been used to find this criti-
satisfies R(w0 , μ) = 0. cal value within a range of Ū . The cases of evaluating the Jacobian
Dynamical system theory gives criteria for an equilibrium to Matrix with linear and nonlinear structural model can be obtained
be stable [20]. In particular, all eigenvalues of the Jacobian matrix with the CFD solution at α = 0. If the equilibrium condition is
of Eq. (12), given by A = ∂ R/∂ w, must have negative real parts. about zero, the evaluation the Jacobian matrix for both linear and
A Hopf bifurcation with respect to the μ parameter occurs in the nonlinear structural models yields the same result, which can be
166 E. Camilo et al. / Aerospace Science and Technology 30 (2013) 163–174

Table 1
Structural model non-dimensional pa-
rameters (cf. Eq. (11)).

Parameter Value
rα 0.539
xα − 0.2
ωR 0.343
μs 100.0
xea − 0.1
y ea 0.0

observed by looking at the element (4, 3) of the Ass matrix in

Eq. (18).

6. Flutter boundaries and time-marching results

Flutter boundaries have been found with the IPM, considering

a Hopf bifurcation point with respect to the Ū parameter. First, in
this section, flutter boundaries for the NACA 0012 airfoil with a Fig. 1. Flutter boundaries prediction for linear structural dynamics.
linear structure and at zero incidence have been considered. In the
second part of this section, torsional concentrated nonlinearities the system response. The flutter point can be associated to con-
have been added to the aeroelastic system and results are com- stant amplitude oscillations, here observed nearly to Ū = 3.5 (see
pared with previously analyzed linear structural model. The basic Fig. 2(b)). Similarly to the previous simulations, in Figs. 3(a) to 3(c),
non-dimensional parameters for the structural model are given in the identification of the flutter point at M ∞ = 0.82 and Ū = 3.68
Table 1, where the ea subscript denotes the elastic axis. Here, (see Fig. 3(b)) is illustrated.
airfoil dimensions are related to its chord length. Moreover, the Grid refinement evaluation has been performed in order to fully
results have been compared with the time-marching methodology. understand and validate results from both IPM (structured grid)
and time-marching (unstructured grid) approaches. Here, three
6.1. Linear structural model results grid discretizations have been considered for the structured and
unstructured mesh analyses. Data on these meshes are presented
Linear concentrated stiffness has been initially admitted to the in Table 2.
structural dynamics. The only source of nonlinear behavior is, Critical flutter conditions have been calculated using the dif-
therefore, the compressibility effects in transonic regime. Flutter ferent mesh discretizations. The comparison of the various flut-
boundaries achieved with both IPM and time-marching approaches ter points at M ∞ = 0.75, obtained with IPM and time-marching
have been compared. The results of such comparison are depicted methodologies and considering different grids in the calculations,
in Fig. 1, where some discrepancies can be observed between are also shown in Table 2.
the stability boundaries, mainly for Mach numbers near the unity. Steady-state calculations can be used to verify the relative qual-
Those discrepancies can be associated to the fact that the IPM ap- ity of the results for different mesh refinements. The respective
proach cannot deal with fully unsteady effects of shock excursions CFD codes for structured and unstructured meshes have been used
in the transonic regime. The time-marching approach, however, is to calculate pressure distribution around the airfoil for α = 1.25◦
based on direct integration of unsteady transonic aerodynamics and M ∞ = 0.8. Pressure coefficient distributions are presented in
and structural coupled dynamic equations and provides more re- Figs. 4 and 5. Fig. 4 presents, for each code, a comparison of the ef-
alistic instability onset predictions. fect of mesh refinement, whereas Fig. 5 shows, for each grid level,
An estimate of the computational costs for a complete flutter the effect of using the structured or unstructured grid code.
boundary, as presented in Fig. 1, has been obtained. The time- It can be observed that, for a given code, there is some effect
marching methodology requires about 1 month of sequential cal- of mesh refinement on the pressure coefficient distributions and,
culations for the number of flight conditions considered in Fig. 1. hence, in the shock position. However, such differences are not at
In the IPM calculations, the computational costs for the same num- all dramatic. In other words, as far as these results are indicating,
ber of flutter points presented in Fig. 1 is about 12 hours of se- one could even use the respective coarse grids for the calculations
quential calculations. of interest here without too much loss in the quality of the results.
The flutter boundaries obtained with the time-marching scheme A code-to-code comparison, for each grid, as indicated in Fig. 5,
are identified by observing when the aeroelastic responses be- reveals that, for the fine grid, there is essentially no difference in
come unstable. Therefore, in this approach, one integrates the the pressure coefficient distributions provided by the two codes
aerodynamic and structural dynamic equations in time, concur- for this test case. For the coarse grids, however, one can observe
rently, and observes the response of the aeroelastic system. For that the upper surface shock wave predicted by the structured
instance, Figs. 2(a) to 2(c) present the time histories for the aeroe- grid code is more smeared than that given by the unstructured
lastic response in the pitch degree-of-freedom at M ∞ = 0.6 and grid code. This behavior is most certainly associated to the rather
for Ū = 3.0, Ū = 3.5 and Ū = 4.0, respectively. For M ∞ = 0.6, low number of points on the airfoil surface for the former grid.
mild compressibility effects are present and, thereby, this is con- Actually, the number of airfoil surface points for the coarse un-
sidered as the lowest Mach number condition of interest for this structured grid is comparable to the number of surface points in
investigation, as far as nonlinear transonic effects are concerned. the medium structured grid. Hence, in summary, and with the ex-
In these computations, most of the runs do not need many simu- ception of a certain additional smearing of the upper surface shock
lation time, because system oscillations are easily identified with wave by the coarse structured grid, the results are indicating that
damped (converging amplitudes) response (see Fig. 2(a)), before all grids are actually quite adequate for the present purposes.
the flutter point, and diverging amplitudes response, after the flut- For the time-marching approach, which uses unstructured
ter point (see Fig. 2(c)), simply by looking at only a few cycles of mesh, the flutter points are obtained from the analysis of time
 E. Camilo et al. / Aerospace Science and Technology 30 (2013) 163–174 167

