Proceedings of ASME Turbo Expo

IGTI 2014
June 16-20, 2014, Düsseldorf, Germany

GT2014-25474

FLUTTER ANALYSIS FOR TURBOMACHINERY USING VOLTERRA SERIES

Meng-Sing Liou Weigang Yao∗

Aeropropulsion Division NASA Postdoctoral Program
NASA Glenn Research Center NASA Glenn Research Center
Cleveland, OH 44135 Cleveland, OH 44135
Email: [email protected] Email: [email protected].

NOMENCLATURE
a Speed of sound.
ABSTRACT A Area vector.
The objective of this paper is to describe an accurate and c Blade chord length.
efficient reduced order modeling method for aeroelastic (AE) Cp Specific heat at constant pressure.
analysis and for determining the flutter boundary. Without los- d structural deformation vector.
ing accuracy, we develop a reduced order model based on the E Specific total energy.
Volterra series to achieve significant savings in computational Fm Modal force vector.
cost. The aerodynamic force is provided by a high-fidelity so- F Inviscid flux vector.
lution from the Reynolds-averaged Navier-Stokes (RANS) equa- Fv viscous flux vector.
tions; the structural mode shapes are determined from the finite hm (n) mth -order Volterra series kernel.
element analysis. The fluid-structure coupling is then modeled K Stiffness matrix.
by the state-space formulation with the structural displacement L Length measured in the x-direction.
as input and the aerodynamic force as output, which in turn acts M Mass matrix.
as an external force to the aeroelastic displacement equation for M∞ Freestream Mach number.
providing the structural deformation. NASA’s rotor 67 blade is p Pressure.
used to study its aeroelastic characteristics under the designated Pκ (ω ) Turbulence production in the κ (ω ) equation.
operating condition. First, the CFD results are validated against Pr Prandtl number.
measured data available for the steady state condition. Then, qi Heat flux vector.
the accuracy of the developed reduced order model is compared
q∞ Dynamic pressure(=ρ∞ a2∞ )
with the full-order solutions. Finally the aeroelastic solutions
U Conservative variables.
of the blade are computed and a flutter boundary is identified,
r Position vector.
suggesting that the rotor, with the material property chosen for
Re Reynolds number.
the study, is structurally stable at the operating condition, free of
S Source term.
encountering flutter.
T Temperature.
U∞ Inflow velocity magnitude.
∗ Currently
ui Cartesian velocity components.
School of Mechanical and Aerospace Engineering, Queen‘s Uni-
versity, Belfast, UK.
x, y, z Axial (horizontal), spanwise and vertical direction.
This
c material is declared a work of the U.S. Government and is not sub- δi j Kronecker delta.
ject to copyright protection in the United States. Approved for public release; ξ,η Modal displacement.
distribution is unlimited.
1 Copyright 
c 2014 by ASME
κ Turbulence kinetic energy. and a structural dynamics code, as a result they in turn determine
μ Viscosity. the reliability and efficiency of the tool.
ω Specific turbulence dissipation rate. For aerodynamics analysis, developments in computational
Ω Rotating speed. fluid dynamics over the past several decades have provided in-
σ Value of step function. creasingly powerful and reliable capabilities. The complex-
τ Pseudo time. ity and fidelity, hence its range of applicability, of analysis is
τi, j Viscous stress tensor. strongly correlated with the fast evolution of computer power:
V Control volume. from the early linearized potential flow solution to the current
Subscripts large eddy simulations using Navier-Stokes equations. Linear
A Aerodynamic. models are still used widely in the design phase. But develop-
∞ Far upstream or “free” stream. ments in computer technology and CFD methods and software
m Modal coordinate have made use of high-fidelity models feasible even in early stage
s Structural dynamics. of a design cycle. However, large eddy simulations are still far
t Total (stagnation) condition. too costly and from being timely to be adopted in the design pro-
turb Turbulence. cess.
In the current study, we employ the Reynolds-average
Navier-Stokes equations for which the turbulence terms are
1 INTRODUCTION closed with the two-equation κ -ω model, specifically the shear
NASA is considering new generations of aircraft that meet stress transport (SST) version by Menter [2]. The second-order
aggressive economic, noise and environmental targets; a spe- backward differencing is used for time-discretization. The non-
cial configuration, called N3-X, employs all electric power and linear inviscid terms are approximated by the AUSM+ -up [3]
propulsion systems by which the thrust force for the vehicle is method while the viscous terms approximated by the usual cen-
generated exclusively with an array of fans housed in compart- tered formulas. The resulting implicit algebraic system is then
mentalized flow paths. Hence, the designing of fans to meet es- solved by the LUSGS method [4].
sential considerations is paramount. Specifically, increasing per- For structural dynamics, one may invoke the full finite ele-
formance and operating life and reducing weight to optimize the ment analysis, as employed in the aeroelastic study of rotor 67
economic objective, while reducing noise and emission to meet by Doi [5]. The resulting fluid-structure system is a time depen-
environmental regulations. In pursuit of higher performance of a dent set of equations describing not only the flow variables in the
compressor/fan, the past design trend is to run at higher pressure entire domain, but also the motion of the structure immersed in
ratios and higher mass flow rates, thus moving close to flutter the fluid. The system can be solved either in the frequency [6] or
boundaries associated with surge or choke as defined in the com- time domain [5]. The frequency domain approach may be pre-
pressor map. Hence, it is important to ensure the compressor ferred for linear problems for its computational efficiency; how-
is structurally sound over the entire operating range, from the ever for a nonlinear problem, it is more efficient and accurate to
choke to the stall conditions. Structural vibration, either caused arrive at solution with the time domain approach.
by natural resonance or forced response, is a major consideration The time-domain computation for flutter analysis can be-
in assessing the devise’s structural integrity. The fluid-induced come costly when a large number of time-dependent solutions
instability of a compressor blade is typically not of concern un- of the fluid-structure system are needed, for example as part of
less it is tuned to the natural vibration frequency. However, it a design process. It is therefore desirable to reduce the com-
becomes an issue in transonic speed regime, because a small dis- putational cost by a significant factor, for example by at least
turbance can result in a large amplitude variation and nonlinear an order of magnitude or more. This can be readily achieved
behavior. The unsteady excursion of a shock wave through the by employing strategies called model order reduction (MOR), of
blade-to-blade passage can intermittently choke or stall the flow, which the harmonic balance, proper orthogonal decomposition,
potentially crossing the flutter boundary. The unsteady forces and Volterra series are among the most popular. [7] Model order
resulting from the shock motion are shown to have either stabi- reduction should not only save computational effort, but also re-
lizing or destabilizing effects, depending on the shock structure tain the fidelity of the original (full) system. This goal has been
and inter-blade phase angle. [1] well realized for linear problems through model reduction, but
With advances in computers and computational fluid dynam- not yet universally for nonlinear problems.
ics (CFD), aeroelastic analysis is fast becoming common for real For nonlinear problems, the Volterra series expansion is used
world designs. To be useful and adopted in practice, a compu- to approximate the input-output relationship of a nonlinear time
tational tool must be reliable for predicting the aeroelastic char- dependent system, with a capability of capturing ”memory” ef-
acteristics and just as importantly be efficient (cheap and fast). fects. This input-output concept is well suited for the aeroelastic
This computational tool will consist of an aerodynamics code analysis in which the aerodynamic force and structural deforma-

