Progress in Aerospace Sciences: Mehdi Ghoreyshi, Adam Jirásek, Russell M. Cummings

Progress in Aerospace Sciences 71 (2014) 167217
Contents lists available at ScienceDirect
Progress in Aerospace Sciences

journal homepage: www.elsevier.com/locate/paerosci
Reduced order unsteady aerodynamic modeling for stability

and control analysis using computational uid dynamics
Mehdi Ghoreyshi n, Adam Jirsek, Russell M. Cummings
High Performance Computing Research Center, U.S. Air Force Academy, USAF Academy, CO 80840-6400, USA
art ic l e i nf o
a b s t r a c t
Article history:
Received 12 August 2014
Accepted 2 September 2014
Available online 3 October 2014
Recent advances and challenges in the generation of reduced order aerodynamic models using
computational uid dynamics are presented. The models reviewed are those that can be used for
aircraft stability and control analysis and include linear and nonlinear indicial response methods,
Volterra theory, radial basis functions, and a surrogate-based recurrence framework. The challenges
associated with identication of unknowns for each of the reduced order methods are addressed. A
range of test cases, from airfoils to full aircraft, have been used to evaluate and validate the reduced
order methods. The motions have different amplitudes and reduced frequencies and could start from
different ight conditions including those in the transonic speed range. Overall, these reduced order
models help to produce accurate predictions for a wide range of motions, but with the advantage that
model predictions require orders of magnitude less time to evaluate once the model is created.
Published by Elsevier Ltd.
Keywords:
Unsteady aerodynamics
Computational uid dynamics
Reduced order model
Indicial functions
Radial basis functions
Volterra functions
Contents
1.
2.
3.
4.
5.
6.
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.
Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.
The need for reduced order unsteady aerodynamics modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.
Current aerodynamic modeling efforts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Review of unsteady aerodynamic prediction methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.
Classical theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.
Volterra theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.
Indicial theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4.
Surrogate-based recurrence-framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5.
Radial basis functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Review of system identication methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.
System identication of CFD simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.
Volterra kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.
Indicial functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.
Training maneuvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Flow solver. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Validation test cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.
NACA 0012 Airfoil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.
SDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.
SACCON UCAV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.
X-31 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.
Airfoil aerodynamics modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.1.
Models for M0.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.2.
Models for M0.764 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Corresponding author.
E-mail addresses: Mehdi.Ghoreyshi@usafa.edu (M. Ghoreyshi), Adam.Jirasek@gmail.com (A. Jirsek), Russ.Cummings@usafa.edu (R.M. Cummings).
http://dx.doi.org/10.1016/j.paerosci.2014.09.001
0376-0421/Published by Elsevier Ltd.
168
168
168
168
170
170
171
172
173
173
174
174
174
174
176
177
177
178
179
182
183
188
188
188
195
168
M. Ghoreyshi et al. / Progress in Aerospace Sciences 71 (2014) 167217
6.2.
SDM aerodynamics modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3.
SACCON aerodynamics modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.4.
X-31 aerodynamics modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1. Introduction
1.1. Overview
The understanding and prediction of unsteady nonlinear aerodynamics attach great importance to advance aircraft aeroelasticity and ight dynamics. Maneuvering ight at high angles of
attack, fast motion rates, and gusty wind conditions, as well as
sudden control surface deections and utter phenomenon, all are
examples of unsteady ight conditions [1]. The unsteady aerodynamic forces and moments in these conditions can have a
signicant effect on aircraft aeroelastic and dynamic stability
characteristics [2]. In particular, ow hysteresis in a rapid maneuvering ghter can lead to pronounced time lags in the increase and
decline of aerodynamic loads with respect to the ow eld
changes [3,4]. The main sources of aerodynamic nonlinearities
are shock wave motions and separated ows which can generate
limit cycle oscillations and have a signicant impact on aircraft
performance and stability [5]. An aerodynamic model that can
accurately predict the nonlinear and unsteady airloads of a
maneuvering aircraft can improve the accuracy of the structural
analysis and ight control design. This translates into reduced
project risk and enhanced analysis of system performance prior to
rst ight.
A wide range of mathematical models have been developed to
represent the unsteady aerodynamic loads for use in aircraft
stability and control (S&C) analysis, including classical linearized
models using rotary and unsteady aerodynamic derivatives (the _
terms) [6]. The unsteady derivatives of static and weakly nonlinear
systems can be determined by a Taylor series expansion about
given equilibrium ight conditions. However, these derivatives are
not suitable for analysis of a high performance aircraft with highly
nonlinear and unsteady aerodynamics [7,8]. There are also only
limited experimental measurements available to determine the
effects of unsteady ow on the aerodynamic forces and moments
acting on an aircraft. This is mainly due to the complexity of
unsteady ow and the limitations of existing test facilities [8].
During the past decade, efforts have been made at formulating
techniques that predict continuous and discrete-time aerodynamic
responses using Computational Fluid Dynamics (CFD) [9,10]. These
recent advances have led to the generation of unsteady nonlinear
aerodynamic models for six-degree-of-freedom (6DoF) aircraft
maneuvers with moving control surfaces. This work reviews the
current state-of-the-art in unsteady aerodynamic modeling for use
in S&C and discusses current challenges for these modeling
approaches.
1.2. The need for reduced order unsteady aerodynamics modeling
With advances in computing techniques, one straightforward
way to predict ow-induced nonlinearities is to develop a fullorder model (often called a time-marching CFD solution) based on
time integration of the structural and aerodynamic equations
[1113]. However, creating a full-order model for S&C is
a computationally expensive approach and is impractical from a
designer's point of view because it requires a large number of
199
201
209
213
215
215
computations for different values of motion frequency and amplitude [14,15]. Also, this approach makes the solution of the aircraft
equations of motion an innite-dimensional problem, where the
current states depend on the evolution of previous states at
innitely many points in time [16]. It is therefore desirable to
use a relatively simple but accurate approximation for the
unsteady aerodynamics by using a Reduced Order Model (ROM)
that allows describing the unsteady ow in the form of a small
number of spatial/temporal modes (typically less than one hundred) compared with the very large number of grid points in the
full-order model (on the order of 550 million or more) [1719].
The ROM can then predict the responses to an arbitrary input
much faster than computing the input in a full CFD solution [20].
The creation of ROMs using CFD is an area of active research.
These models can be classied into nonparametric and parametric
models depending on the identication methods used [21,22]. The
parametric models provide a structure for representing aerodynamic forces and moments in the aircraft equations of motion.
Nonparametric models, on the other hand, are concerned with the
measured inputoutput behavior of the aircraft dynamics [21].
ROMs can also be classied into time domain and frequency
domain approaches [9]. While time-domain models are necessary
for rotary wing applications and predicting responses to a timedependent input, the formulation in frequency-domain is required
for determination of utter instability boundary [23]. Some
popular ROMs are indicial response methods [2427], proper
orthogonal decomposition (POD) [28,29], Volterra theory [22,3032],
surrogate-based recurrence framework (SBRF) [33,34], radial basis
functions (RBF) [3537] and state-space modeling [38,39]. Notice that
some of these models, such as indicial response methods, Volterra
theory, and POD analysis, can be performed either in the time or
frequency domain. Recent applications of POD methods has been
reviewed by Lucia et al. [9] and therefore these methods are not
discussed here. Also, the state space model proposed by Goman and
Khrabrov [38] uses experimental data to estimate the model
unknowns. There are only a few published studies of this model using
CFD and therefore the model is also not discussed. This paper reviews
the recent development and applications of Volterra theory, linear and
nonlinear indicial response methods, RBF, and a SBRF model using CFD
data. These models were selected because they have been used or can
be used in the aircraft S&C analysis.
1.3. Current aerodynamic modeling efforts

Early numerical models of unsteady aerodynamics were mostly
developed for ows in cascades and turbomachineries and were
based on the Fourier series decomposition of the inputs and
responses [4044]. These models can be performed either in the
time or frequency domain. However, for time-periodic ows, such
as ows about helicopter rotors and ows in turbomachinery, the
frequency-domain models can dramatically reduce the computational time compared with time-domain models [45]. The recent
modeling efforts of the Fourier-based methods were published in a
special section of the International Journal of Computational Fluid
Dynamics [4550] and have also been reviewed by He [51].
Nomenclature
A
a
C(k)
CL
CL0
C L
CLq
Cl
Cm
Cm0
C m
Cmq
CN
Cn
CY
c
f
H
h
k
L
M
P
p
q_
q
q1
unit step function

acoustic speed, m/s
Theodorsen function
lift coefcient, L=q1 S
zero-angle of static lift coefcient
lift coefcient derivative with angle of attack, 1/rad
lift coefcient derivative with normalized pitch rate, 1/
rad
roll moment coefcient
pitch moment coefcient
zero-angle of attack pitch moment coefcient
pitch moment coefcient derivative with angle of
attack, 1/rad
pitch moment coefcient derivative with normalized
pitch rate, 1/rad
normal force coefcient
yaw moment coefcient
side force coefcient
mean aerodynamic chord, m
frequency, Hz
plunging motion amplitude, m
plunging motion vertical placement, m
reduced frequency, c=2V
lift force, N
Mach number, V/a
pressure, Pa
normalized roll rate, 1/rad
time rate of pitch rate, rad/s2
normalized pitch rate, 1/rad
freestream dynamic pressure, Pa
For xed wing aircraft, aerodynamic damping derivatives are

usually used to represent unsteady effects. However, a derivativebased model breaks down for modern combat aircraft undergoing
agile maneuver at high angles of attack [8]. A wide range of efforts are
emerging to extend the application of these models to advanced
ghter aircraft. For example, the work of Morton and his colleagues
[5255] at the Air Force SEEK EAGLE Ofce (AFSEO) involves aerodynamic modeling for aircraft ight simulation, control system design,
and optimization. They begin with a global nonlinear parameter
modeling technique proposed by Morelli [56] that describes the
functional dependence between a motion and its computed aerodynamic response in terms of force and moment coefcients. They
then attempt to nd a new model which has adequate complexity to
capture the aerodynamic nonlinearities while keeping the number of
terms in the model low. The latter requirement improves the ability to
identify model parameters, resulting in a more accurate model with
good predictive capabilities. The model unknowns are then identied
using CFD simulations of some training maneuvers. These techniques
have been incorporated into the Department of Defense (DoD)
CREATE-AV code, Kestrel.
Various reduced order unsteady aerodynamic models have
also been used at the University of Michigan. Skujins and Cesnik
[5759], for example, predicted unsteady loads due to the oscillation
of multiple vehicle elastic modes using convolution of modal step
responses along with a correction factor which takes into account
different oscillation amplitudes, ght conditions, and nonlinear
aerodynamic effects due to multi-modal oscillations. Glaz et al.
[33] also described a reduced-order nonlinear unsteady aerodynamic modeling approach suitable for analyzing pitching/plunging
airfoils subject to xed or time-varying freestream Mach numbers.
Re
r
S
St
s
t
tn
t0
V
ut
u; v; w
x
Y
y
169
Reynolds number, Vc=

normalized yaw rate, 1/rad
reference area, m2
Strouhal number, St 2fH=V
normalized time, 2Vt=c
time, s
non-dimensional time step, Vt=c
start time, s
freestream velocity, m/s
internal state vector
velocity components in X, Y and Z-axis, m/s
input vector
side-force, N
output vector
Greek
_
A
0
e
t
t
normalized time rate of angle of attack, 1/rad

angle of attack, rad
pitch amplitude, rad
mean angle of attack, rad
effective angle of attack, rad
side slip angle, rad
circular frequency, rad/s
Wagner function
Kssner function
density, kg/m3
pitch angle, rad
pitching motion amplitude, rad
air viscosity
The reduced-order model uses Kriging surrogates to account for

ow nonlinearities and recurrence solutions to account for timehistory effects associated with unsteadiness. They showed that the
model can accurately predict time-varying freestream Mach number
effects and is therefore applicable to rotary wing and xed wing
applications. Later, they used this modeling approach for the
prediction of unsteady lift, pitch moment, and drag of an airfoil
pitching/plunging under dynamic stall conditions [60].
The aerodynamic modeling studies at Duke University are mainly
focused on frequency domain techniques of Harmonic Balance (HB),
POD-based models, and Volterra series, all applied to different aeromechanics problems including utter, forced response, limit cycle
oscillations, multistage ows, and non-synchronous vibrations. The HB
approach assumes that unsteady ows are temporally and spatially
periodic and then ow variables are represented by a Fourier series in
time with spatially varying coefcients [61]. The full data at each grid
point is then reduced to 2N1 coefcients, where N is the number of
harmonics retained in the Fourier series representation of the ow.
The two other models used at Duke are based on POD and Volterra
series. For example, Thomas et al. [62] applied POD-based models to
an inviscid ow transonic wing conguration and reduced the CFD
model with over three quarters of a million degrees of freedom to a
system with just a few dozen degrees of freedom, while still retaining
the accuracy of the unsteady aerodynamics of the full system
representation. Also, the Volterra series approach by Balajewicz and
his colleagues [63,64] was investigated for utter boundary and LCO
analysis. They used a pseudo-inverse approach and training inputs
from the full-order system to identify Volterra kernels.
At Stanford University, POD methods have been used
for unsteady aerodynamic analysis by Farhat and his colleagues.
170
For example, Lieu and Farhat [65] proposed a ROM based on POD
for evaluating aeroelastic frequencies and damping ratio coefcients of the F-16 aircraft. They extended previous ROMs to a
multiple state variable model by including the variation of Mach
number and angle of attack ight conditions. Jameson and his
colleagues at Stanford have also applied HB techniques to accelerate the convergence of unsteady ows to a periodic steady state
[66,67]. The HB techniques have also been successfully tested by
Badcock and his colleagues at the University of Liverpool [68,69].
Time domain based neural networks have also been used to
model unsteady aerodynamics [7073]. Olcer et al. [70] used neural
networks for modeling the unsteady aerodynamic responses to a
trailing edge ap deection and to extract a linear time-invariant
model of the ap actuator. Marques and Anderson [71] used a
temporal neural network to approximate the unsteady lift and pitch
moment of a two dimensional airfoil with changes in angle of attack
in the transonic regime. Faller and Schreck [73] also used a neural
network to predict the time-dependent surface pressures of a
pitching wing. The network inputs included the instantaneous pitch
angle, angular velocity, and the initial surface pressure coefcients
at t0. The network output then predicted the surface pressures at
time t 0 t. These predictions were fed back as the input to the
network to predict surface pressures at the next time. This process
continues in time until it reaches some nal time.
Some other works to be mentioned include the studies at the
University of Maryland [7476,76], the work of Raveh [77,78,20] at the
Israel Institute of Technology, the work of Beran [9,79] at the Air Force
Research Laboratory, the work of Silva [30,22] at NASA Langley, and
the work of Murman [80,81] at NASA Ames. Finally, researchers at the
U.S. Air Force Academy (USAFA) have made a strong effort toward the
development of unsteady aerodynamic reduced order models applied
to ight mechanics and ow control.
2. Review of unsteady aerodynamic prediction methods

The input to, and output from, a model are dened by x(t) and y
(t), respectively. Note that for a dynamic system (a system with
memory) y is said to be a functional of x, not the function of x [82].
2.1. Classical theories
Classical solutions have been obtained for the airfoil lift and
pitch moment responses to harmonic pitching and plunging
motions and wind gusts by Theodorsen [83], von Karman [84],
Kssner [85], Sears [86], and Wagner [87]. Pitching and plunging
oscillations are shown in Fig. 1 and are described by
t A cos t;
ht H cos h t
where A and are the pitch amplitude and rotational velocity,

respectively; h is the vertical placement of the airfoil, H is the plunge
amplitude, and h denotes the oscillatory frequency. The reduced
frequencies of k and kh are dened as c=2V and h c=2V , where c
is the airfoil chord and V is the freestream velocity. The plunge
motion has no rotation but the angle of attack changes due to the
vertical displacements of the airfoil; this angle is named the

effective angle of attack, which is denoted by e and is dened as
!
h_
e tan 1
2
V
The maximum effective angle of attack for a plunge oscillation
starting at zero degrees angle of attack is determined by the Strouhal
number (dened as St 2fH=V ) such that max
tan 1 St. The
e
Strouhal number is a dimensionless number that is useful for
describing unsteady ow of oscillating geometries [16]. There are a
number of theories to predict the aerodynamic loads of pitching and
plunging motions. Wagner [87] developed a theory based on
potential ow, and detailed the unsteady lift coefcient of an airfoil
undergoing a plunging motion as
Z t
C L 2
t _ e d
3
0
where _ e de =dt; CL is the lift force coefcient and t is Wagner

function with its exact values known in terms of Bessel functions.
The Wagner function was approximated in non-dimensional time by
a two-pole exponential function as [75]
s 1 0:165 exp 0:0455s 0:335 exp 0:3s
where s 2Vt=c is a non-dimensional time. Wagner's approximation

function is shown graphically in Fig. 2(a) which shows that the
unsteady lift increases asymptotically in time to reach the quasisteady value at one degree angle of attack, while it reaches almost
90% of this value after traveling a distance equivalent to 15 semichords [88].
Lomax [89] used the linearized Euler equations to derive exact
initial values of Wagner function for a at-plate airfoil as