Fig. 2. Aeroelastic responses in pitching for M ∞ = 0.6 at different reduced velocities.

Fig. 3. Aeroelastic responses in pitching for M ∞ = 0.82 at different reduced velocities.

Table 2
Grid refinement parameters for both structured and unstructured meshes.

Grid type Grid size Cells/points on airfoil Flutter reduced velocity

surface (M ∞ = 0.75)
Unstructured Coarse 13 744 cells 192 3.55
Medium 17 844 cells 378 3.37
Fine 35 304 cells 642 3.30

Structured Coarse 129 × 33 points 96 3.69

Medium 257 × 65 points 192 3.57
Fine 513 × 129 points 384 3.55

Fig. 4. Comparison of pressure distributions for NACA 0012 airfoil at M ∞ = 0.8 and α = 1.25◦ on respective coarse, medium and fine grids.
168 E. Camilo et al. / Aerospace Science and Technology 30 (2013) 163–174

Fig. 5. Comparison of pressure distributions for NACA 0012 airfoil at M ∞ = 0.8 and α = 1.25◦ for the two different codes.

Fig. 6. Aeroelastic responses for M ∞ = 0.75 – unstructured mesh.

Fig. 7. Aeroelastic responses for M ∞ = 0.75 with fine unstructured grid.

histories of the coupled solution for given reduced velocities. At number of cells from medium to fine grids increased considerably,
fixed M ∞ = 0.75, the aeroelastic response for the coarse grid the flutter point with the fine grid only varied approximately 2%
and Ū = 3.5 is presented in Fig. 6(a). The system exhibits os- with regard to the medium grid result.
cillations with slightly decreasing amplitudes. Such behavior is It is important to note, in these grid refinement studies, the
consistent with the fact that the flutter point has been established computational costs required for the calculations. As one could
at Ū = 3.55 for this case. Fig. 6(b) shows the approximate flutter expect, the computational costs with the fine grid increase con-
boundary with the medium grid at Ū = 3.37. One can observe that siderably. For instance, the time histories calculated up to 800
the increase in the number of cells, from coarse to medium grid, dimensionless time units require 17 hours of CPU time for the
changes the flutter point velocity approximately 3.7%. coarse grid, 22 hours for the medium grid and 43 hours for the
The same analysis with the fine grid shows the flutter point fine grid, in a machine with an Intel Core 2 Duo 2.2 GHz processor
at Ū = 3.3, as indicated in Fig. 7(b). The system exhibits conver- and 2 GB of memory.
gent and divergent amplitude oscillation responses for Ū = 3.2 and From the information in Table 2, and considering that an un-
Ū = 3.4, as shown in Figs. 7(a) and 7(c), respectively. Although the structured triangular grid has approximately three times more cells
 E. Camilo et al. / Aerospace Science and Technology 30 (2013) 163–174 169