2 Copyright 
c 2014 by ASME
tion can be formulated in this framework. Furthermore, flowfield and turbulence generation.
and structural dynamics have different time scales and their in-
teractions often respond with delay in time, i.e., with ”memory” ⎛ ⎞ ⎛ ⎞ ⎛ ⎞
ρ ρ (ui − ugi ) 0
effects. The Volterra series has been applied in various fields ⎜ ρ u1 ⎟ ⎜ ρ (ui − ug )u1 ⎟ ⎜ pδi1 ⎟
of engineering and is mostly used to construct a reduced order ⎜ ⎟ ⎜ i ⎟ ⎜ ⎟
⎜ ρ u2 ⎟ ⎜ ρ (ui − ug )u2 ⎟ ⎜ pδi2 ⎟
model to mimic a complex dynamic system. Unsteady aerody- ⎜ ⎟ ⎜ i ⎟ ⎜ ⎟
U=⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ρ u3 ⎟ , Fi = ⎜ ρ (ui − ugi )u3 ⎟ , Pi = ⎜ pδi3 ⎟ ,
namic force responses to wing motion have been calculated by ⎜ ρE ⎟ ⎜ ρ (ui − ug )E + pui ⎟ ⎜ 0 ⎟
Silva [8] using the Volterra theory. In the present work, we de- ⎜ ⎟ ⎜ i ⎟ ⎜ ⎟
⎝ ρκ ⎠ ⎝ ρ (ui − ugi )κ ⎠ ⎝ 0 ⎠
scribe the application of the Volterra series, based on RANS so-
ρω ρ (ui − ugi )ω 0
lutions, to turbomachinery aeroelastic problems. (2)
The paper is organized as follows. Section 2 gives the and the terms attributed to viscous diffusion
fluid and structure equations employed and outlines the methods
⎛ ⎞ ⎛ ⎞
adopted to solve them, especially including detailed description 0 0
of fluid-structure coupling and model order reduction based on ⎜ τi1 ⎟ ⎜ 0 ⎟
⎜ ⎟ ⎜ ⎟
the Volterra series. Section 3 presents the application to aeroelas- ⎜ τi2 ⎟ ⎜ −ρ u3 ⎟
⎜ ⎟ ⎜ ⎟
tic analysis for NASA’s rotor 67 compressor blade along with the Fvi = ⎜
⎜ τi3
⎟ , S = Ω ⎜ ρ u2 ⎟ ,
⎟ ⎜ ⎟ (3)
validation of the CFD solution against measured data. In Section ⎜ τi j u j + qi ⎟ ⎜ 0 ⎟
⎜ ⎟ ⎜ ⎟
3.2 we show the results of applying the developed reduced-order ⎝ qκi ⎠ ⎝ Pκ ⎠
model to find the flutter boundary of rotor 67. qωi Pω

Here the subscript “i” denotes the direction in Cartesian coordi-

nates, e.g., F = ∑3i=1 Fi ei . For turbomachinery applications, the
2 MATHEMATICAL SYSTEM coordinate x1 is chosen to be aligned with the rotor axial direc-
The mathematical system considered for predicting the flut- tion, and the grid velocity follows a rigid body rotation with Ωi,
ter conditions consists of the fluid and structure equations, as
described in two respective sections. ug =r × Ωi = Ω(zj − yk) (4)

It is noted that in the aeroelastic calculation performed in this

2.1 Fluid Equations study, the position vectorr(t) of a computational cell varies with
time as the blade geometry changes, even though the rotating
The three-dimensional Reynolds-averaged Navier-Stokes
speed remains fixed.
(RANS) equations are employed, with the turbulence described
The above viscous stress terms include both the laminar and
by the two-equation κ -ω SST model [2]. They are written in the
turbulent effects through the use of eddy viscosity (μt , kt ) model,
following integral form over a control volume V (x,t) enclosed
expressed as:
by a control surface ∂ V (x,t):

