11th World Congress on Computational Mechanics (WCCM XI)

5th European Conference on Computational Mechanics (ECCM V)

6th European Conference on Computational Fluid Dynamics (ECFD VI)
E. Oñate, J. Oliver and A. Huerta (Eds)

ON THE DEVELOPMENT OF A HARMONIC BALANCE

METHOD FOR AEROELASTIC ANALYSIS

Graham Ashcroft, Christian Frey, Hans-Peter Kersken∗

German Aerospace Center (DLR)
Institute of Propulsion Technology
Linder Höhe, 51147 Cologne, Germany
email: [email protected], [email protected], [email protected]

Key words: Harmonic Balance Method, Flutter, Aeroelasticity, Turbomachinery

Abstract. To efficiently simulate time-periodic, non-linear flows in turbomachinery

a Harmonic Balance (HB) method has recently been developed within the framework
of DLR’s compressible, Unsteady Reynolds Averaged Navier-Stokes (URANS) solver
TRACE. As a hybrid time- and frequency-domain method, that solves directly for the
complex valued solution harmonics of the URANS equations, the approach allows both
the integration of highly accurate non-reflecting boundary conditions and the efficient
resolution of non-linear flow phenomena. In the current work the HB-solver is extended
and applied to investigate the aeroelastic problem of flutter in turbomachinery. Within
this context the HB-method is developed, under consideration of the Geometric Con-
servation Law (GCL), to support deforming meshes. To validate the approach the well
documented aeroelastic test case Standard Configuration 10 is simulated and results are
compared to those obtained with an inhouse time-linearized solver and reference data
from the literature.

1 INTRODUCTION
Aeroelastic analysis is an integral aspect of the design of compressor and turbine com-
ponents. To investigate such complex phenomena efficient numerical methods capable of
accurately capturing the underlying unsteady interaction between the blades and aerody-
namic flow field are required. As the physical phenomena of flutter and forced-response
are periodic in nature the associated numerical simulations are typically formulated in the
frequency domain. Due to limited computation resources the earliest work on frequency
domain solvers for aeroelasticity investigations was based on the two-dimensional Euler
equations [7]. With the advent of ever more powerful computer resources the underlying
models have been extended to the three-dimensional Euler equations and more recently
to the three-dimensional compressible Navier-Stokes equations.

1
 Graham Ashcroft, Christian Frey, Hans-Peter Kersken

To facilitate the transformation of the governing equations to the frequency domain all
of these approach are based on the linearization of the governing equations. This decouples
the governing equations for the flow perturbations from those governing the mean flow
field and enables highly efficient solvers to be developed. Whilst highly efficient this
approach negates the application of such tools to problems in which nonlinear phenomena
are suspected to play an imporant role.
To overcome these restrictions and allow frequency-domain methods to be applied to
nonlinear, time-periodic problems considerable effort has been expended in recent years
in the development of harmonic balance (HB) methods.
In this work the Harmonic Balance method implemented in the CFD solver TRACE [4]
is extended to investigate the problems of flutter and forced-response in turbomachinery.
The paper is organised as follows. In Section 2 the basic numerical approach is outlined
in terms of the underlying CFD solver and the basic Harmonic Balance method. The
extension of the HB-solver to support deforming meshes is then described in Section 2.2.2.
The validation of the HB-solver for deforming meshes is then presented in Section 3 using
a simple channel flow with a deforming mesh. In Section 4 the HB-method is then applied
to the benchmark aeroelasticity problem Standard Configuration 10 and the ability of the
approach to capture nonlinear effects in the unsteady flow is investigated. Finally, in
Section 5 a summary of the conclusions of the work is given.

2 NUMERICAL METHOD
2.1 Underlying Flow Solver
The harmonic balance method described in this work is part of the CFD code TRACE.
TRACE [1] is a parallel Navier-Stokes flow solver for structured and unstructured grids
that has been developed at DLR’s Institute of Propulsion Technology in Cologne to model
and investigate turbomachinery flows. The code solves the finite-volume discretization
of the compressible Reynolds-averaged Navier-Stokes (RANS) equations in the relative
frame of reference using a multi-block approach. Inviscid fluxes are evaluated using Roe’s
flux-differencing-splitting method. The upwind states are computed using the MUSCL
family of schemes in conjuction with a modified van Albada limiter to avoid unphysical
oscillations in the vicinity of shocks. Viscous terms are discretized using second-order
accurate central differences. To model the effects of turbulence various turbulence and
transition models have been integrated in the code [2]. For the present work it is sufficient
to note that following the discretization of the spatial operators in the Navier-Stokes
equations the following system of ordinary differential equations (ODEs) is obtained
dq
+ R (q(t)) = 0 (1)
dt
where q is the vector of conservative variables, R is the discretized RANS residual vector
and t denotes the physical time.