order to perform further calculations in this paper, the coarse grid

has been chosen to obtain the flutter boundary results with the
IPM methodology. In principle, if only flutter stability margins were
of interest in the time-marching calculations, the coarse unstruc-
tured grid could also be used for such computations, considering
the previous results. However, some preliminary test cases look-
ing at time histories of the aeroelastic responses, which are not
presented here, have indicated very different post-bifurcation non-
linear behavior using the unstructured coarse grid when compared
to the medium and fine grids responses, even if small differences
are found in the flutter speed. Therefore, in the time-marching
calculations, the medium mesh has been chosen to perform aeroe-
lastic analyses in the remaining test cases to be presented in the
paper.

6.2. Nonlinear structural model results

Flutter boundaries in the case of having torsional polynomial

nonlinearities in the structure have been investigated. Torsional
Fig. 8. Comparison of stability boundaries with cubic polynomial structural nonlin-
polynomial nonlinearities are added according to type and inci-
earity (cf. Eq. (19)).
dence considered in the model. The first structural nonlinearity
introduced to the aeroelastic system is obtained within an inci-
than grid points, one can conclude that the unstructured and struc-
dence interval and adopting 3rd-degree polynomial interpolation
tured coarse grids have a similar number of grid points. On the
as a form of smoothing the restoring torsion motion. It is given by
other hand, the structured medium grid has approximately two
times more grid points than the corresponding unstructured grid.
f (α ) = 0.18α + 18 000α 3 . (19)
Finally, the structured fine grid has approximately five times more
grid points than the unstructured fine grid. The results obtained Fig. 8 depicts flutter boundaries computed for linear and non-
with the corresponding grids, however, are comparable because linear structural models, respectively, for time-marching and IPM
grid refinement for the unstructured meshes is concentrated in methods. Flutter boundaries assessed with a linear torsional restor-
the shock regions, whereas structured grid refinement leads to the ing moment using the IPM approach are represented by the dia-
addition of mesh points throughout the entire computational flow- mond symbols in Fig. 8. Flutter boundaries for the nonlinear struc-
field. tural model given by Eq. (19), also using IPM, are indicated by tri-
Grid refinement studies using the IPM methodology, in the angular symbols in Fig. 8. However, the latter results have been ob-
present work, are associated to structure grid arrangements. As in tained using only the linear term of Eq. (19), that is, f (α ) = 0.18α
the time-marching case, the solutions are calculated for flight con- for the torsional function. The reason is in the fact that there is no
ditions at fixed M ∞ = 0.75. Flutter points have been obtained with effect of torsional nonlinear terms in the flutter speed calculations
the coarse grid at Ū = 3.69, with the medium grid at Ū = 3.57, with Hopf bifurcation analysis at α = 0.
and with the fine grid at Ū = 3.55. The difference between the Then, time-marching calculations have been performed with
reduced velocities at the flutter points obtained with coarse and the cubic structural nonlinearity (cf. Eq. (19)). Fig. 9(a) illustrates
medium grids is 3.2%, and this difference is about 0.61% for the convergent pitch response for M ∞ = 0.75 and Ū = 1.2. On the
solutions with medium and fine grids. other hand, Fig. 9(b) shows LCO at Ū = 1.4 for the same freestream
In the IPM code, a steady state CFD calculation is required in Mach number. The results indicate that the flutter point is at
the equilibrium point, before IPM iterations. The computational about Ū = 1.3 for this Mach number. Increasing the reduced ve-
costs of a steady state solution in the implicit CFD code used in locity to Ū = 1.7, the system maintains the LCO response, but at
the IPM methodology is about 32 seconds for the coarse grid, 2 a higher amplitude of motion, as indicated in Fig. 9(c). A compar-
minutes and 10 seconds for the medium grid, and 8 minutes and ison of these results with only the linear term of the polynomial,
30 seconds for the fine grid. All calculations have been performed i.e., f (α ) = 0.18α , can be obtained by observing the linear results,
in an Intel Core 2 Duo 2.2 GHz processor with 2 GB of memory. In which are shown in Fig. 10(a) for M ∞ = 0.75 and Ū = 1.2, and

Fig. 9. Aeroelastic response for cubic nonlinearity, f (α ) = 0.18α + 18 000α 3 , at M ∞ = 0.75 at different reduced velocities.
170 E. Camilo et al. / Aerospace Science and Technology 30 (2013) 163–174

Fig. 10. Aeroelastic time histories for the linear term of the cubic polynomial, f (α ) = 0.18α , at M ∞ = 0.75 at different reduced velocities.