∂uj
τi j = (μ + μturb ) ∂ ui
∂ x j + ∂ xx − 23 ∂∂ uxk δi j
μ μturb ∂ T
k (5)
    qi = −( Pr + Prturb ) ∂ xi
d
UdV + (F + P) · dA = Fv · dA + SdV (1)
dt V ∂V ∂V V
The turbulence eddy viscosity for the study presented herein
is provided through the solution of two transport equations of
scalar quantities (κ , ω ), specifically the so-called κ -ω turbu-
where we have the standard notation for the conservative vari-
lence model [10] is enhanced with Menter’s shear-stress trans-
ables plus the turbulence variables in U. The surface integral on
port (SST) model [2]. The details of the turbulence model, well-
∂ V (x,t) consists of fluxes through the vectorial area dA can be
known and elaborated in the cited reference, are omitted here.
expressed in terms of 3 Cartesian coordinates. The relative con-
vective flux Fi , the pressure flux Pi , and the viscous stresses and
heat flux Fvi in the i-direction, i = 1, 2, 3 are given in Eq. (2-3), 2.2 CFD Solution Methods
written in the relative coordinate system moving with the speed To eliminate accumulative time integration error, we opt for
ug . [9] The source terms includes the rigid-body rotation, Eq. (4) the dual-time stepping approach, in which a time rate of change

3 Copyright 
c 2014 by ASME
of U in pseudo time (τ ) is added to Eq. (1) so that at each determination of the modal (generalized) force Fm = {Fmi ; i =
new time level the unsteady CFD equations are balanced, namely 1, · · · , Nm }, where each component of the modal force is the in-
driving its discretized residual to diminish. The dual time step- ner product of the mode shape Φi and aerodynamic force vectors
ping strategy is formulated as: over the entire CFD nodes. Additionally, the physical deforma-
tion d of a structure can be expressed in terms of the mode shape
and the modal displacement ξ from the above dynamics equa-
 tion, leading to physical displacement, d = ∑Ni=1m
Φi ξi and physi-
d
UdV = R(U) Nm
dτ V cal force f = ∑i=1 Φi Fmi .
    The above second-order differential equation can be recast
d
=− UdV − (F + P) · dA + Fv · dA + SdV.
(6) into the following first-order differential system:
dt V ∂V ∂V V

In the physical time step (t) the residual R(U) is discretized η̇ = As η + Bs Fm (8)
implicitly using a 3-level, backward differencing in order to ob-
tain second-order temporal accuracy, resulting in a highly nonlin-
where
ear system. In the pseudo-time step the implicit system is solved
by performing fixed-point iterations untill the residual R(U) of
the nonlinear physical-time equation has diminished or reduced ξ 0 I 0
η= , As = , Bs = (9)
to specified small values. Then the solution is advanced to the ξ̇ −M−1 K 0 −M−1
next time level. This pseudo-time iteration is carried out by em-
ploying the LUSGS method [4]. and η = {ηi ; i = 1, · · · , Nm }.
The inviscid flux terms F + P arguably have received the The time derivative in Eq. (8) is approximated by the
most attention in past CFD research, especially those under the second-order Crank-Nicolson method, producing a discrete sys-
framework of upwind solvers, yielding many proposed schemes tem for t n ≤ t ≤ t n+1 ,
for approximating it. In this study, We employ the AUSM+ -up
method [11] for the inviscid fluxes. For the viscous terms Fv and
source S terms, a typical centered representation is used. η (n + 1) = (I − 0.5tAs )−1 ((I + 0.5tAs )η (n) + Bsd Fm (n))
The mesh velocity is obtained from the structural motion in (10)
response to the aerodynamic forces provided by the CFD solu- This will form part of the coupled fluid-structure (aeroelas-
tion. The structural model is described in the next section. tic) system to be elaborated below. It specifically provides the
The resulting in-house 3D RANS code has been developed time-dependent modal displacement ξ , hence the needed phys-
and validated for a variety of flow problems over a number of ical displacement of the structure so as to affect fluid flow in
years. For the validation relevant to the problem at hand will be response to the geometry variations. The mathematical system
described Section 3.1. describing the interactions between fluid and structural dynam-
ics is given below.
2.3 Structural Dynamics Equation
The finite element model for describing a structural motion 2.4 Fluid-Structural Coupling
is expressed in terms of its displacement ξ from a neutral posi- The coupling of aerodynamic and structural computations
tion (steady state in our case). In our work, we first carry out must be performed on a common geometry, while they need not
finite element analysis on a given set of nodes via MSC/Nastran be of the same mesh density or matching at the same grid points,
[12] to obtain mode shapes, Φi , i = 1, 2, · · · , Nm , Nm being the as displayed in Fig. 1 for the NASA rotor 67 blade, which is the
number of modes. Neglecting damping, the structural motion in structure that will be considered in this paper. As such, interpo-
terms of the modal displacement vector ξ = {ξi ; i = 1, · · · , Nm } lation/extrapolation procedures must be employed to accomplish
in response to the modal force Fm can be described by the mapping between them, through which the proper transfer of
relevant variables may be carried out. In our case, the structure
deformation provides a new body to the CFD process, thus af-
Mξ̈ + Kξ = Fm (7)
fecting boundary condition and the flow domain mesh. On the
other hand, the aerodynamic force needs to be transferred to the
where (M, K) are the mass and stiffness matrices of the mate- contact points for the finite element analysis. This mapping of
rial of the structure respectively. The modes on the FEM nodes grids must satisfy certain physical requirements, such as conser-
are then interpolated to every CFD node at which the aerody- vation of virtual work, and numerical requirements of accuracy
namic forces are known. This modal information facilitates the and stability.