2
 Graham Ashcroft, Christian Frey, Hans-Peter Kersken

2.2 Harmonic Balance Method

In the current work we are interested in time-periodic solutions of Eq. 1. Often such
problems can be accurately described by a limited number of solution harmonics, e.g.
" K #
X
q(x, t) = Re q̂k (x)eikωt (2)
k=0
where q̂k are the complex valued solution harmonics and ω = 2πf is the fundamental
angular frequency. In such cases it is attractive to formulate the unsteady problem,
Eq. 1, in the frequency domain, i.e. to consider

[ = 0,
ikω q̂k + R(q) (3)
k
for only a finite number of harmonics, k = 0, . . . , K. For configurations in which nonlin-
earity can be assumed to be negligble the coupling between the harmonics of q, imposed by
the nonlinear nature of R, can be neglected and one obtains K + 1 independent equations
for the solution harmonics and the time-mean solution field. However, if one wishes to re-
tain nonlinear effects an alternative approach is required in the modelling or computation
[ . In this work we compute R(q)
of R(q) [ as F (R(F −1 q̂))|k and therefore solve
k k

ikω q̂k + F (R(F −1 q̂))|k = 0, (4)

where F denotes the Discrete Fourier Transform (DFT).
Eqn. 4 is solved directly in the frequency domain to obtain the complex valued har-
monics of the conservative variables q̂k . To compute the harmonics of the RANS residual
vector R the solution field vector is first reconstructed at N sampling points within the
period of oscillation from the Fourier coefficients of the conservative variables q̂k using
an inverse DFT. Using the reconstructed solution vectors the RANS residual vectors R
are then computed at the N sampling points to enable the DFT of R to be computed.
Note, since the RANS residual R is evaluated in the time-domain the standard flux and
discretization schemes from the underlying nonlinear solver can be used without the need
to explicitly model the coupling terms between the solution harmonics.
As a hybrid time- and frequency-domain method the approach has the advantage, over
methods formulated purely in the time- or frequency-domain, of being able to employ
not only the possibly highly nonlinear time-domain flux functions (and their stabilizing
numerical limiters) but also highly accurate nonreflecting boundary conditions formulated
in the frequency domain [5]. Particularly in the context of aeroelasticity or aeroacoustics
should boundary conditions are of the utmost importance.

2.2.1 Harmonic Sets

The application and solution of Eqn. 4 is straightforward if the problem at hand can be
described by harmonics of a single fundamental frequency. On the otherhand, if multiple

3
 Graham Ashcroft, Christian Frey, Hans-Peter Kersken

fundamental frequencies are present, e.g. blade row interaction in a multistage compressor
or turbine, the application of Eqn. 4 is more involved. If the fundamental frequencies do
not share a common base frequency the DFT used in Eqn. 4 is essentially approximative
and great care must be taken in selecting the sampling times and sample duration if the
accuracy and stability of the approach is to be significantly reduced [6]. In certain cases
the fundamental frequencies may be integer multiples (harmonics) of a common third
fundamental frequency. However, if the difference in frequency between the original two
fundamental frequencies ∆f is small then the common fundamental frequency will be
equally low (f = ∆f ) and the harmonics of interest high. This reduces the efficiency of
the approach significantly as the number of sampling points required is proportional to
the order of the harmonics to be resolved.
To overcome these difficulties the concept of harmonic sets Si is introduced [4]. A
harmonic set is defined by a fundamental angular frequency (ω) and interblade phase
angle (σ), fundamental mode, and higher harmonics, e.g.