Fig. 11. Structural nonlinearity representations: 7th-, 11th-, and 13th-degree polynomials.

in Fig. 10(b) for M ∞ = 0.75 and Ū = 1.4. One can see that, for
this Mach number, the results indicated that the calculations are
convergent for Ū = 1.2, and the flutter point is at about the same
value of Ū = 1.3 for both linear and nonlinear structures. As can
be expected, these give the same flutter speed, but the post-flutter
behavior may be different due to the nonlinear terms. On the other
hand, it can be observed from Eq. (19) that the linear term re-
veals the structure is less stiff in torsion, which leads to lower
flutter speeds compared with previous results for linear torsion,
i.e., f (α ) = α , as clearly shown in Fig. 8.
Based on the observation that the polynomial nonlinear terms
do not influence flutter stability boundary calculations, but do in-
fluence the post-flutter behavior, it seems reasonable to investigate
the effect of different nonlinear terms with increasing order. These
are obtained to study the influence of the curve gradient at the
equilibrium position, i.e., the linear term value, in the flutter speed.
Increasing polynomial order can be related to stronger nonlinear-
ity. Three polynomials have been adopted, 7th-degree, 11th-degree,
and 13th-degree polynomials, as shown in Figs. 11(a) and 11(b).
The first structural nonlinearity is a 7th-degree polynomial, de-
Fig. 12. Stability boundaries with the P7 and P11 polynomial structural nonlineari-
noted here as P7, that is,
ties.
 
f (α ) = 3.6728 × 10−1 α − 2.8389 × 10−12 α 2 The second structural nonlinearity is an 11th-degree polynomial,
 6 3
 −5 4 denoted as P11, given by
+ 3.4407 × 10 α + 1.2850 × 10 α
   
− 8.1911 × 1012 α 5 − 17.695α 6 f (α ) = 1.3076 × 10−1 α − 2.3217 × 10−11 α 2
  
+ 6.4405 × 1018 α 7 . (20) + 1.0067 × 107 α 3 + 3.3485 × 10−4 α 4
 E. Camilo et al. / Aerospace Science and Technology 30 (2013) 163–174 171

Fig. 13. Aeroelastic responses for P11 nonlinearity at M ∞ = 0.75 and different reduced velocities.

 
− 6.0815 × 1013 α 5 − 1.6774 × 103 α 6 pitch degree-of-freedom. In Fig. 13(a), the time history results
  for M ∞ = 0.75 and Ū = 1.0 are presented. For this flight condi-
+ 1.7859 × 1020 7
α + 3.3998 × 10 α 9 8
  tion, the system exhibits a convergent response. For M ∞ = 0.75
− 2.4809 × 1026 α 9 − 2.4809 × 1015 α 10 and Ū = 1.2, as shown in Fig. 13(b), one obtains an LCO-type
 response, which is consistent with the results obtained, for this
+ 1.3070 × 1032 α 11 . (21)
test case, with the IPM code as indicated in Fig. 12. The LCO
Finally, the third structural nonlinearity is a 13th-degree polyno- is, however, of fairly low amplitude. The system again shows
mial, denoted as P13, LCO response, at comparatively larger amplitudes, when the re-
  duced speed is increased to Ū = 1.5 and Ū = 2.0. The results are
f (α ) = 3.5799 × 10−2 α + 7.9409 × 10−10 α 2 shown in Figs. 13(c) and 13(d), respectively. In the inverse power
 
+ 1.4512 × 107 α 3 − 1.7529 × 10−2 α 4 method calculations, the flutter point for M ∞ = 0.75 and zero in-
  cidence, and considering the P11 structural nonlinearity, occurs at
− 1.1979 × 1014 α 5 + 1.3849 × 105 α 6 Ū = 1.19.
 