4 Copyright 
c 2014 by ASME
 FIGURE 2: CFD AND CSD INTEGRATED COMPUTATION:
LOOSE COUPLING.

in which both the CFD and computational structural dynamics

(CSD) equations are solved simultaneously. The loosely cou-
pling strategy is easy to implement and computationally effi-
FIGURE 1: ILLUSTRATION OF GRIDS USED FOR cient; the lag in time between CFD and CSD is believed to be
AERODYNAMIC AND STRUCTURAL ANALYSES. BLUE: insignificant, since the time step is usually much smaller than the
STRUCTURAL GRID; RED: AERODYNAMIC GRID. TYPI- characteristic time of the problem under study. This combined
CALLY THE STRUCTURAL GRID IS MUCH COARSER. CFD-CSD full order modeling process is extremely expensive
especially when a large number of computations are committed.
In real-world engineering practice, analysis is not performed
only for one condition, but over many computations. In ad-
The constant volume transformation method [13] was at-
dition, design optimization typically will require hundreds and
tempted to interpolate/extrapolate between the aerodynamic and
thousands similar computations, differing for example in range
structural grids, but it failed to provide a stable and converged
of conditions, parameters, or geometry. A reduced-order model
solution because of severe geometrical twisting involved in the
takes only a small fraction of computational time needed by a
present case. Instead, we employed a surrogate model to provide
full-order model, but is of value if and only if it is capable of
the structural mode shapes.
preserving the accuracy of the full-order system. This can be
The radial-based function (RBF) neural network method, achieved easily for a linear system, but still remains a topic of
which we previously used for reduced-order modeling of flut- intensive research for a nonlinear system [7].
ter and limit cycle oscillations [14], is applied here by taking the
As all the aeroelastic computations for finding flutter/LCOs
finite element nodes as input and modes as output via a training
are similar in kind and repetitive, they differ only by a limited
process by virtue of Nastran calculations. Once the training (the
number of variables and the variations in value. A model order
RBF and neurons) is accomplished, the neural network can take
reduction will be of great value in significantly reducing com-
the aerodynamic grid as input and produce modes as output, the
putational cost and time. In what follows we will describe the
modes are in turn interpolated to the aerodynamic grid.
application of Voterra theory [15] for constructing a reduced or-
The resulting first three mode shapes are displayed in Fig. der model for aeroelastic analysis, based on the fidelity of the
11, revealing indeed a large deformation of the structure. Once RANS equations for aerodynamic calculations.
the blade shape deformation caused by the aerodynamic force
is updated as described above, then the computational mesh for
CFD is changed accordingly using transfinite interpolation (TFI) 2.5 MODEL ORDER REDUCTION BY VOLTERRA SE-
at each time step. As such, it also allows physical variables to be RIES
interpolated onto the new grid in the same manner. The Volterra theory provides a functional relationship repre-
A typical fluid-structure coupling is performed as shown in senting a nonlinear response to a given input function which may
Fig. 2, where the structural deformation is known, hence mesh be time dependent and is capable of capturing the ”memory” ef-
generated at t n , a subsequent CFD solution for the new time step fect. While it has been employed in previous studies for aeroe-
at t n+1 for Un+1 is performed with the structural shape frozen lastic application, for example [16, 17], these have been limited
at qn , then it is followed by a geometry update to get qn+1 to external flows over an airfoil or a wing. To our knowledge, the
with the input of Un+1 by procedure 3. This is the so-called current paper represents the first aeroelastic application of the
loosely coupling strategy, in contrast to the tightly coupling one Volterra theory to turbomachine. The Volterra theory has some

5 Copyright 
c 2014 by ASME
advantages over other ROMs, see [7, 17] for more discussion. a system response after applying a step function,
The Volterra theory can be easily adopted as an alternative proce-
dure without having to modify the baseline full-order procedure. 
σ0 , n > 0,
It is equally applicable to the time and frequency domains and σ (n) = (14)
0, n = 0.
the conversion between them is rather simple. Moreover, the for-
mulation facilitates to retain nonlinearity of the full order model
more easily than other ROMs. A small number is given to σ0 = 1.0 × 10−4 to ensure the prob-
The Volterra series, unlike the Taylor series, includes in the lem remains linear. Then, according to Eq. (13), we have the
output accumulative effects of inputs occurring at previous times. response,
The output y(t) of a continuous time-invariant system in response
to a single input u(t) for t ≥ 0 is expressed by the Volterra theory n
as: y(n) = h0 + σ0 ∑ h1 (n − k), (15)
k=0