1 4
1M
s
1
s
5
2 M
2M
where M is the Mach number. Note that Lomax's theory is valid
only for the initial time of plunging motions. Theodorsen [83],
assuming potential ow, used the method of conformal mapping
to derive the solution of unsteady aerodynamic loads of a thin
airfoil undergoing small pitching and plunging oscillations. His
theory was expanded by Hodges and Pierce [90] for the lift and the
pitch moment about the quarter-chord as
!
"
#
c_
ch c_ ac2
h_
C L 2
2
6
Ck
2V
2 V2 V
V
V
Cm
"
#
_ 1 ac2
1 ch c
2 2 V2 V
8 2 V2
where a is the pitch axis location with respect to the half-chord as

shown in Fig. 1. In this denition, the leading edge and trailing
edge correspond to values of a 1 and a 1, respectively. In
Eq. (6), the rst of the two terms in the lift shows the circulatory
lift and the second term shows non-circulatory effects. Noncirculatory effects, also called added mass, are generated from
motion accelerations. Circulatory effects, on the other hand, are
generated from the differences in the velocity on the upper and
lower surfaces of the airfoil. For the airfoil pitch moment about the
quarter-chord the non-circulatory terms are usually negligible.
The effective angle of attack is dened as
"
#
h_
c_
e Ck
8
V
2V
where C(k) is called Theodorsen function and is given by
Ck
Fig. 1. Airfoil pitch and plunge maneuvers.
H 2
1 k
2
H 1 k iH 2
0 k
171
Fig. 2. Unsteady aerodynamic force functions for incompressible ow over an airfoil. (a) Wagner and Kssner functions and (b) Theodorsen's functions.
where H0 and H1 are Hankel functions of the rst and the second
kind. The Theodorsen function is real and equals one for the steady
case (i.e. k 0) and includes an imaginary and a real part for a
moving airfoil. These parts are expressed as Ck Fk iGk and
are shown in Fig. 2(b) for reduced frequencies up to two. The gure
shows that F(k) decreases with reduced frequency, thus the phase
lag between the motion variables and unsteady loads is increased.
The Theodorsen function is related to the Wagner function as [91]
Z 1
Ck ik
e ik d
10
0
yt 1 F xt; t 1 T rt r t 1
and is approximated as
Ck 1
Taylor series also apply to the Volterra series [94], e.g. the Volterra
series typically is accurate for weak nonlinearities [95]. The theory
of functional expansions was rst introduced by Volterra in a book
published in 1930 [92]. This work and the further developments
by others (for example [9698]) have been extensively used in
electrical and biological systems engineering [99103]. Recently,
there is an increasing interest in using Volterra series in the eld of
nonlinear unsteady aerodynamic loads modeling [10,104].
For a dynamic system, the output in a given time instant of t1
depends on the input values over some nite interval [82], i.e.
0:165ik 0:335ik

ik 0:0455 ik 0:3
11
Also, Kssner [85] dened a Wake Vortex Function, later named

the Kssner function, t, to nd the unsteady lift of an airfoil in
response to a sharp-edged gust with an upwash velocity of g given
as
Z s
C L 2=Vg 0 s
s _ g d
12
0
where _ g dg =dt. The Kssner function can be related to Wagner

function using a Fourier-integral and was approximated in terms of
the Jones exponential expansion as
s 1 0:5 exp 0:15s 0:5 exp s
13
Fig. 2(a) compares Kssner and Wagner functions; it shows that

Kssner function is zero at s0 and reaches unity for very large s
values. While Kssner function predicts the responses to a sharpedged gust, Sears [86] derived the frequency response to a sinusoidal gust as well. Note that these theories are limited only to
potential ow and small-amplitude airfoil motions. For a modern
combat aircraft maneuvering at high angles of attack, a full nonlinear, large-amplitude, unsteady aerodynamic model is required.
2.2. Volterra theory

The Volterra theory [92] is considered one of the most
important tools for the representation of nonlinear dynamic
systems in the time and frequency domains [93]. The approach
is a functional Taylor series expansion for dynamic systems (often
called functional expansions) [94], therefore, the limitations of
14
where F is a continues functional of the input. Volterra used a p-th

order series to estimate the values of y in a given time as
p
yt H0 Hi xt
15
i1
where H0 is the steady-state term and is constant; Hi represents

the i-th order Volterra operator. In Eq. (15), p may be indenite to
make sure that y converges to the exact solution. However, it has
been shown that for a time-invariant system, the output of y for
each t Z 0 can be approximated by the right side of Eq. (15) for
sufciently large but nite values of p as well [82]. For such a
system, the Volterra operator is dened as an i-fold convolution
between the input, xt, and the i-th order Volterra kernel, Hi, i.e.
Z t
Z t
i
H i t 1 ; t 2 ; ; t i xn dn
16
Hi xt
1
1
n1
and using the commutative law, it can also be written as

Z t
Z t
i
H i 1 ; 2 ; ; i xt n dn
Hi xt H0
1
1
17
n1
Korenberg and Hunter [82] mentioned some Volterra kernel

characteristics: (1) the zero order Volterra kernel of H0 is a
constant value and found from the zero-input response of the
system, (2) for a casual system the kernel of H i 1 ; 2 ; ; i 0
when any of 1 ; 2 ; ; i is less than zero, and the lower limits of
the integrals can be set to zero, and (3) it can be assumed that each
kernel is symmetric with respect to any permutation of 1 ; 2 ; ; i .
The Volterra series approximation can be used to represent any
continuous-time, time-invariant, nite memory, and single-input
and single-output system. The nonlinear unsteady aerodynamics is
considered such a system (ignoring those hysteresis effects that
make the solution innite-memory) except that it is a multi-input
172
and multi-output system. For a forced oscillation about the pitch

axis, the output function y is dened as y C L;D;m t (corresponding to lift, drag, and pitch moment) and the input vector includes
_
xt t; qt; qt
18
where q and q_ denote normalized pitch rate and its rst time
derivative, respectively. Classical Volterra series cannot properly
identify any nonlinear cross-coupling effects between multiple
input parameters [5]. Balajewicz et al. [105] applied a multi-input
Volterra theory to the weakly nonlinear aerodynamic modeling of
NACA 0012 airfoil under pitching and plunging motions. Their
results showed a signicant improvement in the accuracy over a
single-input Volterra model. Also, Ghoreyshi et al. [34] proposed a
multi-input Volterra series for aerodynamics prediction of the
X-31 aircraft pitching in the transonic speed range. This series is
written as
p
yt x1 t; x2 t; ; xm t Hm
i
19
i1
where the term Hm

i is the multi-input Volterra operator dened as
a mp-fold summation of p-fold convolution integrals between the
inputs and the p-th order multi-input Volterra kernels [105]. The
output response up to second order is rewritten as
Z t
m
x
yt H0
H 1j t xj d
j1
j1 1 j2 1
1
t
1

O jxj3
t
1
xj ;xj2
H2 1
t 1 ; t 2 xj1 1 xj2 2 d1 d2
20
Note that the superscripts in Eq. (20) identify to which inputs

the kernel corresponds, for example, the second-order kernel
xj ;xj
H 2 1 3 correlates the inputs xj1 and xj3 . Also, it should be noted
that the second and higher-order kernels are symmetric with
xj ;xj
xj ;xj
respect to the arguments, H 2 1 3 H 2 3 1 . To estimate the output
of y using a Volterra series, the kernels of Hm
i must be determined
using a system identication method. Unfortunately, the kernels of
a nonlinear dynamic system (third- and higher-order kernels) are
very difcult to compute [22].
2.3. Indicial theory

The Heaviside response to a unit step change in input function
(angle-of-attack, pitch rate, etc.) is the so-called indicial response.
Assuming a linear relationship between the input function and the
output, a linear ROM is dened as the convolution (or Duhamel's
superposition [106]) of the indicial response with the derivative of
the forcing function [107]
Z t
_
yt x0At
At x
d
21
0
where x_ dx=dt; A represents the unit response, or indicial

response, of the system and t is elapsed time; x is input or
forcing parameter and x0 shows input value at time zero.
The indicial response functions are used as a fundamental
approach to represent the unsteady aerodynamic loads. For aerodynamic loads modeling, the mathematical models are detailed by
Tobak et al. [25,26] and Reisenthel et al. [108,109]. It is assumed
that lift and pitch moment depend on angle of attack and pitch
rate. Indicial responses due to a unit step change in angle of attack,
, and normalized pitch rate, q, are denoted by C j and C j q ,
respectively, where C j C L ; C m representing lift and pitch
moment coefcients (or any other aerodynamic coefcient). The
unsteady lift and pitch moment are then calculated by adding
Duhamel's integrals with respect to and q, i.e.

Z t
Z t
_
_
C j t C j 0 0C j t
C j t
d
C jq t q
d
0
22
where C j 0 denote the zero-angle-of-attack lift and pitch moment
coefcients and are found from static calculations; 0 is the angle
of attack at time zero or the initial time of motion. Note that
0C j t is different from C j 0 . The motions considered here start
from a steady-state solution with q0 0 and therefore the term
of q0C jq t was not added to the equation. Using the differential
theorem of the convolution integral [110], Eq. (22) changes to
Z t

Z t

d
d
C j t d
C jq t q d
23
C j t C j 0
dt 0
dt 0
This equation predicts lift and pitch moment responses in the
linear regime of a ow. A nonlinear model is also considered
where the responses in the angle of attack depend on both the
angle of attack and Mach number. Also, it is assumed that the
response functions with respect to the pitch rate change with
changes in freestream Mach number but do not vary with the
changes in angle of attack at low to moderate angles of attack. This
is a reasonable assumption based on aerodynamic prediction
methods described in aircraft design textbooks [111]. The unsteady
nonlinear lift and pitch moment coefcients at time t are then
obtained using nonlinear indicial response theory as
Z t

d
C j t C j 0 M
C j t ; ; M d
dt 0
Z t

d
C jq t ; Mq d d
24
dt 0
where M denotes the freestream Mach number. The response
functions due to pitch rate, i.e. C jq t; M for j L; m, can be estimated
by using a time-dependent interpolation scheme from the observed
responses. The pitch rate indicial functions are next used to
estimate the second integrals in Eqs. (23) and (24), however, the
estimation of nonlinear C j t; ; M for j L; m needs more explanation. Assume a set of angle of attack samples of 1 ; 2 ; ; n at
freestream Mach numbers of M M 1 ; M 2 ; ; M m , where the spacing can be uniform or non-uniform. The indicial functions at
each angle of i ; i 1; 2; ; n, degrees and Mach number of
M k ; k 1; 2; ; m, are calculated by holding the angle of attack
xed at i degrees, and then performing a small step in the
angle of attack to i . The indicial functions are then
computed by taking the differences between time-varying
responses occurring after the step and the steady-state solution at
i degrees, and dividing them by the magnitude of the step.
Ghoreyshi and Cummings [112] extended the indicial response
method to include lateral aerodynamic coefcients as well.
Assuming that the lateral loads only depend on side-slip angle
(), normalized roll rate (p), and normalized yaw rate (r), the
unsteady lateral forces and moments using indicial functions are
written as
Z t

Z t

d
d
C j t
C j t ; ; M d
C jp t ; Mp d
dt 0
dt 0
Z t

d
C t ; Mr d
25
dt 0 jr
where j Y; l; n denote the side-force, roll and yaw moments,
respectively. The functions of C L t; ; M, C m t; ; M, C Lq t; M,
C mq t; M, C Y t; ; M, C l t; ; M, C n t; ; M, C Yp t; M, C lp t; M,
C np t; M, C Yr t; M, C lr t; M, and C nr t; M are unknown and the
methods for calculating them in CFD will be reviewed. Note that
the integral operation in the indicial theory allows for predicting
airload responses from time-histories of motion. However, similar
to the Volterra series, the indicial theory is typically accurate for

weak nonlinearities.
2.4. Surrogate-based recurrence-framework
The set of nonlinear equations describing the CFD system can
be viewed as a multi-input/multi-output dynamic system with a
mathematical model in state space given by an ordinary differential equation of the form
u_ t f ut ; xt
26
27
with given initial condition u u0 for t t 0 and

u t A Rr ;
yt A Rn ; xt A Rr
where u is a r-dimensional vector of internal state variables over

the eld R of real numbers, x is a vector of the inputs to system,
and y is m-dimensional vector of system outputs. For a discretetime dynamical system, the equations change to
uk 1 f uk ; xk
yk huk for k 0; 1
28
where k is an integer value showing discrete time values. The state

function f is a smooth function that maps the current state uk and
the input xk into a new state uk 1 , and the output function h maps
the state uk into the output yk given as [113]
f : Rr Rn Rr ;
h : Rr Rm
In this system, the outputs can be determined from the states at

time instant k only, therefore the past history of the system is
irrelevant [114]. If the state variables are directly measured, the
functions f and h can be approximated using neural networks or
surrogate-based models. However, in many practical situations,
measuring all state variables is limited.
Referring to unsteady aerodynamic problem during pitching
and plunging motions, the discretized governing equations of uid
dynamics serve as the state space functions and the input vector of
x takes the form of
_
xt t; t;
t
29
also, the output vector is dened as y C L;D;m t. In this system, the

internal state vector includes ; p; u; v; w; E that correspond to the
values of density, pressure, velocity components, and energy at
each grid point. This large amount of data makes the identication
of Eq. (28) a very complex task. Fortunately, there are methods
available that allow us to reconstruct the state space model by
mapping only the input and output data.
For a nite-time interval and a system described by Eq. (28),
Levin and Narendra [115] used the Implicit Function theorem [116]
to write the output vector at any instant as a function of the past n
values of the inputs and the past m values of the outputs, i.e.,
yk 1 yk ; yk 1 ; ; yk m 1 ; xk ; xk 1 ; ; xk n 1
are assumed to be cheaper to measure than primary data and

hence they are available at many combinations of the input vector,
x~ t . An additional mapping is then dened between this system
input and the secondary data, i.e.
y~ k 1 ~s y~ k ; y~ k 1 ; ; y~ k m 1 ; x~ k ; x~ k 1 ; ; x~ k n 1
31
The mapping in Eq. (31) is then used to predict secondary output

values at the n past values of the primary input vector, x. The
function in Eq. (30) is then augmented by this evaluation in the
form of
yk 1 t yk ; yk 1 ; ; yk m 1 ; xk ; xk 1 ; ; xk n
and an output equation as

yt hut
173
30
In Eq. (30), is a vector-valued nonlinear function that maps the

inputs to the outputs, and m and n are integer numbers representing the past values in the output and input, respectively. Eq. (30)
preserves the characteristics of the state-space model but no
longer depends on system internal states. The inputoutput
mapping, , can be learned using neural networks when the time
histories of input and output are available. This network is often
named a recurrent neural network [114], where the network
output becomes part of the next input vector [72].
Ghoreyshi et al. [36] extended the mapping function in Eq. (30)
to allow for primary and secondary measurements. Assume that
secondary measurements of y~ t A Rm are also available. These data
1; y~ k ; y~ k 1 ; ; y~ k n 1
32
such a model brings the information of secondary data into the

mapping function and aids in reducing the computational cost.
Central to the generation of the reduced-order model is the
computation of the function in Eqs. (30)(32). Without a
closed-form analytical expression, a numerical approximation of
is constructed using a number of CFD solutions. For a pitching
aircraft, any motion can be expressed as a functional of three
parameters, e.g., 0, A , and k, that represent the mean angle of
attack, amplitude, and reduced frequency, respectively. These
independent variables form a parameter space, which represents
the envelope of all possible ow conditions in which the aircraft
conguration is expected to operate. To generate a consistent set
of unsteady aerodynamic loads in response to a given aircraft
motion time history, the training cases at which CFD solutions are
calculated should be representative of the expected ow conditions. Several Design of Experiment methods are available in the
literature including the Kriging-based framework [117]. Let NT be
the number of training cases at which CFD solutions are available.
Each training case consists of different combinations of the
independent parameters,
xi i t; _ i t; i t;
for i 1; ; NT
33
and the corresponding aerodynamic loads are indicated by yi t.