+ 5.1166 × 1020 α 7 − 4.8616 × 1011 α 8 A verification of the IPM approach was also performed, for a
  very coarse structured grid with 33 × 9 points, by computing the
− 1.1577 × 1027 α 9 + 7.7577 × 1017 α 10 complete eigenvalue spectrum of the A matrix (see Eq. (17)). This
 
+ 1.3179 × 1033 α 11 − 4.5889 × 1023 α 12 very coarse grid is not indicated in Table 2, because this is a com-
 putational grid even coarser than the so-called coarse grid referred
− 5.9384 × 1038 α 13 . (22) to in the table and it was used solely for the present verification
Fig. 12 shows the flutter boundaries with the P7 and P11 test. The reason for such an approach is a very practical one, and
polynomial nonlinearities compared with the linear structure re- it is related to the fact that the eigenvalues were computed using
sults, f (α ) = α . In this case, the time-marching calculations have Matlab and there was a limit for the size of the matrix that we
been performed with initial incidence equal to 0.5◦ . It can be ob- could input in Matlab to perform this computation. The results of
served from the coefficients of the linear terms of P7 and P11 this calculation are shown in Fig. 14(a), for a range of values of Ū .
that the values are less than unity. Again, the torsion springs in A close-up view of the eigenspectrum for Ū values ranging from
these cases are less stiff than that of the linear test case and, 0.6 to 1.5 is plotted in Fig. 14(b). The arrows indicate increasing
hence, yield lower flutter speeds. Moreover, the linear term of P11, values of Ū . The critical eigenvalue crosses the imaginary axis at
f (α ) = 1.3076 × 10−1 α , has a smaller coefficient than the linear Ū = 1.2, which is similar to the values found from the previous
term of P7, f (α ) = 3.6728 × 10−1 α , and therefore flutter boundary IPM and time-marching analyses.
results of P11 indicate lower flutter speeds. The next case considers the flutter boundary for the P13 poly-
Aeroelastic response calculations, using the P11 polynomial rep- nomial nonlinearity. In Fig. 15, the results are compared with those
resentation of the torsion spring, are presented in Fig. 13 for the obtained with a linear structure. For the range of M ∞ = 0.5 up to
172 E. Camilo et al. / Aerospace Science and Technology 30 (2013) 163–174

Fig. 14. Eigenvalue spectrum of the A matrix with the P11 polynomial structural nonlinearity for M ∞ = 0.75 and a range of Ū values.

Fig. 16. Close-up view of the eigenvalue spectrum with the P13 polynomial struc-
Fig. 15. Comparison of the stability boundaries for the case with the P13 polynomial tural nonlinearity (M ∞ = 0.75 and Ū ranging from 1.0 to 4.35).
structural nonlinearity.

crossing occurs for Ū = 1.3. Such behavior does not occur in the
0.8, the flutter boundaries have presented a strange behavior in the P11 polynomial structural nonlinearity case, as shown in Fig. 14(b).
sense that the linear structure is yielding a lower flutter speed. In Hence, time-marching analyses have also been performed for
order to verify such results, both the system matrix eigenspectrum this test case. Time history results for the P13 polynomial struc-
and time-marching analysis calculations have been performed. tural nonlinearity, for M ∞ = 0.75 and Ū = 1.1, are shown in
The first aspect that can be verified is related to the fact that Fig. 17(a). It is observed from these responses that the system
the flutter point, obtained from the IPM approach and using the exhibits initial transient oscillations, followed by non-oscillatory
coarse grid from Table 2, for M ∞ = 0.75, occurs at approximately equilibrium point different from zero. This is reasonable admitting
Ū = 3.4. Further calculations were also performed with a com- the higher nonlinear order of P13 and the effect of sudden increase
putational grid which is even coarser than the so-called coarse in stiffness determined by the representation of structural nonlin-
earity. Fig. 17(b) shows the time-marching results with an increase
grid from Table 2. This structured grid had 96 × 24 points, and
in the reduced velocity to Ū = 1.2. It can be observed that the
the flutter point, for the same freestream Mach number, occurs
system, initially, exhibits irregular oscillations and, afterwards, fol-
at Ū = 4.29. Therefore, the system matrix eigenspectrum, for this
lows somewhat regular oscillations with constant amplitude. The
even coarser grid, is plotted up to Ū = 4.35 in Fig. 16, in order to
system behavior for Ū = 1.5 is more complex than in the Ū = 1.2
check if any eigenvalues would cross the imaginary axis. One can case, as observed in the time history results in Fig. 17(c).
observe from Fig. 16 that an eigenvalue pair crosses the imaginary From these results, therefore, in the P13 case, the flutter point
axis towards the very end of the reduced velocity range consid- for M ∞ = 0.75 can be estimate with the time-marching analysis at
ered in the present case. Moreover, one must also verify whether Ū = 1.2. IPM calculations do not estimate this value for the flut-
no other (isolated) eigenvalue is in the positive side of the real ter point at this Mach number, as shown in Fig. 15. It is important
axis. Fig. 16 shows that the eigenvalues along the real axis cross to note that Hopf bifurcation calculations are basically concerned
to the positive side. Actually, the calculations indicate that such with finding the point at which the system starts to exhibit limit
 E. Camilo et al. / Aerospace Science and Technology 30 (2013) 163–174 173