∞  t  t
y(t) = h0 + ∑ ··· hk (t − τ1 , · · · ,t − τi )u(τ1 ) · · · u(τi )d τ1 · · · d τi . And the first kernel is readily available as
i=1 0 0
(11) 
where h0 is the steady-state term coincident with the initial con- 0, n = 0,
h1 (n) = (16)
dition and hi , i ≥ 1 are known as the Volterra kernels. As the time (y(n) − y(n − 1))/σ0 , n ≥ 1.
integral is discretized over a n-interval domain, a time-discrete
infinite (or truncated) Volterra series is obtained: The first equality holds because of the initial condition y(0) =
h(0).
n In what follows we show how to construct a reduced order
y(n) = h0 + ∑ h1 (n − k)u(k) model that simply bases on a relationship between the structural
k=0 motion and aerodynamic force, from the viewpoint of relating
n n
input and output data. This is easily facilitated within the state-
+ ∑ ∑ h2 (n − k1 , n − k2 )u(k1 )u(k2 ) + · · · (12)
space theory, as used in control theory. a linear state-space sys-
k1 =0 k2 =0
n n tem can be represented in the following canonical form:
+ ∑ ··· ∑ hm (n − k1 , · · · , n − km )u(k1 ) · · · u(km ) + · · ·
k1 =0 km =0
xa (n + 1) = Aa xa (n) + Ba ξ (n) (17)
Fa (n + 1) = Ca xa (n) + Da ξ (n) (18)
where y(n) is the output with the time index n referring to t n ,
u(k) is the input at preceding times k = 0, 1, 2, ..., n, and hm the
where xa (n) is the state vector at time n. The input ξ is the
mth-order Volterra kernel, m = 1, 2, · · · , ∞. For a linear system, it
structural displacement and the system output Fa denotes the non
suffices to keep only the first-order Volterra kernel, hence
dimensional generalized aerodynamic force.
To set up the above system and solve for the aeroelastic sys-
n tem under consideration, we adopt the Eigensystem Realization
y(n) = h0 + ∑ h1 (n − k)u(k) (13) Algorithm (ERA) [18]. First, we define the finite Hankel matrix
k=0
constructed using the first-order Volterra kernel h1 just described
above,
where h0 corresponds to the response with zero input, or the force
vector at steady state where there is no structural response. To ⎡ ⎤
h1 (k) h1 (k + 1) ··· h1 (k + β − 1)
capture behaviors varying with time variation, one must at least ⎢ h1 (k + 1) h1 (k + 2) ··· h1 (k + β ) ⎥
find the first-order kernel associated with the input at all other ⎢ ⎥
⎢ h1 (k + 2) h1 (k + 3) ··· h1 (k + β + 1) ⎥
times. it turns out that from the continuous system, the first ker- H(k −1) = ⎢ ⎥
⎢ .. .. .. .. ⎥
nel measures the response to an impulse applied at τ1 = 0. To ⎣ . . . . ⎦
include nonlinear effects, higher order kernels are necessary, see h1 (k + α − 1) h1 (k + α ) · · · h1 (k + α + β − 2) α ×β
Silva [8] (19)
In the present study, we make use of the first kernel to build where α and β are the sampling time shift in the row and column
our reduced order model (ROM), for which the necessary step is directions respectively; they control the order (rank) of the sys-
the definition of h1 (n), for n ≥ 0, as will be illustrated below for tem, and are set as α = 1600, and β = 50 in our study. Applying

6 Copyright 
c 2014 by ASME
Singular Value Decomposition to H(0), track starts with the baseline CFD solution as the full-order will,
then builds the Volterra kernel shown in Eq. (16), which forms
H(0) = UΣV T (20) the state space system in Eq. (17). The input and output of which,
ξ and Fa , are coupled with the structural dynamics system in Eq.
(10). It is noted that Fm = q∞ Fa , with q∞ being the dynamic pres-
we find U, Σ and V , which are then used to define the matrices in sure (ρ∞ a2∞ ). These two systems combined form the ROM for the
Eq. (17): aeroelastic analysis discussed next.

Aa = Σ−1/2U T H(0)V Σ−1/2

Ba = Σ1/2V T EL
Ca = EM
T
UΣ1/2
Da = h1 (0) (21)

where

T
EM = [IM 0M · · · 0M ]α M×M
ELT = [IL 0L · · · 0L ]β L×L (22)

with M and L being the number of inputs and outputs respec-

tively. Since only first three modes are retained, we have M = 3
and L = 3. The size of the ROM is 3 × β = 150 for this study. FIGURE 4: FLOW CHART ILLUSTRATING THE PRO-
CESS OF PERFORMING AEROELASTIC ANALYSIS IN
THE PRESENT STUDY.
1

0.1
singular value

3 Aeroelastic Analysis of NASA Rotor 67

0.01 Complex vibration problems arise from the interactions of
nonlinear aerodynamics and structural deformation. These vi-
brations can be either self-induced or caused by flow distortions
1E-3 from upstream and downstream blade rows or tip region; the for-
mer is called flutter and the later forced response. Also mistuning
in blade rows and inter-blade phase angle can force vibration on
1E-4 a blade. [19] Bendiksen [20] gives a comprehensive review of
100 200 300 400 500 aeroelastic problems encountered in turbo machines, in which
n various factors causing flutter are identified. A recent discus-
sion on the progress and challenges of computational aeroelsac-
FIGURE 3: DECAY OF (NORMALIZED) SINGULAR VAL- tic modeling can be found in Bartels and Sayma. [21] Aeroelas-
UES OF THE HANKEL MATRIX. tic analyses of rotor 67 have been conducted by Doi [5], Sadeghi
and Liu [22] and Zhang et al [19], employing full-order mod-
eling of the fluid-structural system where RANS is used in the
CFD procedure. An inlet guid vane row is included in Zhang
In Fig. 3, we show the efficiency of the reduction method. et al to study its effect on the flutter characteristics. Doi found
Singular values of the Hankel function are seen to decay rapidly that operating condition and inter-blade phase angle determine
within the first 50 values, indicating that model order is of rea- the stability of the structural response; Sadeghi found the rotor
sonable size needed to retain accuracy. to be stable using the 10 first eigenmodes. In our study, we also
It is appropriate at this juncture to illustrate the entire aeroe- choose this rotor for our computational test model because the
lastic analysis process in a flow chart shown in Fig. 4. The ROM model reveals most flow complexities seen in the turbomachines

7 Copyright 
c 2014 by ASME
in today’s aircraft and also widely used in the turbo machinery
community for validation of CFD results, thus allowing us to
verify our proposed approach for AE analysis against previous
works, for example [23, 24, 5, 22].
In what follows, we shall first validate the CFD solution for
detailed profiles and performance map against the measured data
taken in [25]. Then the fluid-structure coupling procedure will be
described, followed by the aeroelastic calculation of the blade. A
model order reduction method based on the Volterra series is in-
troduced and applied to rotor 67 to determine the flutter behavior.