The approximation of the function is obtained by interpolating
the sampled data in the form of an input/output relationship.
Kriging interpolation is a common interpolation method, but for
increasing number of independent parameters the problem can be
ill-conditioned. An alternative approach is the multi-linear interpolation technique, which is in general faster than the Kriging
interpolation.
2.5. Radial basis functions
The RBF model used is similar to SBRF because they are both
created by the inputoutput mapping function of given in Eq.
(30), except that the function is now approximated through RBF
neural networks from a set of training data. The RBF network
provides an approximation of the functions based on the location
of data points, and is generally much faster than multi-layer feedforward neural networks [114]. Given an input vector of
fX cj : j 1; ; pg; X cj A R and a corresponding output vector of
fY cj : j 1; ; pg; Y cj A R, the RBF approximates the output at a
new given point as
P
Y^ X n k i X
34
k1
such that
Y^ X cj Y cj
for j 1; 2; ; p
35
where k are the weights of the linear combiners. The functions i

are named Radial Basis Functions and are often described by a
174
Gaussian basis function as

!
J X n X cj J 2
i X exp
2
36
where is a real variable to be chosen by the user, and J J

denotes the Euclidean norm such that the functions i will vanish
at sufciently large values of J X n X cj J . In terms of the network
structure, the RBFNN is a two-layer processing structure with one
hidden layer that approximates i at each node. Then, the output
layer is a set of linear combiners of approximation from hidden
layer nodes. The network is then trained to minimize the error
between the target (desired) values and the network predicted
values.
estimated kernels are very sensitive to the inaccuracies in dening

an impulse function in CFD. Also, a highly nonlinear system needs
higher-order kernels and the impulse approach is mainly limited
to the second-order kernel identication. A way to eliminate the
need for impulse simulations is the indirect identication method.
Tromp and Jenkins [122] used CFD-calculated indicial
responses to identify the rst-order kernel of an airfoil undergoing
a pitch oscillation. For higher-order kernels, Balajewicz and Dowell
[64] proposed a pseudo-inverse approach, where values of aerodynamic coefcients and the time-history of the motion variables
are known from the CFD simulation and are used as a training
input. The CFD solution is discrete in time, and the time step is
indicated by t. Let x denote as xt xnt xn. The discretetime representation of Eq. (20) can be rewritten as
m
3. Review of system identication methods

3.1. System identication of CFD simulations
yn H 1j n kxj k
j1 k0
m
j1 1 j2 1 k1 0 k2 0
System identication (SID) refers to the process of constructing

a mathematical model to describe the relationship between input
and output data of a system under testing [52]. The mathematical
model structure can take various forms depending upon the
intended use, i.e. what is assumed to be known, and what is to
be computed. Usually wind tunnel and ight-test data are used to
obtain accurate and comprehensive mathematical models of aircraft aerodynamics for aircraft ight simulation, control system
design and evaluation, and dynamic analysis [118,56]. Recently,
aircraft system identication has been used in cooperative
approaches with CFD to take advantage of the strength of both
approaches or having one approach ll in the gaps where the
other cannot be used [52,53]. The wide range of SID tools that have
been developed for aircraft system identication (including modeling techniques proposed by Morelli [56]) can easily be used to
analyze CFD data computed for aircraft in prescribed motions. The
modeling effort is potentially global because the independent
_ , etc.) are varied over a large range. These globally
variables (, ,
valid models and their associated smooth gradients are useful for
optimization, robust nonlinear control design and global nonlinear
stability, and control analysis.
3.2. Volterra kernels
The Volterra kernels need to be determined for creation of a
Volterra-based ROM. Several methods exist to identify these
kernels in the time and frequency domains. The techniques based
on a continuous time impulse are proposed by George [119] and
Schetzen [120] to estimate these functions. Silva [31] also used a
method based on the impulse functions to directly calculate rstand second-order kernels using CFD. In his approach the rstorder kernel is a combination of the response to unit and double
unit impulses at time t 1 T. The second-order kernel is a combination of two successive unit impulses at time t 1 T and
t 2 T t and two unit pulses, one at time T, and a second at
time T t. Although this approach is straightforward, it faces
some challenges as well. The main challenge is that an impulse
response is a mathematical concept and has no exact equivalent in
experimental or numerical studies [121]. The numerical simulation
of an impulse is very complicated as the impulse occurs over a
very short period of time (theoretically for t-0). A computational solution is to use a discrete time impulse in CFD. However,
the functions found using these simulations depend on the time
discretization used and are not directly related to Volterra kernels.
Milanese and Marzocca [121] used a distributed impulse instead,
however the simulation of these impulse functions strongly
depends on the time step used. Raveh [20] showed that the
xj ;xj2
H2 1

n k1 ; n k2 xj1 k1 xj2 k2
37
The identication of discrete-time Volterra kernels involves the

resolution of an overdetermined system. Values of aerodynamic
coefcients and the time-history of the motion variables are
known from the CFD simulation and are used as a training input.
Let y y0; y1; ; ynT denote each aerodynamic load computed using CFD, and let A contain the permutations of input
parameters relevant to the unsteady motion. Eq. (37) can be recast
in the form of
y Ab
38
where the vector b contains the unknown Volterra kernels. The

matrix A is in general non-square, with more rows than columns.
Several numerical methods are available to solve least squares
problems, e.g., direct inversion of AT A, Gaussian elimination,
MoorePenrose generalized inverse approach and QR factorization. However, the MoorePenrose approach and QR factorization
are more accurate than Gaussian elimination and the
direct
inversion solutions. The cost of QR factorization is O n2 , and

the MoorePenrose inversion involves O n3 operations. Note that
computational resources attributable to the identication of the
Volterra kernels grow exponentially with order. Increasing the
order of the Volterra series introduces a requirement for a training
maneuver of sufcient duration. A remedy to this is the use of a
simplied form of the kernel parametric structure. For example,
Balajewicz et al. [105] proposed to set all off-diagonal terms of the
kernel to zero, i.e.

xj ;xj ;;xjp
Hp 1 2
n k1 ; n k2 ; ; n kp 0
for k1 a k2 a a kp
39
The Volterra kernels are then identied from Eq. (38) solving for b,
with y and A being known for a training maneuver. The matrix A is
then recomputed for a novel maneuver, and the low-order model
in Eq. (38) is used to predict the resulting unsteady aerodynamic
loads in place of the full-order system.
3.3. Indicial functions
The indicial functions can be estimated via analytical or
computational methods. The experimental tests are practically
nonexistent for impulse and step response functions. Wagner [87]
was the rst to solve the indicial lift response of a thin airfoil due
to a step change in the angle of attack in incompressible ow, the
so-called Wagner function. His function is a single parameter
function and is sufcient to model the plunging motions of a thin
airfoil in incompressible ows.
175
Fig. 3. Fourier transform relationship between oscillatory and indicial aerodynamics [124].
Since Wagner's original work, there have been numerous

contributions to obtain the approximations of response functions
in compressible ow as well; many of these solutions are reviewed
by Bisplinghoff et al. [123]. The most common way of calculating
the indicial functions in compressible ow is by using harmonic
(oscillatory) motions [76]. The Wagner function ( in Eq. (3)) and
Theodorsen function (C(k) in Eq. (9)) are related through a Fourier
integral or Laplace transform as shown in Fig. 3, because both
solutions are based on the same ow equations assuming the
same boundary conditions [124,125]. Similar relations can be
developed for the compressible indicial responses [76]. Mazelsky
and Drischler [126] used oscillatory data to obtain the indicial lift
and pitch moment responses for Mach numbers of 0.5 and 0.6. A
similar approach was used later by Dowell [127] to obtain the
indicial functions for transonic Mach numbers as well. However,
the derived indicial functions using harmonic motions depend
largely on the quality of motion, e.g. amplitude and frequency.
Recently, CFD simulations have been used for the approximation of response functions. The earliest computational efforts
involved the small perturbation potential code of Ballhaus [24]
and the Euler code of Magnus [128], both for applications to
airfoils. Singh and Baeder [129] used a surface transpiration
approach to directly calculate the angle of attack indicial response
using CFD. However, this approach has not been tested for pitch
rate indicial functions and aircraft congurations. Also, the
approach is not available in many commercial ow solvers that
do not provide access to the source code. Recently, Ghoreyshi et al.
[27] described an approach based on a grid motion technique for
CFD-type calculation of linear and nonlinear indicial functions.
This approach will be briey described next.
Assuming that the ow solver allows all translational and
rotational degrees of freedom, the location of a reference point
on the aircraft is specied at each time step. In addition the
rotation of the aircraft about this reference point is dened using
the rotation angles of yaw, pitch, and roll. The aircraft reference
point velocity, va, in an inertial frame is calculated to achieve the
required angles of attack and sideslip and the forward speed. The
velocity is then used to calculate the aircraft location at each time
step. The initial aircraft velocity, v0, is specied in terms of Mach
number, angle of attack, and side-slip angle. The instantaneous
aircraft location for the motion is then dened from the relative
velocity vector, va v0 .
For CFD-type calculation of a step change in angle of attack, the
grid immediately starts to move at t0 to the right and downward
as shown in Fig. 4. The translation continues over time with a
constant velocity vector. Since there is no rotation, all the effects in
aerodynamic loads are from changes in the angle of attack. For a
unit step change in pitch rate, the grid moves and rotates
simultaneously. The grid starts to rotate with a unit pitch rate at
t 0. To hold the angle of attack zero during the rotation, the grid
moves right and upward as shown in Fig. 4.
Fig. 4. The grid motion for modeling a step change in angle of attack and pitch rate.
Ghoreyshi and Cummings [112] later used this approach to

generate longitudinal and lateral indicial functions for a generic
unmanned combat air vehicle (UCAV) and used these functions for
predicting the aerodynamic responses to aircraft 6DoF maneuvers.
However, extending the model to predict airloads of these
maneuvers could be computationally expensive because a large
number of indicial functions need to be calculated for each
combination of input parameters.
In more detail, the indicial method assumes the linearity of
output response with respect to input variables [130]. This means
the predicted indicial functions can be used for only small
amplitude oscillations at a xed Mach number similar to the one
used for predicting indicial functions [130]. For motions at compressible speeds and large amplitude, many indicial functions
need to be generated for each combination of angle of attack
and freestream Mach number. It should be noted, however, that
these models are still cheaper than full-order simulations because
the ROMs based on indicial functions eliminate the need to repeat
calculations for each frequency [24].
In order to reduce computational cost of model creation, a
special time-dependent surrogate-based modeling approach was
used by Ghoreyshi and Cummings [112] that predicts indicial
responses for a new point from available (observed) responses.
These observed responses were viewed as a set of time-correlated
spatial processes where the output is considered a timedependent function. Assume an input vector of xt x1 t;
x2 t; ; xn t where n represents the dimensionality of the input
vector. To construct a surrogate model for tting the inputoutput
relationship, the unsteady aerodynamic responses corresponding
to a limited number of input parameters (training parameters or
samples) need to be generated. Design of Experiment methods, for
example, can be used to select m samples from the input space.
The input matrix Dm n is then dened as
2
3
x11 x12 x1n
6 x21 x22 x2n 7
6
7
D6
40
7
5
4
xm1 xm2 xmn
where rows correspond to different combinations of the design
parameters. For each row in the input matrix, a time-dependent
response was calculated at p discrete values of time, and this
information is summarized in the output matrix of Zm p as
2
3
y11 y12 y1p
6y
7
6 21 y22 y2p 7
7
41
Z6
6
7
4
5
ym1 ym2 ymp
176
Fig. 5. Stability and control aerodynamic loads model-development process [19].
where for aerodynamic loads modeling, p equals the number of

iterations used in time-marching CFD calculations. The objective of
surrogate modeling is to develop a model
that allows for

predicting the aerodynamic response of yx0 y01 ; y02 ; ; y0p at a new
combination of input parameter of x0 . To construct this surrogate
model, the responses at each time step are assumed as a separate
set, such that each column of the output matrix is a partial
realization of the total response. In this sense, p surrogate models
are created; they are denoted as Zi D for i 1; 2; ; p. A universaltype Kriging function [131] is then used to approximate these
models. For more details about creating Kriging models, the reader
is referred to Ghoreyshi et al. [132]. Having created Kriging models
for each Zi D function, the total response at x0 is then combination of predicted values of each model, i.e.

Z~ x0 Z~ 1 x0 ; Z~ 2 x0 ; ; Z~ p x0
42
where ~ shows that Kriging model is an approximation of the
actual function. For modeling aerodynamics of a maneuvering
aircraft, the input matrix of D includes combinations of angle of
attack and Mach number, i.e.
2
3
1 M 1
6 2 M 2 7
6
7
D6
43
7
5
4
m M m
so in this case n 2 and x0 0 ; M 0 . The output vector is also
dened as y C L ; C m ; C Y ; C l ; C n and therefore the output matrices
of Zj are
2
3
C j 11 C j 12 C j 1p
6C
7
6 j 21 C j 22 C j 2p 7
7
44
Zj 6
6
7
4
5
C j m1 C j m2 C j mp
where j L; m; Y; l; n represents lift, pitch moment, side force, roll
moment, and yaw moment coefcients, respectively. The unsteady
effects in drag force were assumed to be small and are not
included.
3.4. Training maneuvers
The focus of S&C research at USAFA and AFSEO has been to
effectively incorporate CFD into the model-development process
Table 1
Special training maneuvers.
Maneuver
Denition
Linear chirp
Spiral
Schroeder
t 0 A sin t 2
t 0 A t sin t
!
r
2
n
1
2kt k
cos

t 0 A
2N
T
N
k1
using dynamic CFD predictions of complete aircraft congurations.

This has lead to an innovative approach for modeling aircraft S&C
characteristics, the details of which can be found in McDaniel et al.
[19]. A schematic of the approach is graphically represented in
Fig. 5, and a summary is as follows:
1. CFD simulations are performed using computational training
maneuvers designed to excite the relevant ow physics
encountered during actual missions.
2. A mathematical ROM is built for the aircraft response.
3. The model is validated by comparing CFD simulations against
model predictions.
4. Predictions of all ight test points are made using the model to
determine the expected behavior of the aircraft.
A training maneuver (or multiple maneuvers) is needed to
provide enough information to learn the mapping between inputs
and outputs in Eq. (30) by using SBRF and RBF methods. Previous
studies to generate training maneuvers for aerodynamic characteristics [56,133136] are limited by the range of the motion
frequency content. A ROM identied from such a maneuver has
limitations with respect to S&C applications. Thus, the basic
requirement for a training maneuver to generate a reliable ROM
in S&C applications is that it sufciently covers the desired
regressor space of state variables. A ROM built on data produced
by such motions can then be used to predict the aircraft aerodynamic behavior within the regressor space. The systematic
coverage of the regressor space can be, in general, treated as an
optimization problem of lling the multidimensional space with
strong constraints resulting from the fact that some axes of the
regressor space do not represent an independent variable. For the
current study, three special training maneuvers are reviewed:
linear chirp, spiral, and Schroeder. These maneuvers are dened
in Table 1 and some examples are shown in Fig. 6.
177
Fig. 6. A spiral, a chirp, and a Schroeder training maneuver. (a) spiral maneuver; (b) chirp maneuver; and (c) Schroeder maneuver.
In Table 1, 0 is the mean angle of attack, A is the amplitude,

and is the angular velocity. The chirp maneuver used has a
constant amplitude and linearly increasing frequency in time. In
the spiral maneuver, the amplitude increases linearly in time, as
does the angle of attack. The Schroeder maneuver is a multi-stage
frequency sweep consisting of multiple cosine terms with a
specied phasing. This maneuver has three parameters that enable
direct control of the regressor space coverage: maneuver amplitude, A , maneuver length, T, and the number of frequencies in the
maneuver, N.
4. Flow solver
The ow solver used for this study is the Cobalt code [137] that
solves the unsteady, three-dimensional and compressible Navier
Stokes equations in an inertial reference frame. In Cobalt, the
NavierStokes equations are discretized on arbitrary grid topologies
using a cell-centered nite volume method. Second-order accuracy
in space is achieved using the exact Riemann solver of Gottlieb and
Groth [138], and least squares gradient calculations using QR
factorization. To accelerate the solution of the discretized system, a
point-implicit method using analytic rst-order inviscid and viscous
Jacobians is used. A Newtonian sub-iteration method is used to

improve the time accuracy of the point-implicit method. Tomaro
et al. [139] converted the code from explicit to implicit, enabling
CourantFriedrichsLewy (CFL) numbers as high as 106. Some
available turbulence models are the SpalartAllmaras (SA) model
[140], Wilcox's k model [141], and Mentor's SST model [142].
5. Validation test cases

A range of test cases, from airfoil to full aircraft, have been used
to evaluate and validate the reduced order methods. Note that the
purpose of this review is not to present validation results of CFD
codes or grid-independence study, but rather to validate ROM
predictions against CFD data. Therefore, only one CFD code and
only one mesh (ne size) was used for each test case based on the
previous experience with CFD and mesh generation tools in
predicting unsteady aerodynamics.
All RANS meshes were generated in two steps. In the rst step,
the inviscid tetrahedral mesh was generated using the ICEMCFD
code. This mesh was then used as a background mesh by TRITET
[143,144] which builds prism layers using an advancing front
technique. TRITET rebuilds the inviscid mesh while respecting the
178
Fig. 7. The NACA0012 grid and static and dynamic validations. The static conditions are M 1 0:3 and Re 5:93 106 . Static and AGARD CT2 experimental data are from
Ladson [145] and Landon [146], respectively. In (f) moment reference point is 0.25c. (a) Euler grid; (b) viscous grid; (c) static normal force coefcient; (d) static pitch moment
coefcient; (e) normal force coefcient of AGARD CT2; and (f) pitch moment coefcient of AGARD CT2.
size of the original inviscid mesh from ICEMCFD. Simulations of

aircraft congurations were run on the Cray XE6 machines at the
Engineering Research Development Center (ERDC) (Garnet with
2.7 GHz core speed and Chugach with 2.3 GHz core speed) which
have approximately 20,000 and 11,000 cores, respectively.
5.1. NACA 0012 Airfoil