S = {(k1 ω, k1 σ), (k2 ω, k2 σ), . . .} . (5)

For each fundamental mode a harmonic set is defined. As a given mode may exist
simultaneously in multiple harmonic sets, e.g. the zeroth harmonic or time-mean solution,
the harmonic sets can not be treated entirely independently without the duplicate modes
being ill-specified. To address this deficiency the RANS residual is extended to couple the
harmonic sets via the duplicate modes and approximated (in the case of two harmonic
sets) by

R(qS1 ∪S2 (t)) ≈ R(qS1 (t)) + R(qS2 (t)) − R(qS1 ∩S2 (t)), (6)
allowing the harmonics of the residual to be computed as the sum of the harmonics of
the residuals on the right-hand side of Eqn. (6). The main advantage of this approach is
that the sampling points for each of the harmonic sets S1 , S2 and S1 ∩ S2 can be chosen
independently. Applying the specific DFT to each of the three summands then yields the
harmonics of the residual. For more details see [4].

2.2.2 Time-dependent Meshes

To simulate the problem of flutter it is necessary to be able accomodate time-dependent
boundaries and therfore meshes. For such problems the RANS residual depends not only
on the time-dependent solution vector q but also on the grid coordinates x and the grid
velocities ẋ, i.e.

R = R(q, x, ẋ). (7)

Furthermore, in constrast to Eqn. 1 the cell volume is time-dependent and must be re-
tained in the temporal derivative term. The governing equations then read

4
 Graham Ashcroft, Christian Frey, Hans-Peter Kersken

∂
(V q) + R (q(t), x(t), ẋ(t)) = 0 (8)
∂t
where V is the, time-dependent, volume of a given cell. Since for flutter problems the
blade surface motion is time-periodic the mesh displacement can be discribed by

δx(x, t) = Re x̂(x)eiωt
 
(9)
where x̂ is the complex valued displacement vector. Using Eqn. 9 the HB-solver can be
extended for periodically deforming meshes by reconstructing the mesh alongside the flow
field at the N sampling points. Furthermore, Eqn. 9 is used to compute the grid velocities
ẋ. With the mesh geometry at the N sampling points known all required metrics as well
as the Jacobian J = 1/V can be calculated and used in the evaluation of the RANS
residual R and the computation of its harmonics.
In contrast to the situation with stationary meshes care must also be taken when eval-
uating the temporal harmonic of the temporal derivative term in Eqn. 8. For stationary
meshes this is simply ikω q̂Jk . However for deforming meshes we require
q
\
ikω . (10)
J k
That is, the harmonics of the volume weighted conservative flow variables are required.
As we solve for the harmonics of the conservative flow variables these data are not directly
available. One approximate, but computationally efficient, approach to compute the term
in Eqn. 10 is to employ its linearized form and compute
J − J0
ikω q̂k , (11)
J02

where J0 is the Jacobian of the undisturbed mesh. This is the approach adopted in the
linear solver linearTRACE [8]. Whilst suitable for small amplitude mesh displacement,
and therefore a linear solver, the approach was found to produce large amplitude spuri-
ous pressure waves at large amplitudes. To overcome this limitation the time-derivative
term is computed by computing directly the harmonics of the weighted conservative flow
variables from the reconstructed mesh and flow time-histories. Although somewhat more
computationally expensive this approach ensures the time-dependent mesh displacement
does not corrupt the solution at large amplitude disturbances.

2.2.3 Solution Techniques

To solve the governing equations an implicit pseudo-time approach is adopted. Dis-
cretizing the pseudo-time operator using the first-order Euler backward method and then
linearizing the harmonic balance residual RHB in pseudo-time the following system of
equations is obtained

5
 Graham Ashcroft, Christian Frey, Hans-Peter Kersken

"   #
1 1 ∂R
∆q̂ (m) = −RkHB q̂ (m) ,


+ ikω + (12)
J0 ∆τ ∂q (m)
q̂0

where q
\
RkHB q̂ (m) = ikω + F (R(F −1 q̂))|k , ∆q̂ (m) = q̂ (m+1) − q̂ (m)


(13)
J k
and m is the pseudo-time iteration counter. Note, the coupling terms have been neglected
(m)
in the left-hand side and therefore it depends directly only on the time-mean solution q̂0 .
As such the residual Jacobian ∂R ∂q
is identical to that employed in the steady or unsteady
nonlinear flow solver. The linear system of equations, Eqn. 12, is solved using either the
incomplete lower upper (ILU) or successive overrelaxation (SSOR) methods.