3.1 Validation at Steady State Operating Points

NASA rotor 67, shown in Fig. 5 is the first stage rotor of a
two-stage fan [26]. It is a low aspect ratio (1.56) transonic axial
flow rotor with a design tip relative Mach number of 1.38; an ex-
periment program was undertaken to provide laser anemometry
and aerodynamic performance data at Glenn (formerly Lewis) FIGURE 5: TEST MOD
MODEL
OD
O DEL OF THE NASA ROTOR 67.
Research Center in 1980s, culminating in an extensive compila-
tion by Strazisar et al [25]. Shown in Fig. 6 is the test model
of the rotor with 22 blades assembled. The rotor was designed
for axial inflow and did not require inlet guide vanes, nor a stator
stage. The design total pressure ratio is 1.63 at a mass flow rate of
33.25 kg/s (choked at 34.96 kg/s) and a rotating speed of 16,043
rpm. Other geometrical dimensions and operating conditions can
be found in the cited reference. The laser anemometry measure-
ments acquired on streamsurfaces, starting at roughly one chord
length upstream of the rotor and continuing through it till some
distance into the wake, providing detailed data in the form of rel-
ative Mach number and relative flow angle. Flow variables were
also available at an upstream and a downstream planes. These
data will be used to validate our computed results first, before
building up the reduced order.
The characteristics boundary conditions are employed for
the inviscid boundaries: at the subsonic inflow boundary the left
running (negative) Riemann variable is extrapolated from the in-
terior domain; at the subsonic outflow boundary the static pres-
sure is specified at the hub and radial pressure equilibrium as- FIGURE 6:BLADE SHAPE OF THE NASA ROTOR 67
sumed to calculate other radial points. The no-slip conditions AS MOUNTED ON THE HUB, TOGETHER WITH THE
are applied at the hub. But the shroud in this study is assumed to ALIGNED COMPUTATIONAL SURFACE.
be an inviscid wall (or a streamsurface), this may cause some dis-
crepancies of our results from the data, as will be remarked when
appropriate. The mesh used for the CFD solution is shown for rections. This mesh density may be considered coarse in today’s
the blade tip in Fig. 7 in an overall view and two enlarged views CFD practice, however, Fig. 8 shows that the computed profiles
showing the dense mesh at the blade surface and in the wake. of static pressure and total pressure and temperature ate the exit
Similar mesh distribution is also generated for all spanwise sec- plane from both grids are essentially indistinguishable. These
tions. Two H-type mesh systems, one coarse and another fine, are solutions are also comparable with the measured data and other
used first to establish whether the coarse mesh is sufficiently ac- published CFD results using finer meshes, for example in [5,22].
curate to be used for further aeroelastic analysis. The two meshes Hence we consider the coarse grid to adequate to provide suf-
respectively consist of 77x43x45 and 104x63x80 grid points (re- ficiently accurate aerodynamic forces to the structural analysis
sulting in 140,448 and 504,494 cells)–the three numbers respec- and thus, to be employed in this study. Moreover, our emphasis
tively refer to the streamwise, blade-to-blade, and spanwise di- in this paper is to show the efficacy and validity of the proposed

8 Copyright 
c 2014 by ASME
 (a) blade tip (b) leading edge (a) 10% span from shroud (b) 10% span from shroud

(c) trailing edge

(c) 30% span from shroud (d) 30% span from shroud
FIGURE 7: CFD MESH AT THE BLADE TIP.

(a) static pressure ratio (b) total pressure ratio (c) total temperature ratio

(e) 70% span from shroud (f) 70% span from shroud
FIGURE 8: PROFILES OF STATIC PRESSURE, TOTAL
PRESSURE AND TEMPERATURE RATIOS AT AN EXIT LO-
FIGURE 9: RELATIVE MACH NUMBER CONTOURS AT
CATION WHEN THE ROTOR IS NEAR PEAK EFFICIENCY.
THREE SPANWISE SECTIONS, RESPECTIVELY 10%, 30%
AND 70% MEASURED FROM SHROUD.
model order reduction method for AE analysis.
It is noted, however, that an overestimation is found in the
static pressure ratio by the computation. This is probably caused
by several simplifications committed in our computational setup: The relative Mach contours at three spanwise sections, re-
(1) we did not assume a boundary layer profile at the inflow spectively 10%, 30% and 70% measured from the tip, are com-
boundary while in the experimental setup a solid surface is con- pared for the peak efficient condition in Fig. 9, revealing the
nected to the hub surface of the rotor, (2) the tip clearance is not nearly normal shock wave across the blade passage at the tip sec-
taken into account and instead an inviscid slip wall is assumed at tion, but subsonic or low supersonic near the root.
the casing, and (3) the hub wall is assumed adiabatic, hence pos- Finally, we plot the rotor 67 performance by the CFD solu-
sibly giving rise to a higher temperature or pressure in the layer tion in comparison with the measured values, as shown in Fig.
at the hub. This low-momentum layer at the inlet will continue 10. At the peak efficiency point, the solution gives a mass flow
to develop, growing through the rotor, resulting in a thickened rate of 33.68 kg/s, a total pressure ratio of 1.651, and an effi-
boundary layer profile, in comparison with the computed result ciency of 0.9178. The calculated results are in good agreement
which indicates a fuller profile in a thinner layer. with the data over the entire operating conditions.