Both Euler and RANS meshes are available as shown in Fig. 7. In
both meshes, the minimal distance from the body to each of the
outer boundaries is 20c, where c is the airfoil chord. The Euler
mesh was generated using Gridgen version 15.0, and is a structured mesh generated by normal extrusion of surface connectors.
The overview of the Euler mesh is shown in Fig. 7(a). The RANS
mesh volume is rectangular with the airfoil geometry centrally
located. The no-slip adiabatic wall boundary conditions are
employed at the body surface, and a modied Riemann-invariant
condition is implemented at the far-eld boundary. The RANS
mesh has prisms cells near the airfoil and tetrahedra cells elsewhere. The pitch axis is set to 0.25c, but the moment reference
point is at the leading edge unless stated otherwise. The overview
of the RANS mesh is shown in Fig. 7(b).
The static Euler and RANS calculations are shown in Fig. 7. Note
that Euler calculations are signicantly faster than RANS calculations (on the order of ve times faster for the meshes used in this
work). The simulations correspond to M 0.3 and Re 8:93 106
in order to match experimental data from Ladson [145]. All RANS
simulations were performed using the SA turbulence model. Euler
and RANS predictions compare well with the experiments at low
angles of attack as shown in Fig. 7. The RANS model accurately
predicts the maximum lift, but the stall region predictions do not
match as well. The Euler model predicts the slopes of lift and pitch
moment fairly well up to moderate angles of attack. Differences
between the Euler predictions and experiments are observed at
high angles of attack due to the inviscid assumption, although the
Euler simulations show ow separation on the upper surface at
higher angles of attack resulting in a fall in lift slope. Cobalt in
Euler mode predicts the secondary boundary layer separation at
high angles of attack due to adverse pressure gradients, however,
these predictions are not accurate.
A dynamic test case (AGARD CT2) was also selected with
available experimental data from Landon [146]. The AGARD CT2
is a pitch forced oscillation with test conditions summarized in
Table 2. The CFD simulation of the RANS grid ran for three cycles
with a non-dimensional time-step of t n Vt=c 0:01. The results
are shown and compared with experimental data in Fig. 7. Even
for this unsteady case, very good agreement was found, although
the angles of attack all fall within the linear range. These predictions gave condence in the ability of the current numerical
approach to predict unsteady aerodynamics.
5.2. SDM
The Standard Dynamics Model (SDM) is a generic ghter conguration based on the F-16 planform. The model includes a slender
strake-delta wing, horizontal and vertical stabilizers, ventral ns, and
a blocked off inlet. The three-view drawing is shown in Fig. 8. This
geometry has been tested extensively at various wind tunnel facilities
to collect aerodynamic data [147149]. Note that slightly different
geometries were used in these previous studies.
A full-span geometry mesh is available as shown in Fig. 9. The
mesh has around 9 million points and 19.5 million cells. The wind
tunnel experiments [150] were used to validate the CFD predictions at low speeds. The conditions of the tests were V 110 m/s,
Re 0.57 million and 51 for 0901. All CFD simulations were
run at free-stream conditions consistent with ow conditions in
the wind tunnel tests. For ow solution, the RANS equations are
discretized by second-order spatial and temporal operators. The
turbulence models used are SA, SARC, and Menter's SST.
The static force and moment coefcients were compared with
experimental data in Fig. 10. The comparison showed that there
is a good agreement between the RANS predictions and the
measurements for angles of attack below 251. However, all
Table 2
Description of the AGARD CT2 test conditions.
Test conditions
Values
Mach number, M
Mean incidence, 0
Pitch amplitude, A
Reduced frequency, k
Reynolds number, Re
0.6
3.161
4.591
0.0811
4.8 106
179
Fig. 9. The SDM aircraft surface mesh model.
Fig. 8. Standard dynamic model (SDM) layout [150].
180
Fig. 10. Static aerodynamic predictions at V0 100 m/s and 5. (a) normal force coefcient; (b) pitch moment coefcient; (c) side force coefcient; (d) roll moment
coefcient; and (e) yaw moment coefcient.
turbulence models predicted a positive pitch moment slope at

zero degrees angle of attack, while experiments show a falling
trend at this angle. This was likely due to different inlet geometries
used in the wind tunnel models. At 51, a pair of vortices emanating
from the strake and wing leading-edge have formed as shown in
Fig. 10(a). These vortices do not exhibit breakdown and do not

interact over the airframe. The vortex formation causes an additional increase of lift and pitch moment coefcients. In experimental tests, the sign of the pitch moment is reversed at this
angle, while CFD shows a jump in the moment rate of increase. No
181
Fig. 11. SDM ow-eld visualization. The calculations are for a Mach number of 0.3 and 5 using SARC turbulence model. (a) 51; (b) 101; (c) 141; (d) 181;
(e) 221; and (f) 251.
signicant changes on lateral force and moment coefcients were

observed at this angle.
The SDM vortices grow in size and strength with increasing
angle of attack. At 101, the center of the wing vortex is shifted
laterally, while the shedding point is moved forward as shown in

Fig. 11(b). There are still no signs of vortex breakdown, and the
lateral moments slightly change due to the wing vortex movement. Around 141, the two vortices wind around each other
182
towards the trailing edge of the wing as shown in Fig. 11(c). With a
slight increase in angle of attack, the wing vortex appears to
breakdown quickly as shown in Fig. 11(d). The vortex breakdown
leads to a smaller lift rate of increase and a negative pitch moment
slope. The vortex breakdown phenomenon is asymmetric and
hence the lateral force and moment coefcients suddenly start to
change very rapidly. At 221 the strake vortex is also burst as shown
in Fig. 11(e). Finally, at 251 there is no sign of a wing vortex as
shown in Fig. 11(f).
5.3. SACCON UCAV
The SACCON geometry and experimental data were provided to
the partners participating in NATO RTO Task Group AVT-161
(Assessment of Stability and Control prediction Methods for NATO
Fig. 12. The SACCON geometry [151].
Air and Sea Vehicles) [151]. The objective of this task group was to
evaluate CFD codes against wind tunnel results. The vehicle
planform and section proles were dened in cooperation
between the German Aerospace Center (DLR) and EADS-MAS.
DLR adjusted the pre-design geometry for wind tunnel design
purposes which actually led to a higher overall thickness at the
root chord to provide enough space for the internal strain gauge
balance. The aircraft has a lambda wing planform with a leading
edge sweep angle of 531 as shown in Fig. 12. The root chord is
approximately 1 m, the wing span is 1.53 m, the reference chord is
0.48 m, and the reference area is 0.77 m2. The main sections of the
model are the fuselage, the wing section, and wing tip. The
conguration is dened by three different proles at the root
section of the fuselage, two sections with the same prole at the
inner wing, forming the transition from the fuselage to wing and
the outer wing section. Finally, the outer wing section prole is
twisted by 51 around the leading edge to reduce the aerodynamic
loads and shift the onset of ow separation to higher angles of
attack.
The wind tunnel model was designed and manufactured at
NASA Langley Research Center (LaRC). The model was designed to
accommodate a belly sting mount for tests in the German-Dutch
Low Speed Wing Tunnel (DNW-NWB) in Braunschweig and the
14 ft 22 ft low speed wind tunnel at NASA LaRC [151,152]. The
high-angle-of-attack ow around SACCON is very complicated and
unsteady due to vortical ow formation, vortex interaction, and
vortex breakdown. SACCON also has complicated aerodynamic
characteristics at non-symmetric ow conditions. Some experimental aerodynamic behavior of the SACCON conguration in the
lateral direction are shown in Fig. 13. The experimental results
[151] show that side force, yaw and roll moments are a nonlinear
function of angle of attack, most signicantly above an angle of
attack of 101. Fig. 13 also shows that the angle of attack dependency can be seen in the roll moment even at low angles of attack.
The lateral coefcients are nearly linear with side slip angle for
angles of attack below 151, and become increasingly dependent on
side-slip for angles of attack above 151 as shown in Fig. 13.
Two meshes are available, the rst uses a belly mounted sting
present in the experiments and the second has no sting. The grid
including sting geometry is shown in Fig. 14. This grid contains
Fig. 13. SACCON side force, roll moment, yaw moment versus angle of attack for different angles of side slip [151].
183
Fig. 14. The SACCON grid and static validations. The static conditions are M 1 0:144 and Re 1:61 106 . Experiments are obtained from Cummings and Schtte [151] and
shown with lled circles. The solid lines with delta markers show the Cobalt predictions for a grid without sting geometry. The lines with square marker show predictions of
a grid including sting geometry. Turbulence model is SARC. (a) SACCON grid; (b) static lift coefcient; (c) static drag coefcient; and (d) static pitching moment coefcient.
around 4.5M points and 13M cells. The total run time of 1000
iterations for the model without sting geometry was 5 h using 128
processors. The SA with Rotation/curvature Correction (SARC)
[153] turbulence model was used for all CFD simulations. The
predicted static coefcients of lift, drag, and pitch moment were
compared with experiments in Fig. 14. The gures show that CFD
predictions closely follow the trends of experimental data up to
moderate angles of attack. The offsets in low angle of attack pitch
moment in the model have been shown to be due to the effects of
the belly sting mounting present in the experiments. For the ROM
studies, the mesh without sting geometry was used since it has
less grid points.
Some of the SACCON aerodynamic features are shown in Fig. 15.
Two emanating vortices from the wing tip and apex are present at
141 angle of attack (Fig. 15(a)). These vortices lead to a negative
pressure region on the upper wing surface, and hence augments the
lift force. As the angle of attack increases above 161, the onset point
of the outboard vortex starts to travel toward the wing apex due to
increasing adverse pressure gradients. At 19.51 angle of attack the

vortices are already interacting as shown in Fig. 15(b). Further
increasing of angle of attack causes the inboard vortex to start to
break down (Fig. 15(c)). At higher angles of attack the tip vortex also
breaks down. The interaction of the vortices produces a strong
recirculation zone over the upper wing (Fig. 15(d)) and results in
wing stall and the aerodynamic center backward movement.
For generation of 6DoF SACCON maneuvers, the wind tunnel
model was scaled up to t the characteristics of a real aircraft.
Initially, estimations of the mass and moments of inertia were
made, through work carried out in the NATO group, based on the
Northrop YB-46 aircraft. Table 3 summarizes the SACCON geometry parameters and mass and moments of inertia.
5.4. X-31
The X-31 aircraft is considered in this paper as well. The aircraft
geometry and wind tunnel data were again provided to the
184
Fig. 15. The SACCON vortical ows using SA turbulence model. The conditions are M 1 0:144 and Re 1:61 106 . The vortices core lines are extracted and shown by black
lines. For case (d), the ow separations lines are shown by red lines. (a) 141; (b) 19:51; (c) 20:51; and (d) 231. (For interpretation of the references to color in
this gure caption, the reader is referred to the web version of this paper.)
Table 3
Geometry parameters and mass/inertias of SACCON
yable model.
Mean aerodynamic chord, c (m)
Wing area, S (m2)
Wing span, b (m)
Ixx (kg m2)
Iyy (kg m2)
Izz (kg m2)
Maximum take-off weight, MTWO
(kg)
5.011
55.0
13.0
8014
6564
8937
2000
participants in NATO RTO Task Group AVT-161. The vehicle is a

super-maneuverable ghter which was built by the United States
and West Germany in the 1990s. The test aircraft has been a
subject of extensive ight and wind tunnel tests (see for example
[154157]), as well as CFD simulations (an example is the work of
Schtte et al. [11]). A three-view drawing of the vehicle is shown
in Fig. 16. The aircraft has a fuselage length of 13.21 m, a canard,
and a double delta wing with total wing span of 7.26 m. The inner
delta wing has a sweep angle of 571 and the outer sweep is 451.
The inner wing sweep places the wing behind the supersonic
shock wave, while the outer one improves the vehicle stability and
control [158]. The canard is a cropped delta wing with a sweep
angle of 451. Additional characteristics of the model are the inner
Fig. 16. The X-31 aircraft geometry.
and outer leading edge aps, the trailing edge aps, the front wing,
and rear fuselage strakes.
The mesh overview is shown in Fig. 17. The grid is a symmetric
conguration and contains 4.9 million points and 11.7 million cells.
Three boundary conditions were imposed to the surfaces: a fareld, symmetry, and solid wall. The low-speed experiments were
made available by the DLR, German Aerospace Center [157]. The
wind tunnel model has a closed inlet and is tted with moving lift
and control surfaces. The experiments are composed of two
185
Fig. 17. The X-31 aircraft mesh model. (a) Half-Model Mesh and (b) Surface mesh around LEX and canard.
Fig. 18. The X-31 static loads validations. The static conditions are: M 0.18 and Re 2 106 . (a) lift coefcient; (b) drag coefcient; (c) pitching moment coefcient.
setups. The rst setup uses a belly mounted sting attached to the
model directly under the main wing. This setup allows six degree
of freedom motions. The second setup uses an aft mounted sting
connected to an arm in the wind tunnel. The values of lift, drag,
and pitch moment from the second setup were used to validate
CFD predictions. Four turbulence models were tested: SA, SARC,

SST, and SARC with Detached Eddy Simulation (SARC-DES). Fig. 18
compares the lift, drag, and pitch moment coefcients obtained
from each turbulence model with the available measurements.
All these models yielded similar predictions at low angles of
186
Fig. 19. The X-31 vortical ows using SARC-DES turbulence model. The conditions are M 0.18 and Re 2 106 . (a) 101; (b) 141; (c) 181; (d) 201; (e) 221;
and (f) 251.
attack, but they resulted in a wide spread of predictions at

moderate to high angles. For angles between 15231 SARCDES and SST models performed quite well compared to SA and
SARC. DES and SST accurately predicted unsteady separated ows
occurring at these angles, but for angles above 231 all models
failed to predict accurately the massively separated ows.
Some features of aerodynamic characteristics from the SARCDES turbulence model predictions are explained. There is an
emanating vortex from the canard tip at small angles of attack.
This vortex is the source of the small non-linearity in the pitch
moment at low angles of attack. As the angle of attack is increased,

the canard vortex becomes stronger, resulting in a negative
pressure on the upper surface and forward movement of the
aerodynamic center. Therefore, the pitch moment slope suddenly
increases from the slope value at zero degrees angle of attack. This
vortex is shown in Fig. 19(a) for 101 angle of attack. Around an
angle of attack of 141, the canard vortex starts to breakdown and
the wing vortex is formed as shown in Fig. 19(b). The wing vortex
helps to further forward movement of aerodynamic center and
increase of pitch moment. At 181 angle of attack, the canard vortex
breakdown point is nearly moved to the leading edge and then the
wing vortex starts to breakdown as shown in Fig. 19(c). This results
in an aft movement of the aerodynamic center and a change in the
pitch moment slope sign. The wing vortex breakdown point moves
187
towards the leading edge by increasing angle of attack (Fig. 19(d),

(e)). The canard vortex is fully separated at these angles. As the
vortex breakdown point becomes close to the wing leading edge
(Fig. 19(f)), the pitch moment starts to rise again.
Fig. 20. The NACA 0012 lift and pitch moment response functions at M 0.3 and Re 5:93 106 . (a) linear C L , q 0; (b) linear C Lq , 0; (c) C L at different , q 0; (d) C Lq at
different ; (e) C m at different , q 0; (f) C mq at different .
188
6. Analysis
The models are evaluated by comparing the predictions with
time-accurate CFD solutions which are labeled Time-Marching in
the plots. The prediction capability of ROMs is reviewed rst for
the airfoil, SDM, and SACCON test cases. These models are then
tested for a pitching motion of the X-31 conguration to permit
direct comparison of costs and errors. In this work, the error norm
was dened as
E
cfd
N
i yrom ij
i 0 jy
cfd
ycfd
max ymin
100
45
where N is the total number of time-steps used in the CFD

solutions, and the superscripts cfd and rom indicate, respectively,
the time-marching and the model predictions .
6.1. Airfoil aerodynamics modeling

All airfoil calculations were run on 10 cores with 2.3 GHz core
speed. The ROMs reviewed for the airfoil predict only lift and pitch
moment coefcients and were used at a xed Mach number
similar to the one used in the training maneuver/response
function calculations.
6.1.1. Models for M 0.3