3 VALIDATION
To quantify the errors induced by the adopted approximation of the Geometric Conser-
vation Law (GCL) the impact of time-dependent mesh deformation on a uniform subsonic
flow in a three-dimensional channel is investigated. The computation mesh and domain
used in this work are shown in Fig. 1. The channel has a radial height of ∆r = 0.1 m,
a pitch ∆φ = 2π 64
, an outer radius ro = 1 m and an axial length lx = 1 m. The compu-
tational domain, with a total of 2 × 104 cells, is resolved with 200, 20 and 5 cells in the
axial, radial and circumferential directions, respectively. In simulating the problem the
nonlinear Euler equations are solved and a slip boundary condition is applied along all
solid wall boundaries. At the inflow boundary the flow is purely axial with the uniform
values of stagnation pressure and temperature of 101325 Pa and 288.15 K, respectively.
At the outflow boundary a uniform value of static pressure of 85400 Pa is prescribed to
give a flow with Mach number M = 0.5.
Time-dependent mesh motion is induced by prescribing axial displacement vectors on
the inner and outer radial boundaries. The displacement vectors are given by
(   
α π(x−x0 )
1 + cos , if |x − x0 | ≤ ∆,
x̃ = 2 ∆
(14)
0, otherwise,
where ∆ is the axial width of the disturbance, x0 is its center and α ∈ [1, 0.5, 0.1] is the
disturbance amplitude. The three-dimensional grid deformation is then computed using
an elliptic mesh deformation algorithm [10]. The resulting distribution of the relative
volume change (∆V /V ) is shown in Fig. 1 for the amplitude α = 1. It can be seen
that this corresponds to a maximum relative volume change of approximately 30%. As
a disturbance vectors prescribed on the inner and outer radial boundaries have only
components in the axial direction the relative volume change varies linearly with the
amplitude of the disturbance vectors. E.g. for α = 0.5 and α = 0.1 the maximum relative
volume changes are approaximately 15% and 3%, respectively.

6
 Graham Ashcroft, Christian Frey, Hans-Peter Kersken

V|max
V -0.3 -0.2 -0.1 0 0.1 0.2 0.3
Y

Z X

Figure 1: Channel Flow: Computation domain, topology, mesh and cell deformation.

The impact of mesh deformation on solution accuracy is shown in Fig. 2. Error is mea-
sured in terms of the maximum amplitude of the spurious pressure fluctuations (|∆p|max ).
Ideally, the time-depedent mesh motion should produce no changes in the solution field.
However, as the implemented approach only satisfies the GCL approximately, some in-
fluence on the solution quality is to be expected. In Fig. 2 the influence of displacement
amplitude and frequency, between 10 Hz and 10000 Hz, is shown. It can be seen that
the error varies approximately linearly with both disturbance frequency and displacement
amplitude. Using the linear approximation to evaluate the time derivative of the Jacobain
according to Eqn. 11 large spurious pressure fluctuations are obtained. When using the
HB approach the spurious fluctuations are reduced by up to two orders of magnitude.
For this approach, at the largest investigated frequency (f = 10000 Hz) and the largest
relative volume change of 30% spurious pressure fluctions with amplitudes of approxi-
mately 50 Pa are observed. At more moderate frequencies and amplitudes the amplitude
of the suprious pressure fluctuations is of the order 1 Pa. This emphasises that the HB
formulation has to be used in the Harmonic Balance method to avoid the corruption of
the results by spurious perturbations.

4 APPLICATION
To validate the HB-solver for flutter problems the extended solver is applied in this
section to compute the well documented aeroelastic testcase Standard Configuration 10
(STCF10) [3]. STCF10 is a two-dimensional compressor cascade. The profile is con-
structed by superimposing the thickness distribution of a modified NACA0006 airfoil
onto a circular-arc camber line. The stagger angle is 45 and the pitch to chord ratio
is 1. The computational domain and mesh topology used in the simulations are shown
in Fig. 3. The two-dimensional computational domain comprises a total of 5600 cells
distributed over 6 blocks. The entry and exit boundaries of the computational domain
are located approximately one chord length upstream and downstream of the compres-
sor blade, respectively. In accordance with most data in the literature the test case is
simulated by solving the nonlinear Euler equations.
To investigate the accuracy of the HB-solver STCF10 is simulated at two flow con-
ditions: subsonic flow (M1 = 0.7 and β1 = 55◦ ), and transonic flow (M1 = 0.8 and
β1 = 58◦ ). Figures 4 and 5 show the computed time-mean flow fields in terms of the
isentropic Mach number Mis . At both operating points the numerical results agree very