9 Copyright 
c 2014 by ASME
 0.96 Experiment TABLE 1: MATERIAL PROPERTIES.
caculated
adibatic efficiency

Material Young’s modulus Poisson’s ratio Density

0.92
(Pa) (kg/m3 )
0.88 Titanium Alloy 1.172×1011 0.3 4539.5

0.84 TABLE 2: MODAL FREQUENCY (HZ).

0.80 1st mode 2nd mode 3rd mode

0.92 0.94 0.96 0.98 1.00
mass flow rate/mass flow rate at choke Present 369.8 1009.4 1622.9
(a) adiabatic efficiency Doi [5] 401.9 1096.0 2093.7

We now apply the model order reduction technique es-

tablished above to rotor 67 at the peak efficiency condition.
1.7 The time-dependent aerodynamic force is built using the above
total pressure ratio

Volterra series with the first mode displacement given as:

1.6
ξ1 = 5.0 × 10−5 sin ω t (23)
1.5
Experiment where ω is the first natural frequency of the structure.
1.4 calculated It is noted that the small amplitude is chosen in Eq. (23) to
ensure linearity assumed for the current ROM formulation. Since
0.92 0.94 0.96 0.98 1.00 the flutter boundary estimated by the linear theory is independent
mass flow rate/mass flow rate at choke of the perturbation magnitude, it is not critical what value is used
(b) total pressure as long as the value is small. The time step used in the time
integration is chosen to be sufficiently small that time accuracy is
FIGURE 10: ROTOR 67 PERFORMANCE VS MASS FLOW maintained; in this study the time step is 2.0 × 10−5 (s), allowing
RATIO: EFFICIENCY AND TOTAL PRESSURE RATIO. about 30 time-intervals in the period of the highest frequency
mode considered.
In Fig. 12, we validate the accuracy of the ROM-CSM
model (system of Eqs. (17) and (7)), by comparing the first
3.2 FLUTTER ANALYSIS FOR ROTOR 67 three modal forces of rotor 67 blade. The aerodynamic force
The flutter characteristics depends on the structural proper- is obtained either by solving the full Reynolds-averaged Navier-
ties, in addition to the aerodynamic conditions. The material cho- Stokes equations or by using the ROM with 150 degrees of free-
sen for consideration is titanium alloy whose properties are given dom (i.e. the order of matrix Aa in the state space model) for
in Table 1, same as those used in [5, 22]. (The material in Doi’s an input defined by Eq. (23). It reveals that the third mode
work was altered to give the Young’s modulus of 1.422×1011 (Pa) (torsion) is most dominant and the weakest is the second mode
to place the first natural frequency away from the rotating fre- (second bending). The close agreement between the full and
quency or its double.) reduced models confirms the accuracy of the current Volterra-
The first three modal frequencies, calculated with the com- series-based ROM. Discrepancy is seen in the second mode, but
mercial software MSC/Nastran [12] on a 15×15 mesh, are listed this mode is less important than the other two.
in Table 2, along with the results by Doi. Despite using different The aeroelastic ROM system consists of 150 degrees of free-
values of Young’s modulus, these two predicted modal frequen- dom in the aerodynamic ROM and 6 in the structural represen-
cies are quite close. These three mode shapes are shown in Fig. tation (displacement and velocity), thus resulting in 156 DOF in
11; the first mode representing the bending, second mode the total for the entire AE ROM. The instability critical point can
second bending, and third mode the torsion. be determined by increasing total density at inlet boundary. As

10 Copyright 
c 2014 by ASME
 FIGURE 12: COMPARISON OF MODAL FORCE OBTAINED
BY THE FULL ORDER AND REDUCED ORDER SOLU-
TIONS.

(a) first mode: bending (b) second mode: second bending

(c) third mode: torsion

FIGURE 11: MODE SHAPES OF THE ROTOR 67 STRUC-

TURE AND INTERPOLATION BETWEEN THE STRUC-
TURAL AND AERODYNAMIC GRIDS. BLACK: NON-
DEFORMED GRID; BLUE: STRUCTURAL GRID; RED:
AERODYNAMIC GRID.
FIGURE 13: EIGENVALUES OF THE 156-ROM.

shown in Fig. 13, the 1st structural mode eigenvalue crosses the
imaginary axis, i.e., the eigenvalue becomes positive, indicating ified in Table 1 is determined to be structurally stable under the
an amplification of structural displacement. Figure 14 displays chosen operating condition, with a high margin of safety, when
the blade displacement predicted by the 156 aeroelastic ROM at only an isolated blade is considered, this finding consistent with
the flutter condition, the third and first modes are the two most that in [5,22]. However, blade row interactions, such as the effect
dominant ones while the second mode is nearly negligible. The of upstream inlet guid vane, can induce forced vibration in rotor
dynamic pressure needed to induce flutter is q∞ = 1.455x106 Pa, blade, thus altering its flutter characteristics, see study in [19].
nearly 10 times larger than the baseline operating condition at The ROM strategy presented here can also serve as an efficient
q∞ = 1.416x105 Pa. Hence the rotor made with the material spec- and reliable way of investigating the effect of inter-blade interac-

11 Copyright 
c 2014 by ASME
 tions. ACKNOWLEDGMENT
Finally we remark on the primary motivation of employing This paper presents part of the multidisciplinary design anal-
ROM, while under the foremost requirement of preserving ac- ysis and optimization (MDAO) effort contributing to the research
curacy. For performing an aeroelastic analysis over a complete towards next generation transport sponsored by the subsonic
sinusoidal cycle (Eq. (23)), the full-order (CFD-CSD) model fixed wing (SFW) project, under the fundamental aeronautics
takes 10.8 hours on a Xeon(R) W3530 computer with Intel(R) program (FAP) in NASA’s Aeronautics Mission Directorate. The
Compiler compared to 0.56 seconds used by the ROM, a whop- authors are grateful for the support of SFW management team.
ping savings by over 19,200 times. This shows the tremendous
value of using the ROM when searching for the flutter bound-
ary shown in Fig. 14, or when conducting design optimization, REFERENCES
both of which will otherwise require enormous computational [1] Bendiksen, O. O., 1986. “Role of shocks in tran-
resources. sonic/supersonic compressor rotor flutter”. AIAA J., 24,
pp. 1179–1186.
[2] Menter, F., 1994. “Two-equation eddy-viscosity turbulence
st
models for engineering applications”. AIAA Journal, 32,
1 mode
nd
pp. 1598–1605.
-6 2 mode
4.0x10 [3] Liou, M. S., 2006. “A sequel to AUSM, Part II: AUSM+–up
rd
3 mode
for all speeds”. J. Comput. Phys., 214, pp. 137–170.
-6
2.0x10
displacement(m)