The RANS grid with the SA turbulence model was used for the
response calculations of the NACA 0012 airfoil. The step responses
of the airfoil are shown in Fig. 20 for M 0.3. These simulations
were run for 0.125 s with a time step of 5 10 5 s. Fig. 20(a) and
(b) shows that the lift has a peak at s 0 followed by a rapid falling
trend. The lift again builds up and asymptotically reaches the
steady-state value at the nal time. The initial peak can be
explained based on the energy of acoustic wave systems created
by the initial perturbation [89,106]. Fig. 20 shows that the lift
values at the initial and nal times of the response are favorably
comparable with the analytical results of Lomax, Wagner, and
Leishman. Fig. 20 also shows the effects of angle of attack on
response functions. The results show that initial values of the
response functions are invariant with the angle of attack but the
transient trend and steady state values change depending on the
angle of attack.
A linear ROM was created using Eq. (23) and indicial responses
at zero degrees angle of attack. This model was then used for the
response prediction of two pitching motions with reduced frequencies of k 0.077 and 0.307, a mean incidence of zero degrees,
and an amplitude of one degree. The results of the linear model
are compared with the time-marching solutions in Fig. 21. The
comparisons show that the linear ROM provides very good
agreement for these slow and fast motions. The error norm values
Fig. 21. The linear indicial response ROM model for the airfoil, M 0.3 and Re 5:93 106 . (a) pitch oscillation with k 0.077; (b) pitch oscillation with k 0.077; (c) pitch
oscillation with k 0.307; and (d) pitch oscillation with k 0.307.
for the lift and pitch moment coefcients are around 0.9% and
2.2%, respectively.
A nonlinear ROM was also created using Eq. (24) and nonlinear
response functions calculated at M0.3. Fig. 22(a) and (b) compares
the linear and nonlinear ROM predictions for a ramp increase of
189
angle of attack from 1.51 to 151 with a rate of 10 deg/s. The grid is
undergoing only a translational motion, so the second terms in Eqs.
(23) and (24) are zero. Notice that the forces and moments acting on
the airfoil during translation are different from the static values. The
ow change is not as fast as the angle of attack change and hence
Fig. 22. The nonlinear indicial response ROM model for NACA 0012, M 0.3 and Re 5:93 106 . The moment reference point in (b) is on the quarter chord point. (a) ramp
motion with rate of 10 deg/s; (b) ramp motion with rate of 10 deg/s; (c) pitch oscillation with k 0.077; (d) pitch oscillation with k 0.077; (e) pitch oscillation with k 0.307;
and (f) pitch oscillation with k 0.307.
190
the ramp motion underestimates the static coefcients. The results

show that for low angles of attack both linear and nonlinear ROMs
are in good agreement with values from the time-marching simulation. For large angles of attack, the nonlinear model closely follows
the time-marching solution, but the linear model is off since the
model formulation is valid in linear regimes. The error norm values
of the linear model are 2.82% and 5.0% for the lift and pitch moment
coefcients. These errors drop to 1.4% and 2.2% using the
nonlinear model.
Fig. 22(c)(f) also shows two pitch oscillation cases with
reduced frequencies of k 0.077 and 0.307. Both motions start
from 101 and have an amplitude of 41. Fig. 22(c)(f) shows that
the nonlinear ROM predictions agree well with time-marching
simulation data. The error norm values are around 2.3% and 2.7%
for the lift and pitch moment coefcients, respectively.
Some of the model strengths and weaknesses are now briey
highlighted. An advantage of the indicial response model is that
responses are calculated as a function of time and include the
_ therefore they can be used for any frequency of
effects of _ and q,
interest. These models can also predict the initial transient
behavior seen in CFD solutions. However, response functions do
not work for the full range of Mach numbers. For aerodynamic
modeling of large amplitude motions, two types of problems were
also reported [159]: (1) a large number of simulations is required
to model a highly nonlinear motion, therefore, the reduced order

model becomes quiet expensive to generate, and (2) the second
problem is related to the stability of response type CFD calculations at large angles of attack. Fig. 23 shows the effects of angle of
attack on the responses of the NACA 0012 airfoil at M 0.3. Fig. 23
shows that as the angle of attack reaches the stall angle, small
oscillations are found in the response solutions. For angles of
attack beyond the stall angle, the oscillations become large and
persist for a long time. This behavior makes the nonlinear indicial
functions infeasible to model motions at large incidence angles.
In terms of computational cost, the linear and nonlinear models
required about 0.6 and 6.0 h of CPU time, respectively. The nonlinear model requires response calculations at different angles of
attack and hence comes with more computational cost. However,
once these models are constructed, they can be used to predict
airfoil responses to a wide range of pitching and plunging motions
in order of a few seconds. In comparison to model costs, each
time-accurate simulation (full-order model) costs around 1.8 h
of CPU.
RBF neural networks were also used for aerodynamic modeling
of the NACA 0012 airfoil [36]. The networks used had a hidden
radial basis layer with a Gaussian transfer function. All training
data were generated using CFD with either the Euler or RANS
equations. Six training datasets were dened to nd a mapping
Fig. 23. The NACA 0012 airfoil angle of attack response functions. In these
simulations M 0.3 and Re 5:93 106 .
Fig. 25. RBF neural network training performance. The y-axis displays root mean
squared error and the x-axis shows number of iterations.
Fig. 24. The plunge and pitch training maneuvers for airfoil Euler simulations. (a) plunge training motions and (b) pitch training motions.
191
Fig. 26. RBF network predictions of testing motions; simulations correspond to an Euler mesh and ran at M 0.3. (a) plunge oscillation with f 1 Hz; (b) ramp motion with
rate of 10 deg/s; (c) plunge oscillation with f 0.5 Hz; and (d) pitch oscillation with f 2.5 Hz.
Fig. 27. Training data for the airfoil RANS simulations; M 0.3. (a) pitching motion and (b) ramp motion.
Fig. 28. RBFNN predictions of the airfoil pitching motion using both Euler and RANS equations; M 0.3. (a) normal force coefcient and (b) pitch moment coefcient.
192
Fig. 29. NACA 0012: training maneuver for the generation of a discrete-time multi-input Volterra model (M 0:764, 0 0:01 and k 0:10); in (a), variation of angle of
attack with time; in (b), Model refers to the discrete-time multi-input Volterra model. (a) angle of attack and (b) pitch moment coefcient.
Fig. 30. NACA 0012: predictions of pitching moment dynamic dependence

(M 0.764, 0 0:01, A 8:51, and k 0.10); Model refers to the discrete-time
multi-input Volterra model.
Fig. 31. NACA 0012: carpet plot of the pitching moment coefcient (M 0.764,
0 0:01, and k 0.10); large spheres indicate the 4 CFD solutions used to construct
the SBRF model.
Fig. 32. NACA 0012: predictions of pitching moment dynamic dependence

(M 0.764, 0 0:01, A 8:51, and k 0.10); Model refers to the surrogate-based
recurrence-framework.
Fig. 33. Design samples.
193
Fig. 34. Nonlinear lift and pitch moment indicial solutions due to angle of attack for M 0.3 and 0.6. (a) nonlinear lift responses at M 0.3; (b) nonlinear lift responses at
M 0.6; (c) nonlinear pitch moment responses at M 0.3; (d) nonlinear pitch moment responses at M 0.6.
Fig. 35. Lift and pitch moment indicial solutions due to pitch rate for 01. The pitch axis and moment reference point are located at 35% MAC. (a) Pitch rate lift indicial
functions and (b) Pitch rate pitch moment indicial functions.
194
between aerodynamic coefcients with motion variables at a xed

Mach number of 0.3. These motions are linear chirp type and have
an initial frequency of 1 Hz but started at different angles of attack
as shown in Fig. 24. The starting angles of attack are 51, 01, and
51 to include the effects of steady-state conditions into the
pitching and plunging motions.
All six training motions were rst simulated using the Euler
equations. The CFD calculations started from a steady solution and
ran for three seconds of physical time with a time step size of
5 10 5 s. This gives a total computational cost equivalent to that
of 240 Euler steady-state calculations or 17 CPU hours; this is
much higher than the cost of response functions because training
motions take a longer time to complete than indicial responses.
Euler calculations are ve times faster than calculations using the
RANS grid. Therefore, the cost of simulating all six training
motions using the Euler approach is equivalent to that of only 48
RANS steady-state calculations.
The reported normal force and pitch moment coefcients of
training maneuvers were re-arranged according to Eq. (30) and
then fed to the RBF neural network for training with different
values of n and m and with a performance error threshold of
1 10 6 . All networks computed and converged to the threshold
error as shown in Fig. 25. The results showed that using m 2 and
n 4 in Eq. (30) is sufcient for modeling the studied motions. The
costs to train the networks and that of executing the networks for
new maneuvers were around 88 s and 17 s, respectively.
Once the networks were trained, they were used to predict the
aerodynamic loads of new maneuvers again simulated by the
Euler equations. The predicted normal force and pitch moment of
these maneuvers are compared with CFD data in Fig. 26. Fig. 26
(a) shows the airfoil responses to a plunge oscillation with a
frequency of 1 Hz started at zero degrees angle of attack. The
results show that the ROM predictions match well with the timemarching calculations. The error norm values for normal force and
pitch moment are 0.75% and 5%, respectively. The cost of simulating this motion in a time-marching fashion is around 0.36
CPU hours.
Fig. 26(b) compares the ROM predictions for a ramp increase of
angle of attack with a rate of 10 deg/s started at zero degrees angle
of attack. The results show that ROM predictions agree well with
the CFD prediction (the error norm values are 0.34% and 5.45% for
the normal force and pitch moment coefcients, respectively).
Also, two pitch oscillation motions with frequencies of f 0.5 Hz
and f 2.5 Hz were considered. The low-frequency motion predictions are compared with the CFD data in Fig. 26(c). Again a good
match is found (the normal force and pitch moment error norms
are 0.46% and 5.98%). Likewise, Fig. 26(d) shows that the ROM
predictions closely match with the time-marching simulations of
the motion with f 2.5 Hz.
The development of a ROM from RANS/Euler was also investigated by Ghoreyshi et al. [36]. For network training purposes,
these data are also called primary/secondary or expensive/cheap
data. The cheap data are assumed to provide information at least
about the trend of the target function, whereas the expensive
calculations give quantitative information. To illustrate the
approach the NACA 0012 airfoil was used again. The timemarching Euler predictions of all six training sets described in
Fig. 24 are considered as cheap data. It is assumed that these
predictions had similar trends to the RANS predictions, but with
different values. The new RBF model tries to correct these trends
by using much less expensive data.
The expensive (or RANS) calculations were run only for two
training sets. The rst set is a chirp-pitch that runs for one second
of physical time as shown in Fig. 27(a). The second set is a ramp
motion with a rate of 10 deg/s, starting at 101 and progressing
for two seconds of physical time as shown in Fig. 27(b). Both
motions have a sweep of angle of attack from the low to high

angles, and therefore are found very helpful to learn the differences between Euler and RANS predictions. Both training sets
were simulated using time-marching RANS equations and the SA
turbulence model. Likewise for the Euler predictions a time step of
5 10 5 s was used. This results in a computational cost equivalent to that of 40 RANS steady-state calculations and a total cost
(both Euler and RANS) of 88 RANS steady-state calculations. Note
that using RANS for time-marching simulations of all six training
sets is equivalent to that of 240 RANS steady-state calculations. In
this sense, a ROM generated using both Euler and RANS calculations is 63% cheaper than a ROM generated using the RANS
calculations of six training maneuvers.
The RBFNN was trained for these maneuvers using both RANS
and Euler evaluations and then was tested for new maneuvers
simulated using RANS equations. A pitch oscillation was used to
demonstrate the network performance. The motion is a constant
frequency type with f 0.5 Hz started at zero degrees angle of
attack. The unsteady normal force and pitch moment coefcients
were obtained using simulations of unsteady RANS equations and
are shown in Fig. 28(a) and (b). Fig. 28(a) and (b) shows that the
output of a network using both Euler and RANS training data
match well with time-marching values. Such a network has similar
trends as predicted by the network trained from Euler data and
then corrects them. The network using only Euler data match
time-marching solutions at low angles of attack, but overestimates
the values at high angles due to the inviscid assumption. Fig. 28
(a) and (b) shows that using only RANS data of two training sets
does not help much to approximate the time-marching values,
since these maneuvers are short and do not properly cover the
frequency and starting angle of attack space.
In comparison to response function models, a training maneuver that even partially covers the input space will take much
longer to simulate in CFD than response functions. RBF based
model predictions are also inuenced by the training maneuver
type and the range of frequency and angle of attack used in the
training maneuver as well as the network settings. On the other
hand, the RBF based models can be applied for aerodynamic loads
prediction in the post-stall regime, assuming an appropriate
maneuver was used in the network training.
Fig. 36. Validation of Cmq values calculated from pitch-rate indicial functions.
Experimental data are from Da Ronch et al. [161].
6.1.2. Models for M0.764

Models based on Volterra theory and SBRF were also used for
aerodynamic modeling of the NACA 0012 airfoil. One problem to
present is the pitch moment response of the airfoil to a pitching
motion at M0.764 for 8:51 sin t with k 0.1.
195
For the identication of the Volterra kernels, Da Ronch et al. [32]

used a CFD-generated training maneuver as the data source. The
variation of the maneuver angle of attack with time is shown in
Fig. 29(a). The Mach number of this maneuver was set to 0.764
similar to the target motion and the reduced frequency was 0.1.
Fig. 37. ROM prediction of plunging motions.
196
The mean angle of attack was zero, but the amplitude changes in
time; the maximum angle of attack in the maneuver is 141 to excite
nonlinear aerodynamics due to shock-induced separation. The pitch
moment coefcient of this maneuver computed using the RANS grid
is illustrated in Fig. 29(b). The Volterra model is also included in the

same gure for comparison with the training signal used in the
identication process. The overall agreement is reasonable. However,
deviations can be observed at both ends of the time interval.
Fig. 38. ROM prediction of pitching motions.
As the airfoil is symmetric and oscillates around zero degrees

mean angle of attack, the resulting aerodynamic loads are odd
functions of the angle of attack change. As a result, any odd
kernel was neglected in the identication process. The Volterra
model used in the results includes kernels up to third order

as follows:
H 1
H 1_
_
H ;
2
_ ;_
H ;
3
Fig. 39. Design space samples.
197
H 1
_
H 2;
_ ;
_
H 3;
H ;;
3
46
The upper bound of the integer k in Eq. (37), which indicates

the number of previous time-steps to account for, was set to one
for the terms H 1 and H 1_ ; for the other kernels, it was set to zero.
The effect of including more terms in the model was assessed
during the identication of the Volterra model, and in the
subsequent comparison for the target maneuver.
Fig. 30 conveys the unsteady pitch moment coefcient for the
target maneuver. The CFD solution exhibits nonlinear characteristics
at the higher angles of attack. The Volterra model including the
kernels shown in Eq. (46) is illustrated in the same gure for
comparison. Including rst-order kernels, the error norm in the
prediction of the target maneuver was 6.78%. Second-order kernels
were then added and the Volterra model deviated by 6.87% from the
CFD solution. By introducing third-order kernels (this model is
shown in Fig. 30), the error norm slightly drops to 6.18%. Despite
using kernels up to third order, the agreement is not excellent.
The overall t of the model is reasonable, but there is no sign of the
prediction of any non-linear features.
Fig. 40. The lift and pitch moment indicial functions. (a) linear C L functions; (b) linear C m functions; (c) nonlinear C L functions at M 0.3; and (d) nonlinear C m functions
at M 0.3.
198
The mediocre performance of the Volterra model may be

attributable to two aspects. To obtain a robust predictive model,
in combination with kernels of the second and third order, the
inuence of the previous time-history on the current time instant
was limited. Second, the training case considered may not be
optimal for the model identication.
An SBRF model was also constructed to predict the pitch
moment response of similar target maneuvers. This model was
created from CFD solutions corresponding to 30 different combinations of A , constrained to vary between 01 and 101. With an
increment in A of 0.251, the parameter space was initially covered
for 75% of all the possible combinations. Fig. 31 depicts the carpet
plot of the pitch moment coefcient for each value of the parameter
A obtained using the SBRF model. The axis Iteration indicates the
time evolution through the last cycle of the simulations.
Next, the number of initial training CFD solutions was iteratively reduced to focus on the predictive capabilities of the model.
Four CFD solutions were eventually retained, and these are

included as large spheres in Fig. 31. Note that two training cases
are located at the borders of the parameter space to avoid
extrapolation. The remaining two cases were automatically
sampled by the algorithm described in reference [117]. Note also
that there is no close proximity of the four training cases to the
amplitude of 8.51 used in the target maneuver.
The ROM predictions are evaluated for the target maneuver with
a Mach number of 0.764. The comparisons are presented in Fig. 32
which shows the reduced-order model closely approximates the
reference solution (the error norm is 1.7%), featuring very similar
nonlinear characteristics at higher angles as well. Note that this
model can only predict responses to motion at a constant reduced
frequency of k 0.1. The number of training maneuvers, and therefore the computational cost of this model, will signicantly increase
for aerodynamic modeling throughout the angle of attack/Mach
number/frequency space.
Fig. 41. The side-force, roll and yaw moments indicial functions. (a) C Y functions at M 0.1; (b) C Y functions at M 0.3; (c) C Y functions at M 0.5; (d) C n functions at
M 0.1; (e) C n functions at M 0.3; (f) C n functions at M 0.5; (g) C l functions at M 0.1; (h) C l functions at M 0.3; and (i) C l functions at M 0.5.
6.2. SDM aerodynamics modeling