7
 Graham Ashcroft, Christian Frey, Hans-Peter Kersken

5
10
V / V|max = 0.30 (Linear approx.)
10
4 V / V|max = 0.30
V / V|max = 0.15
Error (| P|max) [Pa]
10
3 V / V|max = 0.03

102

1
10

0
10

-1
10

-2
10 1 2 3 4
10 10 10 10
Frequency [Hz]

Figure 2: Channel Flow: Variation of error (in terms of pressure disturbance amplitude) with disturbance
frequency and relative cell volume change.

well with the corresponding data from the literature [9]. The aeroelastic stability of the
two-dimensional compressor cascade is investigated for a torsional mode with reduced fre-
quency ω ∗ = cω/2|U |Inlet = 0.5, amplitude 0.1◦ and axis of rotation located at midchord.
The computed values of aerodynamic damping, normalized against the value computed at
σ = 180◦ , are shown in Fig. 7 and compared against reference data from the literature [9].
The HB simlations have been performed here using a single harmonic at the frequency
of the eigenmode. As can be seen the HB solutions show excellent agreement with the
linearTRACE results and the data from the literature, with the acoustic resonances at
approximately σ = −120◦ and σ = −60◦ being well predicted. A comparison of the com-
puted pressure disturbance fields is presented in Fig. 6. Here the real part of the complex
valued pressure field at the first harmonic of the eigenmode is plotted for the interblade
phase angle σ = −130◦ . Again the results show excellent agreement and demonstrate
that the pseudo-time formulation of the nonreflecting boundary conditions used in the
HB-solver perform equal well as their GMRES counterpart used in linearTRACE.
Under the subsonic flow conditions the computed values of aerodynamic damping were
found to be largely insensitive to the amplitude of the torsional motion. In contrast, at the
transonic flow condition the presence of the shock on the suction side of the compressor
blade made the HB results very sensitive to the amplitude of the torsional motion. With
an amplitude of 0.1◦ significant discrepencies (of the order 25%) were observed in the com-
puted values of aerodynamic damping at interblade phase angles near ±180◦ . By reducing
the amplitude of the torsional motion by an order of magnitude the results shown in Fig. 8
could be obtained. As can be seen the degree of agreement between linearTRACE and
the HB-solver is for this amplitude excellent. Furthermore, good agreement is observed

8
 Graham Ashcroft, Christian Frey, Hans-Peter Kersken

Z X

Figure 3: Standard Configuration 10: Computational domain, topology and mesh (every second mesh
point only).

between the HB results and the data from the literature.

1.0 TRACE
Petrie-Repar
0.8

0.6
Mis [-]

0.4

0.2

0.0

0.0 0.2 0.4 0.6 0.8 1.0

xc/c [-]

Figure 4: Standard Configuration 10: Comparison of isentropic Mach number distributions at subsonic
flow conditions.

5 CONCLUSIONS
A Harmonic Balance solver has been extened and validated for the simulation of aeroe-
lastic problems in the field of turbomachinery. The HB-algorithm and its extension for
deforming meshes has been described in detail. Using a simple channel flow the impact of
time-dependent mesh deformation on the accuracy of numerical method has been quanti-

9
 Graham Ashcroft, Christian Frey, Hans-Peter Kersken

1.4
TRACE
1.2 Petrie-Repar

1.0
Mis [-]

0.8

0.6

0.4

0.2

0.0
0.0 0.2 0.4 0.6 0.8 1.0
xc/c [-]

Figure 5: Standard Configuration 10: Comparison of isentropic Mach number distributions at transonic
flow conditions.

fied. Results show that although time-dependent mesh deformation causes the generation
of spurious pressure fluctuations, their amplitudes are only a small fraction of the ampli-
tudes of typical physical pressure fluctuation induced for example by blade vibrations are
therefore can be neglected in practical applications. Application of the HB-solver to the
reference testcase Standard Configuration 10 has shown that the method can be used to
accurately predict the flutter characteristics of the two-dimensional compressor cascade
at both subsonic and transonic flow conditions. At transonic flow conditions numerical
investigations indicate the presence of nonlinear flow effects for amplitudes of oscillation
greater than 0.01◦ .