Fig. 17. Aeroelastic responses for the P13 nonlinearity for M ∞ = 0.75 and different reduced velocities.

Fig. 18. Time-histories for the linear term of the P13 polynomial for M ∞ = 0.75 different Ū .

cycle oscillations at an equilibrium position of the system. Some P13 polynomial can be considered approximately at Ū = 1.4. This
basic problems can be found in the Hopf bifurcation analyses for is quite similar to the value of reduced velocity for the stability
the P13 nonlinear case. One such problem can be associated with boundary, Ū = 1.2, obtained for the complete P13 polynomial tor-
the equilibrium position used for the Hopf bifurcation analysis, sional spring. It also seems to be consistent with the crossing of
that must be further verified in the P13 nonlinear case. As one the eigenvalues along the real axis, as shown in Fig. 16, that oc-
can observe in Fig. 17(a), the system seems to stabilize in an equi- curs for Ū = 1.3. Hence, it seems that the strong nonlinearity in
librium point different from zero and the equilibrium position is this case yields more dynamics to the system and allows equi-
assumed to be zero in the present IPM calculations. librium points different from zero incidence, despite the apparent
Hence, with the initial equilibrium position at α = 0, the results symmetrical flow conditions. One should further realize that, in the
in Fig. 16 are showing that there is also an isolated real eigen- unstructured grid calculations, the flow solution is never absolutely
value which is crossing the imaginary axis as Ū increases, besides symmetric due to the nature of the unstructured grids, which are
the complex eigenvalue pair. Such zero frequency mode eigenvalue not completely symmetric. Therefore, additional Hopf bifurcation
indicates divergence at α = 0. Furthermore, one can also check studies assuming different equilibrium points might be required in
in Fig. 16 that, for smaller values of Ū , the complex conjugate order to fully explain the differences between these computations
eigenvalue pair is in the left-hand plane and, therefore, oscillations and the original IPM results, shown in Fig. 15. However, this de-
must be damped, as confirmed by the time marching response in feats the purposes of the approach in the sense that, if one has to
Fig. 17(a). It is interesting, though, that system oscillatory response perform unsteady coupled calculations in order to find the equilib-
has stabilized at an equilibrium position different from zero, de- rium point prior to the IPM analyses, it is better to use the fully
spite the fact that the airfoil is symmetric and the initial conditions nonlinear time-integration analysis to define the flutter point for
considered α = 0. In any event, this is a perfectly possible result such highly nonlinear cases.
considering the strong nonlinearity of the P13 polynomial. Finally, it must be reported that other types of strong poly-
For comparison purposes, time-marching calculations consider- nomial nonlinearities have been evaluated. The overall behavior,
ing only the linear term of the P13 polynomial, f (α ) = 3.5799 × obtained in these additional cases, is similar to that observed with
10−2 α , have also been performed. For M ∞ = 0.75, at Ū = 1.1 and the P13 polynomial spring. Such behavior seems to be associ-
Ū = 1.2, time histories are shown in Figs. 18(a) and 18(b), respec- ated with the relatively small coefficient of the polynomial linear
tively. The response initially increases and, then, goes to a non-zero term. Furthermore, for all time-marching calculations performed
equilibrium point. The system behavior for Ū = 1.5 is indicated in with such a small coefficient for the linear term of the structural
Fig. 18(c). The system exhibits oscillations around a non-zero equi- model, the system exhibits oscillations about a non-zero equilib-
librium point. Hence, the flutter point for the linear term of the rium point. An appropriate method to calculate the equilibrium
174 E. Camilo et al. / Aerospace Science and Technology 30 (2013) 163–174