A ROM based on an indicial response function, along with a timedependent surrogate approach, is reviewed for the SDM aerodynamic
modeling in the angle of attack/Mach number/frequency space.
199
The indicial response functions of the SDM aircraft are interpolated from some available samples in the angle-of-attack and
free-stream Mach number space. Note that these functions only
need to have dependency on angle-of-attack and Mach number,
and once they have been calculated they could be used to predict
Fig. 42. The lift and pitch moment indicial functions with a unit step change of normalized pitch rate at different Mach numbers.
Fig. 43. The lateral coefcients with a unit step change in the normalized roll rate.
200
the aerodynamic response to any frequency of interest. For

example the samples could be generated using the methods of
factorial designs, Latin hypercube sampling, low discrepancy
sequences, or designs based on statistical optimality criteria
(A-, D- and G-optimal designs) [160]. Factorial designs are extremely simple to construct and have been used in this work. The
SDM motions considered encompass and M values in the range
of [ 101, 101] and [0.3, 0.7], respectively. Assuming symmetrical
ow solutions with respect to the angle of attack, the indicial
functions are only calculated for positive angles of attack.
A set of samples including 50 points is dened on the and M
space using factorial design; these points are shown in Fig. 33. The
indicial functions are calculated using CFD with the grid motion
approach for each sample condition. All these calculations started
from a steady-state solution such that the Mach number in the
steady-state simulations corresponds to each sample Mach number. The step function calculations are second order in time with a
non-dimensional time step of t n t:V=c 0:01. The calculated
indicial functions due to angle of attack are shown in Fig. 34 for
M 0.3 and M 0.6. Fig. 34(a) and (b) shows that the indicial lift
has a peak at s 0 followed by a rapidly falling trend. The lift again
builds up and asymptotically reaches the steady-state value. The
pitch moment also has a negative peak at s 0 as shown in Fig. 34
(c) and (d). Fig. 34 shows that the initial values of indicial functions
are invariant with angle of attack, but the intermediate trend and
steady-state values change, depending on the angle of attack.
Although, the nal values of indicial lift are nearly unchanged

for angles of attack below 51, but the pitch moment nal values
are different even at small angles of attack due to vortices on
the wing.
The effects of Mach number on the lift and pitch moment
indicial functions with respect to pitch rate are shown in Fig. 35.
Fig. 35 shows that increasing Mach number results in the increase
of steady-state CLq and the decrease of steady-state Cmq (the
so-called pitch damping derivative). Fig. 36 compares the steadystate Cmq values calculated from the CFD code with the out-ofphase components of pitch moment derivative, i.e. C mq C m_
measured at different Mach numbers and zero degrees angle
of attack; these experimental data are detailed by Da Ronch
et al. [161]. Like the static predictions, the CFD values slightly
underestimate the experimental pitch moment data, although
the CFD predictions do not include the effects of the rate of
change of angle of attack, i.e. C m_ . Cmq is the largest factor in
the sum C mq C m_ , typically accounting for 90% of the sum.
Again, the underestimation of experimental data is likely due to
different inlet geometries in the wind-tunnel and the SDM
geometry used in this work. Note that the indicial function
approach allows the direct calculation of pitch damping derivatives but extraction of dynamic derivatives from harmonic motions
results in the in-phase and out-of-phase components [161] and
additional work required to separate each derivative from these
components.
Fig. 44. The lateral coefcients with a unit step change in the normalized yaw rate.
A nonlinear ROM based on indicial response theory is created

along with a time-dependent surrogate model to determine the
terms in Eq. (24) at each time step. The validity of the ROM is
tested for several arbitrary pitching and plunging motions. These
motions start from different steady-state conditions (not being
used during ROM creation) and run for different amplitudes and
frequencies. The ROM predictions are compared with timeaccurate CFD simulations in Figs. 37 and 38. Figs. 37 and 38 show
that the ROM lift and pitch moment predictions agree well with
the time-marching simulation values. Small discrepancies are
found in the pitch moment predictions at negative angles of
attack. This is likely due to the fact that SDM pitch moment is
not symmetric with angle of attack and hence the response
201
functions generated at positive angles cannot predict the slope

changes correctly. Note that the cost of ROM creation and the
average cost of generating each time-marching simulation are
around 72,500 and 5280 CPU hours using 128 processors
(2.7 GHz), respectively. However, the model cost is upfront and
once the model is created it can rapidly predict the aerodynamic
responses of many motions, all dened within the input
space range.
6.3. SACCON aerodynamics modeling
Most previous reduced order aerodynamic modeling efforts
have been limited to longitudinal motions with the angle of attack
Fig. 45. Half lazy-8 maneuver. (a) ground trajectory; (b) ight velocity; (c) aero angles; and (d) Euler angles; and (e) ight trajectory.
202
and pitch rate as input variables. Recently, models based on

indicial functions have been used for aerodynamic prediction of
both longitudinal and lateral forces and moments [112]. Here,
these models are reviewed for prediction of the SACCON aero-
dynamic forces/moments during 6DoF maneuvers in the subsonic

speed range.
An optimal control approach [162,163] was used to generate
6DoF maneuvers for the SACCON with the feasible solutions based
Fig. 46. Immelmann turn maneuver. (a) ight altitude; (b) ight velocity; (c) aero angles; (d) Euler angles; and (e) ight trajectory.
on the vehicle control and state constraints. The input angle of

attack of maneuvers is in the range of 101 to 101 with a Mach
number range of 0.10.5. The side-slip angle also ranges from 51
to 51. The side-force, yaw and roll moments are nearly linear with
203
side-slip for small values of side-slip and angles of attack below

151 as shown in Fig. 13. However, the generation of response
functions in the dened input space using CFD is still expensive
and makes the creation of a ROM very time consuming. A time-
Fig. 47. Aerodynamic modeling of half lazy-8 maneuver. (a) lift coefcient; (b) pitch moment coefcient; (c) side force coefcient; (d) roll moment coefcient; and (e) yaw
moment coefcient.
204
dependent surrogate model is presented to model SACCON

response functions for a new ight condition from some available
(observed) responses.
For surrogate modeling of response functions, a set of samples
including 33 points is dened on the and M space using full factorial
design. Also, it is assumed that the aerodynamic forces and moments
are symmetric about zero degrees angle of attack and hence only the
angles between zero and ten degrees are considered for response
function calculations. This assumption is not entirely correct but it aids
in reducing the computational cost of generating response functions.
The sample points are uniformly spaced over positive values for
Mach numbers of 0.1, 0.3, and 0.5 and are shown in Fig. 39.
Fig. 48. Aerodynamic modeling of Immelmann turn maneuver. (a) lift coefcient; (b) pitch moment coefcient; (c) side force coefcient; (d) roll moment coefcient and
(e) yaw moment coefcient.
All response computations started from a steady-state solution

and then advanced in time using second-order temporal accuracy.
The motion les were generated for step changes in aircraft
forcing parameters (angle of attack, side-slip angle, and angular
rates). These les dene the rotations and displacements at
discrete time instants and Cobalt then interpolates motion data
using cubic-splines and moves the grid for each computational
time step. The grid undergoes only translation motion for and
responses, where the relative velocity between grid and ow at
each time instant denes the angle of attack and side-slip. For
angular rate responses, the motions start from a steady-state
solution with zero degrees angle of attack and side-slip angle.
The grid then rotates and translates simultaneously. The rotation
corresponds to a unit step change in the angular rate, while the
translation motion is used in order to keep angles of attack and
side-slip zero during rotations.
In angle of attack response simulations the solutions start from
a steady-state condition at an angle of attack of i and a Mach
number of Mi, and then performs a small step in the angle of attack
for all t 4 0. In these calculations, Mi and i values correspond to
the samples shown in Fig 39, and the side-slip angle is zero
degrees at all times and the grid does not rotate at any time. The
response functions are then computed by taking the differences
between time-varying forces and moments occurring after the
step and the steady-state solution at i degrees, and dividing
them by the magnitude of the step . For a weakly nonlinear
system, the response will be nearly independent of the step
magnitude (assuming that i r i 1 ). The step value used
was a unit step.
The lift and pitch moment indicial responses to a unit step
change in the angle of attack from 0 are shown in Fig. 40
(a) and (b) for Mach numbers of 0.1, 0.3, and 0.5. The lift and pitch
moment are plotted against the nondimensional time s 2Vt=c.
Fig. 40(a) shows that the lift responses have a peak at s 0.
Likewise, the pitch moment predicts a negative peak at this time
as shown in Fig. 40(b). As the steady ow around the vehicle is
disturbed by the grid motion, a compression wave and an expansion wave are formed on the lower and upper surfaces of the
vehicle that cause a sharp peak in the responses [27]. As the
response time progresses, the waves begin to move away from the
vehicle, the lift starts to fall, and the pitch moment starts to
increase, and then the responses asymptotically reach the steadystate values. Note that nal time responses match with the static
data slopes at zero angle of attack. Fig. 40(a) and (b) also shows
that the initial peak becomes smaller for compressible ow. An
explanation is given by Leishman [106]; this is due to the
propagation of pressure disturbances at the speed of sound,
compared to the incompressible case, where the disturbances
propagate at innite speed.
Fig. 40(c) and (d) shows the lift and pitch moment responses at
different angles of attack at Mach number of 0.3. These gures
show that the initial values of responses are invariant with angle
of attack, but the transient trend and steady state values change
depending on the angle of attack. The changes in the indicial
functions are small since the lift and pitch moment are in the
nearly linear regime of angle of attack.
The lateral loads response to a unit step change in the side slip
angle at different angles of attack are shown in Fig. 41 for Mach
numbers of 0.1, 0.3, and 0.5. In these simulations, the solution
starts from a steady-state condition at zero degrees side slip angle
and an angle of attack of i at a Mach number of Mi, and then
iterates such that the side slip angle is held constant to one degree
and angle of attack is held constant to i for all t 40, where Mi and
i correspond to the samples shown in Fig. 39. Likewise, for the lift
and pitch moment, the initial peaks in lateral responses become
smaller for compressible ow. Fig. 41 shows that side-force
205
responses C y remain almost unchanged with the changes in

the angle of attack for the range of angles studied. This gure
shows that the yaw moment responses C n slightly change with
the changes in the angle of attack, but signicant differences are
found for the roll moment. The differences become more apparent
as the Mach number increases.
Typically, the angle of attack effects are negligible for the
responses due to the angular rates at low to moderate angles of
attacks. Fig. 42(a) and (b) shows the lift and pitch moment
responses respectively with a unit step change in pitch rate for
Mach numbers of 0.1, 0.3, and 0.5. Again there is an initial jump in
lift as the grid starts to rotate, and the value decreases as Mach
number increases. The lift response starts to fall a short time after
initial excitation and then it asymptotically reaches a steady-state
value, the so-called pitch dynamic derivative. Fig. 42(a) and
(b) shows that increasing Mach number results in the slight
decrease of the lift and pitch damping derivatives. These calculations, along with a time-dependent surrogate model, were used to
estimate the functions of C Lq t; M and C mq t; M in the ROM
equations. Also, the indicial functions with respect to roll and
yaw rates are shown in Figs. 43 and 44, respectively. These
calculations, along with a time-dependent surrogate model, were
used to estimate the functions of C Yp t; M, C lp t; M, C np t; M,
C Yr t; M, C lr t; M, and C nr t; M in the ROM equations.
The response modeling equations were used for prediction of
two SACCON maneuvers: a half Lazy-8 [164] and an Immelmann
turn [165]. In both maneuvers, the aircraft enters and terminates
the maneuver from a straight and level condition. The maneuvers
were generated using the DIDO code [166] to minimize the nal
maneuver time subject to vehicle aerodynamics, as well as mass
properties, state, and control constraints. The angle of attack and
side slip angles are limited to 101; 101 while the maximum
Mach number is 0.5. Fig. 45 shows that the Lazy-8 maneuver
makes a 1801 turn. The airplane starts a climb steeply to reduce
Fig. 49. Surface pressure solutions during Lazy-8 and Immelmann Turn maneuvers: (a) half Lazy-8 and (b) Immelmann turn.
206
ight speed as shown in Fig. 45(b). This reduced speed helps to

have a smaller radius turn and total traveled time [165]. Next, the
airplane starts to roll as the pitch angle decreases, where at 901
yaw angle, the vehicle is at zero pitch and maximum roll angle as
shown in Fig. 45(d). This is followed by a descent trajectory and
decreasing roll angle, increasing pitch angle, and increase in the

speed until the vehicle reaches the initial velocity and altitude. The
Immelmann turn as shown in Fig. 46 comprises a half loop with a
half roll at the end. The maneuver starts with a steep climb and
thus decreases the speed as shown in Fig. 46(b). At the maximum
Fig. 50. Target motion is a pitch harmonic motion as 71 sin t, M 0.9, k 0.01, and Re 2 106 . In (a) static data are shown with a solid line. The chord-wise pressure
distributions are shown for the wing sections at y=b=2 of 0.3, 0.6 and 0.9, where b is the wing span. (a) target motion; (b) point A; (c) point B; (d) point C; (e) point D;
(f) point E; and (g) point F.
pitch angle, the aircraft heading suddenly changes from 01 to 1801,

which makes the aircraft nal ight path exactly opposite of the
initial path. As the heading starts to increase, the aircraft performs
a half roll to level the wing as shown in Fig. 46(d). The nal
altitude is slightly higher than the starting altitude as shown in
Fig. 46(a).
The time-marching simulations of these maneuvers were calculated using RANS simulations with grid motion. The computations
start from a steady-state solution corresponding to the initial state of
207
the maneuver and then advance in time using second-order temporal

accuracy. The grid moved and rotated at each time step according to
the maneuver descriptions shown in Figs. 45 and 46. Cobalt reports
time-dependent aerodynamic forces and moments in an inertia axis;
these forces and moments are then transformed to the wind axis and
normalized by the reference area and length and the dynamic
pressure at each time step. Fig. 47 depicts the predicted aerodynamic
loads of half Lazy-8 maneuver. The comparisons between the created
ROM with the time-marching model show good agreements for all
Table 4
Comparison of model costs and errors to predict the X-31 target motion at M 0.9.
Volterra using spiral

Volterra using chirp
RBF using spiral
RBF using chirp
RBF using Schroeder
Linear indicial function
Nonlinear response function
SBRF in M space
Response function in M space
a
b
Costa
Error (%)
Comments
21,500
21,500
21,500
21,500
21,500
768
5400
344,000
27,000
1.42
3.47
3.24
1.10
1.53
12.67
0.48
0.98
0.48
Limited to the frequency of maneuver

Might be applied to other frequencies
Limited to the frequency of maneuver
Small amplitude motions with any frequency
Any amplitudeb and frequency
Any amplitude, Machb at a xed frequency
Any amplitude, Machb and frequency
CPU hours.
Amplitude and Mach number should be within the range of input space.
Fig. 51. Volterra reduced order modeling using spiral and chirp training maneuvers. The ow conditions of training maneuvers are M 0.9 and Re 2 106 . (a) spiral
maneuver; (b) ROM prediction; (c) chirp maneuver; and (d) ROM prediction.
208
coefcients. The predictions of aerodynamic loads for the Immelmann

turn are also shown in Fig. 48. Again, the ROM is in close agreement
with the time-marching model for all coefcients; small discrepancies
are found in the roll and yaw moments for maneuver times between
12 and 16 s, where the angle of attack is negative. This shows that the
roll and yaw moments of SACCON are not symmetric about zero
degrees angle of attack and hence the indicial functions need to
include negative angles of attack as well. Note that the cost of
generating each time-marching model is approximately 50,000 CPU
hours using 256 processors (2.7 GHz), but the model predictions are
Fig. 52. RBF reduced order modelingtraining maneuver are a spiral, a chirp, and a Schroeder motion. The ow conditions are M 0.9 and Re 2 106 . (a) spiral maneuver;
(b) ROM prediction; (c) chirp maneuver; (d) ROM prediction; (e) Schroeder maneuver; and (f) ROM prediction.
generated within a few seconds. Finally, the surface pressure distributions during maneuvers are shown in Fig. 49. The leading edge vortex
can be seen behind the leading edge on the upper wing around 101
angle of attack during the maneuvers.
6.4. X-31 aerodynamics modeling
All ROMs considered were evaluated for a similar target motion
to permit direct comparison of costs and errors. The objective of all
models is to predict the unsteady pitch moment resulting from a
sinusoidal pitch oscillation at a freestream Mach number of
0.9 and reduced frequency of k0.01. The amplitude of oscillation
is held constant at 71 and the mean angle of attack is zero degrees.
Fig. 50(a) shows the computed pitch moment coefcient, Cm, by
solving the RANS and SARC-DES turbulence model equations in a
time-accurate fashion. The cost of simulating three pitch cycles is
approximately 13,000 CPU hours using 256 processors (2.7 GHz).
Fig. 50(a) shows that the pitch moment curve of target motion
makes a nonlinear loop on the moment versus angle of attack
gure due to the occurrence of shock waves and vortices. This
gure also shows that the pitch moment curve is not symmetric
about zero degrees angle of attack. The moment curve shows a
negative slope during the pitch cycle such that it has more
negative slope values at negative angles of attack compared with
the slope values at positive angles of attack. Some ow features
209
during the pitch oscillation are shown in Fig. 50(b)(g) and will be
briey discussed here. In Fig. 50(b) the angle of attack is 2.11 and
a vortex can be seen emanating from the wing root on the lower
surface which spirals towards the wing tip. This vortex causes a
sharp negative pressure peak to occur close to the wing leading
edge as shown in the surface pressure plots of Fig. 50(b)(d).
Fig. 50(b) also shows that a shock wave is formed on the lower
surface of the wing which is nearly perpendicular to the fuselage
before it interacts with the leading edge vortex. At the minimum
angle of attack in the pitch cycle, i.e. 71, the leading edge
vortex becomes much stronger and the wing surface pressure
close to the leading edge drops further as shown in Fig. 50(c). This
gure also shows that as the angle of attack becomes smaller, the
shock moves downstream and therefore changes the pitch
moment curve slope. No vortices were observed on the wing
during pitching at positive angles of attack, but a vortex was
formed on the canard tip at the maximum angle of attack in the
pitch cycle, i.e. 71, as shown in Fig. 50(f). Fig. 50(e)(g) shows
that a shock wave is formed over the upper surface which is no
longer perpendicular to the fuselage and moves slowly with
increasing in the angle of attack during upstroke. The model costs
and errors to predict this target motion are summarized in Table 4
with more details given below:
For the identication of the Volterra kernels, the chirp and
spiral training maneuvers were generated using CFD (RANS and
Fig. 53. ROM using indicial functions. The ow conditions are M 0.9 and Re 2 106 . (a) linear responses; (b) ROM based on linear responses; (c) non-linear responses;
and (d) ROM based on non-linear responses.
210
SARC-DES turbulence model) as the source of the data. The