[4] Yoon, S. K., and Jameson, A., 1988. “Lower-upper

symmetric-gauss-seidel method for the euler and navier-
0.0
stokes equations”. AIAA Journal, 26, pp. 1025–1026.
-2.0x10-6
[5] Doi, H., 2002. “Fluid/structure coupled aeroelastic compu-
tation for transonic flows in turbomachinery”. PhD Thesis,
-4.0x10-6 Stanford University, Stanford, CA.
[6] Moffatt, S., and He, L., 2003. Blade forced response pre-
0.00 0.05 0.10 0.15 0.20 dictions for industrial gas turbines, Part I:methodologies.
time(s)
ASME Paper GT2003–38640.
[7] Lucia, D. J., Beran, P. S., and Silva, W. A., 2004.
FIGURE 14: FLUTTER RESPONSE IN TERMS OF DIS-
“Reduced-order modeling: New approaches for compu-
PLACEMENT OF THREE MODES WHERE THE TORSON
tational physics”. Progress in Aerospace Sciences, 40,
AND BENDING MODES ARE DOMINANT AND THE SEC-
pp. 51–117.
OND BENDING MODE IS MINIMAL.
[8] Silva, W. A., 1997. Identification of linear and nonlin-
ear aerodynamic impulse responses using digital filter tech-
niques. AIAA Paper 1997–3712.
[9] Chima, R. V., and Liou, M.-S., 2003. Comparison of ausm+
and h-cusp schemes for turbo machinery applications. 16th
AIAA Computational Fluid Dynamics Conference AIAA
CONCLUDING REMARKS Paper 2003–4120, Orlando, FL, 23-26 June.
We have presented an accurate and efficient method for per- [10] Wilcox, D. C., 1988. “Reassessment of the scale-
forming aeroelastic analysis of a modern transonic compressor determining equations for advanced turbulence models”.
blade, NASA rotor 67. The CFD code, using κ -ω -SST tur- AIAAJ, 26, pp. 1299–1310.
bulence model and AUSM+ -up numerical fluxes, for providing [11] Liou, M.-S., 2006. “A sequel to AUSM, part II: AUSM+ -up
aerodynamic forces has been validated against measured data. for all speeds”. J. Comput. Phys., 214, pp. 137–170.
The structural motion based on finite element analysis is coupled [12] MSC Software Corp., 2011. MSC/NASTRAN user manual.
with fluid motion. the coupling is further modeled by the state- [13] Goura, G. S. L., 2001. “Time marching analysis of flutter
space representation to achieve considerable reduction in compu- using computational fluid dynamics”. PhD Thesis, Univer-
tational cost, while preserving the solution accuracy. The linear sity of Glasgow, Glasgow, UK.
state-space system is formulated by keeping only the first-order [14] Yao, W., and Liou, M.-S., 2012. Reduced-order model-
Volterra kernel. The obtained reduced order model is shown to ing for flutter/lco using recurrent artificial neural network.
be in excellent agreement with the full (original) model. Hence it AIAA Paper 2012–5446.
can be employed to provide an effective aeroelastic analysis tool, [15] Volterra, V., 1959. Theory of functionals and of integral
specifically for defining the flutter boundary. and integro-differential equations. Dover, New York.

12 Copyright 
c 2014 by ASME
[16] Raveh, D. E., Levy, Y., and Karpel, M., 2000. Aircraft
aeroelastic analysis and design using cfd-based unsteady
loads. AIAA Paper 2000–1325.
[17] Balajewicz, M., and Dowell, E., 2012. “Reduced-order
modeling of flutter and limit-cycle oscillations using the
sparse volterra series”. J. of Aircraft, 49, pp. 1803–1812.
[18] MathWorks Inc., 2012. Control system toolbox manual.
[19] Zhang, C., Ye, Z., and Liu, F., 2009. Numerical researches
on aeroelastic problem of a rotor due to igv/fan interaction.
AIAA Paper 2009–865.
[20] Bendiksen, O. O., 1990. Aeroelastic problems in turboma-
chines. AIAA Paper 1990–1157.
[21] Bartels, R., and Sayma, A., 2007. “Computational aeroe-
lastic modeling of airframes and turbomachinery: Progress
and challenges”. Phil. Trans. Royal Soc. A, 365, pp. 2469–
2499.
[22] Sadeghi, M., and Liu, F., 2005. Coupled fluid-structure
simulation for turbomachinery blade rows. AIAA Paper
2005–0018.
[23] He, L., and Denton, J. D., 1994. “Three-dimensional
time-marching inviscid and viscous solutions for unsteady
flows around vibrating blades”. J. Turbomachinery, 116,
pp. 469–476.
[24] Chuang, H. A., and Verdon, J. M., 1998. A numerical simu-
lator for three-dimensional flows throughou vibrating blade
rows. Nasa cr 1998-208511.
[25] Strazisar, A. J., et al., 1989. Laser anemometer measure-
ments in a transonic axial-flow fan rotor. NASA TP 2879.
[26] Cunnan, W. S., Stevens, S., and Urasek, D. C., 1978. De-
sign and performance of a 427-meter-per-second-tip-speed
two-stage fan having a 2.40 pressure ratio. Nasa tp-1314.

13 Copyright 
c 2014 by ASME