points, before Hopf bifurcation analyses, is necessary in order to Acknowledgements

continue the study of flutter boundaries with the IPM methodol-
ogy in these cases. Other test cases have also been analyzed in The authors acknowledge the financial support provided by
the context of the present research, but they are not shown here Conselho Nacional de Desenvolvimento Científico e Tecnológico,
for the sake of brevity. They have indicated that flutter boundaries CNPq, through the Research Grants Nos. 312064/2006-3 and
for zero incidence, calculated considering complete nonlinear poly- 471592/2011-0. Additional support to this research was provided
nomials, are equal to those obtained with only the linear term of by Fundação de Amparo à Pesquisa do Estado de São Paulo, FAPESP,
the respective polynomials. In other words, the nonlinearity in the through Grants Nos. 2004/16064-9 and 2013/07375-0. The au-
torsional degree of freedom does not have an effect on the flut- thors gratefully acknowledge that Prof. Ken Badcock and Dr. M.
ter onset, at least for the polynomial nonlinearities analyzed in the Woodgate kindly provided the Hopf bifurcation code used in the
present work. present research during the tenure of Dr. Elizangela Camilo at Liv-
erpool University, UK.
7. Concluding remarks

Concentrated nonlinearities are shown to have significant ef- References

fects on the aeroelastic responses and to yield limit cycle oscil-
[1] A. Abdelkefi, R. Vasconcellos, F.D. Marques, M.R. Hajj, Bifurcation analysis
lations below the linear flutter speed. For the computations per-
of an aeroelastic system with concentrated nonlinearities, Nonlinear Dynam-
formed here, direct Hopf bifurcation predictions at α = 0, for the ics 69 (1–2) (2012) 57–70.
pitch-plunge typical section airfoil with torsional nonlinearities, in- [2] J.J. Alonso, A. Jameson, Fully-implicit time-marching aeroelastic solutions, in:
dicate that the nonlinear terms do not influence the bifurcation 32nd AIAA Aerospace Sciences Meeting and Exhibit, Reno, NV, Jan. 1994.
point. This is also somewhat evident by looking at the (4, 3) el- [3] J.L.F. Azevedo, L.C. Oliveira, Unsteady airfoil flow simulations using the Euler
equations, AIAA Paper No. 94-1892 in: Proceedings of the 12th AIAA Applied
ement of the Ass matrix. Time-marching analyses are compared
Aerodynamics Conference, Part 2, Colorado Spring, CO, June 1994, pp. 650–660.
to direct calculations of the Hopf bifurcation points and the re- [4] K.J. Badcock, M.A. Woodgate, Prediction of bifurcation onset of large order
sults also indicate that structural nonlinear terms have no effect on aeroelastic models, AIAA Paper No. 2008-1820, 2008.
flutter speed calculations. The results agree well and the computa- [5] K.J. Badcock, M.A. Woodgate, Bifurcation prediction of large-order aeroelastic
models, AIAA Journal 48 (6) (2010) 1037–1046.
tional tools have show to be powerful enough to analyze nonlinear
[6] K.J. Badcock, M.A. Woodgate, P.S. Beran, Fast prediction of wing rock limit cycle
effects in aeroelasticity. However, further analyses are still neces- oscillations based on computational fluid dynamics, Reno, NV, AIAA Paper No.
sary in order to characterize the complete behavior of such nonlin- 2007-1071, Jan. 2007.
ear systems, even for the typical section model. Post-bifurcations [7] K.J. Badcock, M.A. Woodgate, B.E. Richards, Hopf bifurcation calculations for
calculations, on the other hand, show the influence of nonlinear symmetric airfoil in transonic flow, AIAA Journal 42 (5) (May 2004) 883–892.
[8] K.J. Badcock, M.A. Woodgate, B.E. Richards, Direct aeroelastic bifurcation anal-
structural terms on LCO. In the same flight conditions, the systems
ysis of a symmetric wing based on Euler equations, Journal of Aircraft 42 (3)
with nonlinear structural models exhibit oscillations with finite (May–June 2005).
amplitudes, while those with the corresponding linear structural [9] F. Cantariti, L. Dubuc, B. Gribben, M. Woodgate, K. Badcock, B. Richards, Ap-