ACKNOWLEDGEMENTS
This work was supported by the German Federal Ministry of Economics and Technology
under grant number 20T1104B.

REFERENCES
[1] K. Becker, K. Heitkamp, and E. Kügeler. Recent progress in a hybrid-grid CFD
solver for turbomachinery flows. In Proceedings Fifth European Conference on Com-
putational Fluid Dynamics ECCOMAS CFD 2010, 2010.

[2] M. Franke, T. Röber, E. Kügeler, and G. Ashcroft. Turbulence treatment in steady

and unsteady turbomachinery flows. In Proceedings Fifth European Conference on
Computational Fluid Dynamics ECCOMAS CFD 2010, Lisbon, Portugal, June 2010.

[3] T. H. Fransson and J. M. Verdon. Updated report on Standard Configurations for

10
 Graham Ashcroft, Christian Frey, Hans-Peter Kersken

Figure 6: Comparison of the real parts of the computed pressure perturbation fields for Standard
Configuration 10 with an inter-blade phase angle of -130 deg.: time-linearized method (left) and harmonic
balance method (right).

the Determination of unsteady Flow Through Vibrating Axial-flow Turbomachine-

Cascades. Technical Report TRITA/KRV/92.009, KTH, Stockholm, 1992.

[4] C. Frey, G. Ashcroft, H.-P. Kersken, and C. Voigt. A Harmonic Balance Technique
for Multistage Turbomachinery Applications. In Proceedings of the ASME Turbo
Expo 2014, 2014.

[5] M. B. Giles. Nonreflecting boundary conditions for Euler calculations. AIAA Journal,
28(12):2050–2058, 1990.

[6] T. Guedeney, A. Gomar, F. Gallard, F. Sicot, G. Dufour, and G. Puigt. Non-uniform

time sampling for multiple-frequency harmonic balance computations. Journal of
Computational Physics, 236:317–345, 2013.

[7] K. C. Hall and W. S. Clark. Linearized Euler predictions of unsteady aerodynamic

loads in cascades. AIAA Journal, 31(3):540–550, 1993.

[8] H.-P. Kersken, C. Frey, C. Voigt, and G. Ashcroft. Time-Linearized and Time-
Accurate 3D RANS Methods for Aeroelastic Analysis in Turbomachinery. J. Turbo-
mach., 134(5), 2012.

[9] P. J. Petrie-Repar, A. McGhee, and P. A. Jacobs. Three-Dimensional Viscous Flutter

Analysis of Standard Configuration 10. ASME Paper No. GT2007-27800, 2007.

11
 Graham Ashcroft, Christian Frey, Hans-Peter Kersken

2.5

Petrie-Repar

Aerodynamic Damping [-]

linearTRACE
2.0
HB

1.5

1.0

0.5

0.0
-180 -120 -60 0 60 120 180
Interblade Phase Angle [Deg.]

Figure 7: Normalized Aerodynamic Damping at Subsonic Flow Conditions and ω ∗ = 0.5.

[10] C. Voigt, C. Frey, and H.-P. Kersken. Development of a generic surface mapping
algorithm for fluid-structure-interaction simulations in turbomachinery. In J. C. F.
Pereira, A. Sequeira, and J. M. C. Pereira, editors, V European Conference on Com-
putational Fluid Dynamics ECCOMAS CFD 2010, June 2010.

12
 Graham Ashcroft, Christian Frey, Hans-Peter Kersken

2.0
Petrie-Repar
linearTRACE
Aerodynamic Damping [-]

HB (Amplitude 0.01 Deg.)

1.5 HB (Amplitude 0.10 Deg.)

1.0

0.5

0.0

-180 -120 -60 0 60 120 180

Interblade Phase Angle [Deg.]

Figure 8: Normalized Aerodynamic Damping at Transonic Flow Conditions and ω ∗ = 0.25.

13