variations of angle of attack with time for these maneuvers are
shown in Fig. 51(a) and (c). Both maneuvers ran for 2.4 s of
physical time and started from a steady-state solution. The chirp
maneuver has an amplitude of 71 starting from zero degrees angle
of attack and pitching with a frequency of 1 Hz at t0. The chirp
motion frequency increases linearly with time. The spiral maneu-
ver has an initial amplitude of 3.51, starting from zero degrees

angle of attack and pitching at constant frequency of k 0.01. The
oscillation amplitude in the spiral motion increases as time
progresses. Note that the spiral maneuver is at the reduced
frequency of the maneuver to be predicted. The cost of generating
each training maneuver is approximately 21,500 CPU hours using
256 processors.
Fig. 54. ROM using non-linear indicial functions for target motions at different angles of attack and frequency. The ow conditions are M 0.9 and Re 2 106 .
(a) 41 sin t, k 0.01; (b) 61 sin t, k 0.01; (c) chirp maneuver; (d) spiral maneuver; and (e) Schroeder maneuver.
The rst and second order kernels of the Volterra model were
estimated from time-history simulations of chirp and spiral training maneuvers. These estimations were used next to predict the
target motion. The ROM predictions based on spiral and chirp
training maneuvers are compared with target data in Fig. 51
(b) and (d). The comparisons show a good agreement with CFD
data for a ROM identied from spiral data (error norm is 1.42%),
but the ROM identied from chirp data does not match everywhere, in particular, around the maximum and minimum angles of
attack. The instantaneous frequency in the chirp maneuver varies
with time and hence it might not have sufcient information to
identify the Volterra kernels corresponding to a swept amplitude
motion at constant frequency. However, the ROM based on chirp
data possibly could be used for predicting aerodynamic responses
from pitch oscillations at other frequencies within the range of
input, but the ROM based on the spiral is possibly only valid for the
motions at a xed reduced frequency.
The generated chirp and spiral training maneuvers were also
used to nd a mapping between the pitch moment coefcient and
the instantaneous pitch motion variables. This mapping was next
learned using a RBF neural network. Also, a Schroeder maneuver
was dened by a multi-stage frequency sweep. This maneuver
started from an initial angle of attack of 4.951. The number of
frequencies in the maneuver, N, was set to 20 with an initial
amplitude of 71. This maneuver ran for 2.4 s of physical time as
well and is shown in Fig. 52(e). The aircraft responses to these
three maneuvers were generated using the URANS equations and
the SARC-DES turbulence model. The training data were next
normalized using the mean and standard deviation of each input.
The data were then rearranged according to Eq. (30) and the RBF
network performance was tested for different values of m and n,
with a performance error threshold of 1 10 6 . All networks
computed converged to the threshold error. The results showed
that using m n 4 is sufcient for modeling the motions studied.
The trained networks were then tested against the target motion;
the ROM predictions are shown in Fig. 52(b), (d) and (f). These
gures show that the predicted ROM values agree well with the
time-marching solution, although the ROM based on chirp and
Schroeder maneuvers showed better accuracy than models based
on the spiral maneuver.
The indicial pitch moment responses of the X-31 aircraft with a
unit step change in angle of attack and pitch rate are shown in
Fig. 53(a). These functions correspond to the xed Mach number of
0.9. In C L simulations, the angle of attack is zero degrees at t 0
and is held constant to one degree for all other times. In CLq
simulations, the grid starts to pitch up with a normalized pitch
rate of q 1 rad at t 0 and the angle of attack is held to zero
degrees during simulations with the aid of grid translation. All
computations started from a steady-state solution and then are
advanced in time using second-order temporal accuracy with ve
Newton subiterations. As shown in Fig. 53(a), the pitch moment
responses have a negative peak at t 0 followed by an increasing
trend. As the response time progresses the pitch moment
responses start to increase and then asymptotically reach the
steady-state values. The cost of generating each indicial function is
around 384 CPU hours using 256 processors. This model has much
lower cost than RBF and Volterra theory, because response functions reach the steady state solutions in one order of magnitude
less time than the time used to complete training maneuvers.
A linear ROM was created using Eq. (23) and used for prediction of target maneuver. The results are compared with timemarching model in Fig. 53(b). The gure shows that linear ROM
fails to accurately predict the pitch moment values at all angle of
attack. The error norm value is 12.67%. The functions of C L vary
largely with angles of attack in the transonic speed range and thus
a linear ROM cannot predict these effects.
211
Next, the X-31 C L functions were simulated at different angles

of attack and at a freestream Mach number of 0.9 and are shown in
Fig. 53(c). Note that the pitch moment slope is not symmetric with
zero degrees angle of attack and hence the simulations included
both positive and negative angle of attack responses. The total cost
of generating a nonlinear ROM is now increased to approximately
5400 CPU hours (this is still cheaper than the cost of simulating
chirp, spiral and Schroeder maneuvers).
Fig. 53(c) shows that the responses at the initial time are
invariant with angle of attack, but the intermediate and nal
values change depending on the angle of attack. Fig. 53(c) shows
that the pitch moment responses have more negative values than
positive angles of attack due to vortex formation on the lower
surface of the wing. A nonlinear ROM was created, and then using
a linear interpolation scheme, the prediction of the target maneuver was evaluated. Fig. 53(d) shows that the nonlinear ROM
predictions agree very well with the time-marching simulation
values with an error of 0.48%.
Such a nonlinear ROM could be used for computing the pitch
moment responses from many other motions with different
amplitudes and frequencies. For example, the ROM was used to
predict two pitch oscillations with 41 and 61 amplitude at M 0.9
and k 0.01. The predictions are compared with time-marching
solutions in Fig. 54(a) and (b). Again a very good match was found.
Also, the ROM predictions were evaluated for the chirp, spiral and
Fig. 55. Samples and training pitch motions for a SBRF model. (a) samples design
for SBRF model and (b) training pitch motions for SBRF model.
212
Schroeder maneuvers used in RBF work. Fig. 54(c)(e) shows that

even for this varying amplitude and frequency motions, the
created ROM matches very well with the CFD data.
For the generation of the SBRF model, time-accurate simulations
were pre-computed for various combinations of pitch amplitude and
Mach number at a xed reduced frequency. The two-dimensional
parameter space was lled using Design of Experiment methods and
is shown in Fig. 55(a). Also, the pitch motion simulations of all
samples are shown in Fig. 55(b). The SBRF model was used for the
prediction of the pitch moment coefcient time history for sinusoidal
forced motions about zero degrees angle of attack, amplitude of 71
and values of Mach number of 0.78, 0.825, and 0.88. Model
predictions are compared to time-accurate results in Fig. 56. Tests
were performed to evaluate the dependency of the model predictions on the number of previous steps in the inputs (angle of attack
time history, rst and second time derivatives, and Mach number)
and output (the prediction itself). No signicant dependence was
found for values of m and n up to 2. This may be attributed to the
small time step increments used in the time-accurate simulations.
Although the method robustness could degrade for higher values of
m and n, it is considered relevant that the predictions are unaffected
for a range of values. In this work, the Kriging interpolation was used
to approximating the mapping function between inputs and outputs.
A good agreement is noted in Fig. 56 for all ight conditions, with the
model predictions being generated in few seconds.
An issue regarding the SBRF model is that the cost of simulating

three pitch cycles for each sample shown in Fig. 55(a) is around
13,000 CPU hours using 256 processor and the model still cannot
predict the aerodynamic responses to motions at other frequencies.
A ROM based on indicial functions, along with a time-dependent
surrogate approach, is proposed for aerodynamics modeling in the
angle of attack/Mach number/frequency space. In this model, the
indicial functions in the angle of attack and Mach number space are
interpolated from some available samples. A ROM based on these
functions is still cheaper than the time-marching model and SBRF
model because the indicial functions eliminate the need of repeating
calculation for each frequency.
The X-31 motions considered encompass and M values in the
range of 71; 71 and [0.75, 0.9], respectively. A set of samples
including 56 points is dened on the and M space using factorial
design. These points are shown in Fig. 57(a).
The indicial functions with respect to angle of attack are
calculated using the CFD and grid motion approach for each
sample condition. The pitch rate indicial functions are calculated
for a unit step change in the pitch rate for each Mach number
shown in Fig. 57(a). The total cost of generating all functions is
now approximately 27,000 CPU hours. The calculated indicial
functions due to a unit step change in angle of attack are shown
in Fig. 57(b) for each Mach number in the sample design. This
gure shows that the pitch moment initial, intermediate, and nal
Fig. 56. Transonic loads modeling in Mach number/angle of attack space at xed reduced frequency. The ROM is a SBRF model. In above is angular velocity and k c=2V
is reduced frequency. (a) 71 sin t, k 0.01, M 0.78; (b) 71 sin t, k 0.01, M 0.825; and (c) 71 sin t, k 0.01, M 0.88.
loadings are different at each Mach number. The initial peak in the
pitch moment becomes smaller for higher Mach number. Fig. 57
(c) also shows that pitch rate indicial functions decrease as the
Mach number increases.
A new ROM is now created along with a time-dependent
surrogate model to determine the terms in Eq. (24) at each time
step. The validity of the ROM is tested for several motions in the
angle of attack/frequency/Mach number space and compared with
time-accurate CFD simulations in Fig. 58. This gure shows that
the ROM predictions agree well with the CFD data, although small
discrepancies are found in the high-speed motions. This is likely
due to the sample design used with a uniform spacing and the fact
that the pitch moment changes suddenly at high speeds. More
samples at high speeds could improve the model predictions.
7. Conclusions
Aircraft stability and control analysis using CFD requires a very
large number of CFD simulations to determine appropriate forcing
parameters within the frequency/amplitude/Mach number space.
213
Typically, the time-accurate CFD simulations start from a steady

state solution and are marched (iterated) in pseudo time within
each physical time step using a dual-time stepping scheme. Also,
to have a free decay response to the initial grid perturbation, it is
often necessary to march time-accurate solutions for several
oscillations. The aircraft congurations used in this work have
highly swept slender wings resulting in complex vortical ow
under various conditions. A highly rened mesh and small time
step are required to accurately resolve the unsteady ow around
the aircraft in space and time. Because of the combination of large
grids and small time steps, and a large number of simulations, the
full-order modeling approach is expensive for stability and control
analysis of aircraft. This paper reviews the use of reduced order
models that signicantly reduces the CFD simulation time
required to create a full aerodynamics database, making it possible
to accurately model aircraft static and dynamic characteristics
from a limited number of time-accurate CFD simulations.
The models considered were based on linear and nonlinear
indicial response theory, Volterra theory, radial basis functions,
and a surrogate-based recurrence framework. The response functions were directly calculated from unsteady RANS simulations
Fig. 57. Time-dependent surrogate modeling of indicial functions. (a) samples design for indicial functions; (b) angle of attack indicial functions; and (c) pitch rate indicial
functions.
214
starting from an initial steady-state condition with a prescribed

grid motion. An important feature of this approach is uncoupling
the effects of angle of attack and side-slip from angular rate
responses. A method to efciently reduce the number of step
function calculations within the angle of attack/Mach number

space was described. This method uses a time-dependent surrogate model to t the relationship between ight conditions (Mach
number and angle of attack) and step functions calculated for a
Fig. 58. Transonic loads modeling in Mach number/angle of attack/frequency space. The ROM is based on a time-dependent surrogate model that approximates the nonlinear indicial functions at different ight conditions. In above is angular velocity and k c=2V is reduced frequency.
limited number of samples. These responses are a function of time

and include the effects of input time derivatives, therefore they
can be used for any frequency of interest. Models based on
response functions can predict the initial transient behavior seen
in the CFD solutions. However, these models are limited to weakly
nonlinear systems.
An indirect method was described to estimate the nonlinear
Volterra kernels from time-accurate computational uid dynamic
simulations of some special training maneuvers. These maneuvering simulations were also used to estimate the unknown parameters in a model based on RBF. The results showed that both
model predictions largely depend on the training maneuver type.
A Design of Experiment method was used to generate several
pitching motions at different amplitudes and free-stream Mach
numbers. The model based on a SBRF then approximated the
aerodynamic responses induced by pitching motions at new
amplitudes and Mach numbers. However, this model is usually
more expensive than other models to generate. Overall, the
reduced order models were found to produce accurate results
for low to moderate angles of attack and at reasonable
computational cost.
Acknowledgments
Mehdi Ghoreyshi is supported by USAFA under Contract
FA70001320018; their nancial support is gratefully acknowledged. Acknowledgements are expressed to the DoD High Performance Computing Modernization Program (HPCMP) for providing
computer time. The SACCON and X-31 geometries and experimental data were provided by the NATO RTO Task Group AVT-161
on Assessment of Stability and Control Prediction Methods for
NATO Air & Sea Vehicles. The authors appreciate the support
provided by the High Performance Computing Research Center
at USAFA.
References
[1] Shilu C, Xiaofei X. Parameter identication of unsteady aerodynamic forces.
AIAA paper 92-4505; August 1992.
[2] Pamadi BN. Performance, stability, dynamics, and control of airplanes. AIAA
education series, AIAA, Reston, Virginia; 2004.
[3] Myatt JH, Schumacher CJ, McKeehen PD, Bufngton JD. Modeling and
simulation of unsteady aerodynamic effects on a tailless aircraft. AIAA paper
98-4455; August 1998.
[4] Abramov NB, Goman MG, Greenwell DI, Khrabrov AN. Two-step linear
regression methods for identication of high incidence unsteady aerodynamic model. AIAA paper 2001-4080; August 2001.
[5] Zhang W, Wang B, Ye Z, Quan J. Efcient method for limit cycle utter
analysis by nonlinear aerodynamic reduced-order models. AIAA J 2012;50
(5):101928.
[6] Venkatesan C, Friedmann PP. New approach to nite-state modeling of
unsteady aerodynamics. AIAA J 1986;24(12):188997.
[7] Kyle H, Lowenberg M, Greenwell D. Comparative evaluation of unsteady
aerodynamic modeling approaches. AIAA paper 2004-5272; August 2004.
[8] Greenwell DI. A review of unsteady aerodynamic modelling for ight
dynamics of manoeuvrable aircraft. AIAA paper 2004-5276; August 2004.
[9] Lucia DJ, Beran PS, Silva WA. Reduced-order modeling: new approaches for
computational physics. Prog Aerosp Sci 2004;40(1):51117.
[10] Silva WA. Application of nonlinear systems theory to transonic unsteady
aerodynamics responses. J Aircr 1993;30(5):6608.
[11] Schtte A, Einarsson G, Raichle A, Schoning B, Mnnich W, Forkert T.
Numerical simulation of manoeuvreing aircraft by aerodynamic, ight
mechanics, and structural mechanics coupling. J Aircr 2009;46(1):5364.
[12] Ghoreyshi M, Vallespin D, Da Ronch A, Badcock KJ, Vos J, Hitzel S. Simulation
of aircraft manoeuvres based on computational uid dynamics. AIAA paper
20108239; August 2010.
[13] Ghoreyshi M, Jirsek A, Cummings RM. CFD modeling for trajectory predictions of a generic ghter conguration. AIAA paper 20116523; August 2011.
[14] Sitaraman J, Baeder JD. Computational-uid-dynamics-based enhanced indicial aerodynamic models. J Aircr 2004;41(4):798810.
[15] Chin S, Lan CE. Fourier functional analysis for unsteady aerodynamic
modeling. AIAA J 1992;30(9):225966.
215
[16] Liebe R. Flow phenomena in nature: a challenge to engineering design. Great