model exhibit divergent oscillatory responses. proximate Jacobians for the solution of the Euler and Navier–Stokes equations,
Grid refinement studies indicate good correlation between Aerospace Engineering Report No. 9704, University of Glasgow, 1997.
[10] E.H. Dowell, J. Edwards, T. Strganac, Nonlinear aeroelasticity, Journal of Air-
time-marching and IPM solution procedures. The rise of the com- craft 40 (5) (2003) 857–874.
putational costs, especially with finer grids, indicates the advantage [11] E.H. Dowell, D. Tang, Nonlinear aeroelasticity and unsteady aerodynamics, AIAA
of Hopf bifurcation analysis for the flutter boundary calculations. Journal 40 (9) (2002) 1697–1707.
In this fashion, the problem of locating a one-parameter stabil- [12] L. Dubuc, F. Cantarini, M.A. Woodgate, B. Gribben, K.J. Badcock, B.E. Richards,
Solution of the Euler unsteady equation using deforming grids, Department of
ity boundary is reduced from multiple, unsteady, time-marching
Aerospace Engineering Report 9705, University of Glasgow, 1997.
calculations to a single steady-state calculation. Hence, the expe- [13] A. Jameson, Time dependent calculations using multigrid, with applications to
rience has shown that it is possible to reduce the computational unsteady flows past airfoils and wings, AIAA Paper No. 91-1596, 1991.
costs from one month of sequential calculations with the time- [14] A. Jameson, W. Schmidt, E. Turkel, Numerical simulation of the Euler equations
marching code to some hours of sequential calculations with the by finite volume methods using Runge–Kutta time stepping schemes, AIAA Pa-
per No. 81-1259, 1981.
IPM code for the complete flutter boundary results for a given [15] K.A. Kousen, O.O. Bendiksen, Limit cycle phenomena in computational tran-
flight condition. sonic aeroelasticity, Journal of Aircraft 32 (2) (1994) 1257–1263.
Strong nonlinearity results, however, have demonstrated that [16] B.H.K. Lee, S.J. Price, Y.S. Wong, Nonlinear aeroelastic analysis of airfoils: Bifur-
the Hopf bifurcation analysis could not find the complete flut- cation and chaos, Progress in Aerospace Sciences 35 (1999) 205–334.
[17] S.A. Morton, P.S. Beran, Hopf-bifurcation analysis of airfoil flutter at transonic
ter boundary for all cases, as evidenced in some of the present
speeds, Journal of Aircraft 36 (2) (March–April 1999).
calculations for the 13th-order polynomial structural nonlinear- [18] L.C. Oliveira, A state-space aeroelastic analysis methodology using compu-
ity. The difficulties observed seem to be associated with different tational aerodynamics techniques, Master Thesis, Instituto Tecnológico de
equilibrium positions of the aeroelastic system. In other words, Aeronáutica, São José dos Campos, SP, Brazil, 1993 (in Portuguese).
the time-marching calculations have shown non-zero equilibrium [19] S. Osher, S.R. Chakravarthy, Upwind schemes and boundary conditions with ap-
plications to Euler equations in general coordinates, Journal of Computational
points when there is a combination of relatively small coefficients Physics 50 (1983) 447–481.
of the polynomial linear term and an otherwise strongly nonlinear [20] R. Seydel, Practical Bifurcation Analysis and Stability Analysis, 2nd edition,
representation of the torsional spring. Further work is necessary Springer-Verlag, New York, 1994.
in order to provide a more detailed comparison of time domain [21] C.F.C. Simões, J.L.F. Azevedo, The influence of numerical parameters on un-
steady airfoil inviscid flow simulations using unstructured dynamic meshes, in:
predictions and direct Hopf bifurcation results for complete flut-
Proceedings of the 15th Brazilian Congress of Mechanical Engineering – COBEM
ter boundaries with nonlinear structural models. Variations of the 99, ABCM, Águas de Lindóia, SP, Brazil, Nov. 1999.
mean, steady state equilibrium conditions, from which computa- [22] M.A. Woodgate, K.J. Babcock, Fast prediction of transonic aeroelastic stability
tions are performed, are also needed. and limit cycles, AIAA Journal 45 (6) (2007) 1370–1381.