Britain: WIT Press; 2007.
[17] Stark VJE. Indicial coefcients for a cropped delta wing in incompressible
ow. J Aircr 1985;23(5):3705.
[18] Lisandrin P, Carpentieri G, van Tooren M. Investigation over CFD-based
models for the identication of nonlinear unsteady aerodynamics responses.
AIAA J 2006;44(9):204350.
[19] McDaniel DR, Cummings RM, Bergeron K, Morton SM, Dean JP. Comparisons
of CFD solutions of static and maneuvering ghter aircraft with ight test
data. In: 3rd international symposium on integrating CFD and experiments
in aerodynamics, U.S. Air Force Academy, Air Force Academy, CO; June 2007.
[20] Raveh DE. Reduced-order models for nonlinear unsteady aerodynamics.
AIAA J 2001;39(8):141729.
[21] Tischler MB. Aircraft and rotorcraft system identication: engineering
methods with ight test examples. AIAA education series, Reston, Virginia;
2006.
[22] Silva W. Identication of noc systems based on the Volterra theory: progress
and opportunities. Nonlinear Dyn 2005;39(12):2562.
[23] Marzocca P, Librescu L, Kim D, Lee I, Schober S. Generalized transonic
unsteady aerodynamics via computational-uid-dynamics/indicial approach.
AIAA J 2005;43(4):9157.
[24] Ballhaus WF, Goorjian PM. Computation of unsteady transonic ows by the
indicial method. AIAA J 1978;16(2):11724.
[25] Tobak M. Chapman GT, Schiff LB. Mathematical modeling of the aerodynamic
characteristics in ight dynamics. NASA TN-85880; 1984.
[26] Tobak M, Chapman GT. Nonlinear problems in ight dynamics involving
aerodynamic bifurcations. NASA TN-86706; 1985.
[27] Ghoryeshi M, Jirsek A, Cummings RM. Computational investigation into the
use of response functions for aerodynamic loads modeling. AIAA J 2012;50
(6):131427.
[28] Hall KC, Thomas JP, Dowell EH. Reduced-order modelling of unsteady smalldisturbance using a frequency-domain proper orthogonal decomposition
technique. AIAA paper 1999655; January 1999.
[29] Willcox K, Peraire J. Balanced model reduction via the proper orthogonal
decomposition. AIAA J 2002;40(11):232330.
[30] Silva WA, Bartels RE. Development of reduced-order models for aeroelastic
analysis and utter prediction using the CFL3Dv6.0 code. J Fluids Struct
2004;19(6):72945.
[31] Silva WA. Discrete-time linear and nonlinear aerodynamic impulse
responses for efcient CFD analysis [Ph.D. dissertation]. VA: Faculty of the
Department of Applied Science, The College of William and Mary in Virginia;
September 1997.
[32] Da Ronch A, McCracken A, Badcock KJ, Ghoreyshi M, Cummings RM.
Modeling unsteady aerodynamic loads. AIAA paper 20116524; August 2011.
[33] Glaz B, Liu L, Friedmann PP. Reduced-order nonlinear unsteady aerodynamic
modeling using a surrogate-based recurrence framework. AIAA J 2010;48
(10):241829.
[34] Ghoreyshi M, Cummings RM, Da Ronch A, Badcock KJ. Transonic aerodynamic loads modeling of X-31 aircraft pitching motions. AIAA J, also available
as AIAA paper 2012-3127, in press.
[35] Jirsek A, Jeans TL, Martenson M, Cummings RM, Bergeron K. Improved
methodologies for maneuver design of aircraft stability and control simulations. AIAA paper 2010515; January 2010.
[36] Ghoreyshi M, Jirsek A, Cummings RM. Computational approximation of
nonlinear unsteady aerodynamics using an aerodynamic model hierarchy.
Aerosp Sci Technol 2013;28(1):17690.
[37] Cummings RM, Jirsek A, Ghoreyshi M, Post ML. Numerical prediction of
stability derivatives with various reduced order models. AVT-189 specialists
meeting on assessment of stability and control prediction methods for NATO
air and sea vehicles, Portsmouth, England; October 2011.
[38] Goman MG, Khrabov AN. State-space representation of aerodynamic characteristics of an aircraft at high angles of attack. AIAA paper 924651; August
1992.
[39] Kyle H, Lowenberg M, Greenwell D. Comparative evaluation of unsteady
aerodynamic modeling approaches. AIAA paper 20045272; August 2004.
[40] Ashley H. Some recent developments in interference theory for aeronautical
applications. MIT report; July 1963.
[41] Whitehead DS. Force and moment coefcients for vibrating aerofoils in
cascade. Aeronautical Research Council Report and Memoranda, Report No.
3254; 1962.
[42] Ashley H, Brunelle E, Moser HH. Unsteady ow through helicopter rotors. Z
Angew Math Phys 1958;9(5):5780.
[43] Miller RH. Rotor blade harmonic air loading. AIAA J 1964;2(7):125469.
[44] Daughaday H, Piziali RA. An improved computational model for predicting
the unsteady aerodynamic loads of rotor blades. J Am Helicopter Soc 1966;11
(41):310.
[45] Hall KC, Ekici K, Thomas JP, Dowell EH. Harmonic balance methods applied
to computational uid dynamics problems. Int J Comput Fluid Dyn 2013;27
(2):5267.
[46] Sicot F, Gudeney T, Dufour G. Time-domain harmonic balance method for
aerodynamic and aeroelastic simulations of turbomachinery ows. Int J
Comput Fluid Dyn 2013;27(2):6878.
[47] McCracken A, Da Ronch A, Timme S, Badcock KJ. Solution of linear systems in
Fourier-based methods for aircraft applications. Int J Comput Fluid Dyn
2013;27(2):7987.
216
[48] Vilmina S, Lorrain E, Tartinville B, Capron A, Hirsch C. The nonlinear

harmonic method: from single stage to multi-row effects. Int J Comput Fluid
Dyn 2013;27(2):8899.
[49] Gerolymos GA. Filtered chorochronic interface as a capability for 3-D
unsteady throughow analysis of multistage turbomachinery. Int J Comput
Fluid Dyn 2013;27(2):10017.
[50] He L. Fourier spectral modelling for multi-scale aero-thermal analysis. Int J
Comput Fluid Dyn 2013;27(2):11829.
[51] He L. Fourier methods For turbomachinery applications. Prog Aerosp Sci
2010;46(1):32941.
[52] Dean J, Clifton J, Bodkin D, Morton S, McDaniel D. Determining the
applicability and effectiveness of current CFD methods in store certication
activities. AIAA paper 2010-1231; January 2010.
[53] Morton S, McDaniel D, Sears D, Tillman B, Tuckey T. Rigid, maneuvering, and
aeroelastic results for Kestrel a CREATE simulation tool. AIAA paper 20101233; January 2010.
[54] Lillian C, Morton S, McDaniel D. Aircraft loads characteristics determined by
system and proper orthogonal decomposition of CFD simulations. AIAA
paper 2010-4368; July 2010.
[55] Dean J, Morton S, McDaniel D, Clifton J, Bodkin D. Aircraft stability and
control characteristics determined by system identication of CFD simulations. AIAA paper 2008-6378; August 2008.
[56] Morelli EA. System IDentication programs for AirCraft (SIDPAC). AIAA paper
20024704; August 2002.
[57] Skujins T, Cesnik CES. Toward an unsteady aerodynamic ROM for multiple
Mach regimes. AIAA paper 2012-1708; April 2012.
[58] Skujins T, Cesnik CES. Reduced-order modeling of hypersonic unsteady
aerodynamics due to multi-modal oscillations. AIAA paper 2011-2341; April
2011.
[59] Skujins T, Cesnik CES. Reduced-order modeling of hypersonic vehicle
unsteady aerodynamics. AIAA paper 2010-8127; August 2010.
[60] Glaz B, Liu L, Friedmann PP. A surrogate based approach to reduced-order
dynamic stall modeling. AIAA paper 2010-3042; April 2010.
[61] Hall KC, Kielb RE, Ekici K, Thomas JP. Recent advancements in turbomachinery aeroelastic design analysis. AIAA paper 2005-14; January 2005.
[62] Thomas JP, Dowell EH, Hall KC. Three-dimensional transonic aeroelasticity
using proper orthogonal decomposition based reduced order models. AIAA
paper 2001-1526; April 2001.
[63] Balajewicz M. Reduced order modeling of transonic utter and limit cycle
oscillations using the pruned Volterra series. AIAA paper 2010-2725; April
2010.
[64] Balajewicz M, Dowell E. Reduced-order modeling of utter and limit-cycle
oscillations using the sparse Volterra series. J Aircr 2012;49(6):180312.
[65] Lieu T, Farhat C. Adaptation of aeroelastic reduced-order models and
application to an F-16 conguration. AIAA J 2007;45(6):124457.
[66] Jameson A, Alonso J, McMullen M. Application of a non-linear frequency
domain solver to the Euler and NavierStokes equations. AIAA paper 200213625; January 2002.
[67] McMullen M, Jameson A, Alonso J. Acceleration of convergence to a periodic
steady state in turbomachinery ows. AIAA paper 2001-0152; January 2001.
[68] Woodgate MA, Badcock KJ. Implicit harmonic balance solver for transonic
ow with forced motions. AIAA J 2009;47(4):893901.
[69] Da Ronch A, McCracken AJ, Badcock KJ, Widhalm M, Campobasso MS. Linear
frequency domain and harmonic balance predictions of dynamic derivatives.
J Aircr 2013;50(3):694707.
[70] Olcer FE, Prasad JVR, Sankar LN, Bain JJ, Zhao J, He C. Development and
evaluation of reduced order models of on-blade control for integrated ight
and rotor control. In: Proceedings of the American Helicopter Society, 66th
Annual Forum, Phoenix, Arizona; May 2010.
[71] Marques FD, Anderson J. Identication and prediction of unsteady transonic
aerodynamic loads by multi-layer functionals. J Fluids Struct 2011;1
(6):83106.
[72] Faller WE, Schreck SJ. Neural networks: applications and opportunities in
aeronautics. Prog Aerosp Sci 1996;32(5):43356.
[73] Faller WE, Schreck SJ. Real-time prediction of unsteady aerodynamics:
application for aircraft control and maneuverability enhancement. IEEE
Trans Neural Netw 1995;6(6):14618.
[74] Jose AI, Baeder JD. Steady and unsteady aerodynamic modeling of trailing
edge aps with overhang and gap using CFD and lower order models. AIAA
paper 2009-1071; January 2009.
[75] Leishman JG. Subsonic unsteady aerodynamics caused by gusts using the
indicial method. J Aircr 1996;33(5):86979.
[76] Leishman JG, Nguyen KQ. State-space representation of unsteady airfoil
behavior. AIAA J 1990;28(5):83644.
[77] Raveh DE. CFD-based gust response analysis of free elastic aircraft. AIAA
paper 2009-2539; May 2009.
[78] Zaide A, Raveh DE. Numerical simulation and reduced-order modeling of
airfoil gust response. AIAA paper 2005-5128; June 2005.
[79] Beran P, Stanford B, Kurdi M. Sensitivity analysis for optimization of dynamic
systems with reduced order modeling. AIAA paper 2010-1503; January 2010.
[80] Murman SM, Chan WM, Aftosmis MJ, Meakin RL. An interface for specifying
rigid-body motions for CFD applications. AIAA paper 2003-1237; January
2003.
[81] Murman SM, Aftosmis MJ, Berger MJ. Simulations of 6-DOF Motion with a
Cartesian method. AIAA paper 2003-1246; January 2003.
[82] Korenberg MJ, Hunter IW. The identication of nonlinear biological systems:
Volterra kernel approaches. Ann Biomed Eng 1996;24:25068.
[83] Theodorsen T. General theory of aerodynamic instability and the mechanism
of utter. NACA Technical report 496; 1935.
[84] Sears WR, von Karman T. Airfoil theory for non-uniform motion. J Aeronaut
Sci 1938;5(10):37990.
[85] Kssner HG. Zusammenfassender Bericht ber den Instationaren Auftrieb
von Flgeln. Luftfahrtforschung 1939;13(12):41024.
[86] Sears WR. Some aspects of non-stationary airfoil theory and its practical
applications. J Aeronaut Sci 1941;8(1):1048.
[87] Wagner H. ber die Entstehung des dynamischen Auftriebes von Traggeln. Z Angew Math Mech 1925; 1(Book 1):1735.
[88] Wright JR, Cooper JE. Introduction to aircraft aeroelasticity and loads. UK:
John Wiley & Sons Ltd.; 2007.
[89] Lomax H. Indicial aerodynamics. AGARD manual of aeroelasticity, part 2;
October 1960 [chapter 6].
[90] Hodges DH, Pierce GA. Introduction to structural dynamics and aeroelasticity. Cambridge University Press; 2002.
[91] Chen X, Kareem A. Advances in modeling aerodynamic forces on bridge
decks. J Eng Mech 2002;128(1):1193205.
[92] Volterra V. Theory of functionals. London: Blackie; 1930.
[93] Silva WA. Identication of linear and nonlinear aerodynamic impulse
responses using digital lter techniques. NASA Technical report 112872.
Hampton, Virginia; August 1997.
[94] Wray J, Green GR. Calculation of the Volterra kernels of non-linear dynamic
systems using an articial neural network. Biol Cybern 1994;71:18795.
[95] Sechi F, Bujatti M. Solid-state microwave high-power ampliers. Norwood,
MA, USA: Artech House INC; 2009.
[96] Wiener N. Nonlinear problems in random theory. Cambridge, MA: MIT Press;
1958.
[97] Barrett JF. The use of functionals in the analysis of non-linear physical
systems. J Electr Control 1963;15:567615.
[98] Kalman RE. Pattern recognition properties of multilinear response functions.
J Control Cybern 1980;9:531.
[99] Stark L. Neurological control system: studies in bioengineering. New York:
Plenum Press; 1968.
[100] Weiner DD, Spina JF. Sinusoidal analysis and modeling of weakly nonlinear
circuits. New York: Van Nostrand Reinhold; 1980.
[101] Hunter IW, Korenberg MJ. The identication of nonlinear biological systems:
Wiener and Hammerstein cascade models. J Biol Cybern 1986;55:13544.
[102] Korenberg MJ. A fast orthogonal search method for biological time-series
analysis and system identication. In: Proceedings of IEEE international
conference on systems, man, and cybernetics; 1989. p. 45965.
[103] Korenberg MJ. A robust orthogonal for system identication and time-series
analysis. J Biol Cybern 1989;60:26776.
[104] Jirsek A, Cummings RM. Reduced order modeling of X-31 wind tunnel
model aerodynamic loads. Aerospace Science and Technology 2012;20
(1):5260.
[105] Balajewicz M, Nitzsche F, Feszty D. Application of multi-input Volterra theory
to nonlinear multi-degree-of-freedom aerodynamic systems. AIAA J 2010;48
(1):5662.
[106] Leishman JG. Indicial lift approximations for two-dimensional subsonic ow
as obtained from oscillatory measurements. J Aircr 1993;30(3):34051.
[107] Duffy DG. Advanced engineering mathematics with MATLAB. second ed.
Florida: Chapman and Hall, CRC; 2003.
[108] Reisenthel PH. Development of a nonlinear indicial model using response
functions generated by a neural network. AIAA paper 970337; January 1997.
[109] Reisenthel PH, Bettencourt MT. Data-based aerodynamic modeling using
nonlinear indicial theory. AIAA paper 970337; January 1999.
[110] Findeisen D. System dynamics and mechanical vibrations: an introduction.
Berlin: Springer-Verlag; 2000.
[111] Roskam J. Airplane design. Kansas, USA: Roskam Aviation and Engineering
Corporation; 1990.
[112] Ghoreyshi M, Cummings RM. Unsteady aerodynamics modeling for aircraft
maneuvers: a new approach using time-dependent surrogate modeling. AIAA
paper 2012-3327; June 2012.
[113] Babuska R. Fuzzy modeling for control. Norwell, USA: Kluwer Academic
Publishers Group; 1998.
[114] Arbib MA. The handbook of brain theory and neural networks. second edThe
MIT Press; 2003.
[115] Levin AU, Narendra KS. Control of nonlinear dynamical systems using neural
networks: controllability and stabilization. IEEE Trans Neural Netw 1993;4
(2):192206.
[116] Krantz SG, Parks HR. The implicit function theorem: history, theory, and
applications. Boston: Birkhauser; 2002.
[117] Da Ronch A, Ghoreyshi M, Badcock KJ. On the generation of ight dynamics
aerodynamic tables by computational uid dynamics. Prog Aerosp Sci
2011;47:597620.
[118] Jategaonkar RV. Flight vehicle system identication. AIAA educational series,
vol. 216, Reston, VA; 2006.
[119] George DA. Continuous nonlinear systems. Technical report 335. Cambridge,
Massachusetts: M.I.T. Research Laboratory of Electronics; 1959.
[120] Schetzen M. Measurement of the kernels of a nonlinear system of nite
order. Int J Control 1965;1:25163.
[121] Milanece A, Marzocca P. Volterra kernels identication using continuous time

impulses applied to a nonlinear aeroelastic problem. AIAA paper 2009-2321;
May 2009.
[122] Tromp J, Jenkins JE. A Volterra kernel identication scheme applied to
aerodynamic reactions. AIAA paper 1990-2803; August 1990.
[123] Bisplinghoff RL, Ashley H, Ralfman RL. Aeroelasticity. Mineola, NY: Dover
Publications; 1996.
[124] Garrick IE, H RW. Historical development of aircraft utter. J Aircr 1981;18
(11):897912.
[125] Librescu L, Song O. Thin-walled composite beams: theory and application.
Dordrecht, Netherland: Springer; 2006.
[126] Mazelsky B, Drischler JA. Numerical determination of indicial lift and
moment functions for a two-dimensional sinking and pitching airfoil at
mach number 0.5 and 0.6. NACA technical note, 2739; 1952.
[127] Dowell EH. A simple method for converting frequency domain aerodynamics
to the time domain. NASA report TM-81844; 1980.
[128] Magnus RJ. Calculation of some unsteady transonic ows about the NACA
64006 and 64A010 airfoils. AFFDL-TR-77-46; July 1977.
[129] Singh R, Baeder JD. Direct calculation of three-dimensional indicial lift
response using computational uid dynamics. J Aircr 1997;34(4):46571.
[130] Ueda T, Dowell EH. Flutter analysis using nonlinear aerodynamic forces. AIAA
J 1984;21(2):1019.
[131] Ghoreyshi M, Badcock KJ, Woodgate MA. Accelerating the numerical generation of aerodynamic models for ight simulation. J Aircr 2009;46
(3):97280.
[132] Ghoryeshi M, Post ML, Cummings RM. CFD calculation of aerodynamic
indicial functions for a generic ghter conguration. ITEA J Test Eval
2012;33(6):34864.
[133] Grtz S, McDaniel D, Morton SA. Towards an efcient aircraft stability and
control analysis capability using high-delity CFD. AIAA paper 2007-1053;
January 2007.
[134] O'Neill C, Arena A. Time domain training signals comparison for computational uid dynamics based aerodynamic identication. J Aircr 2005;42
(2):4218.
[135] Murphy PC, Klein V. Validation of methodology for estimating aircraft
unsteady aerodynamics parameters from dynamic wind tunnel tests. AIAA
paper 20035397; August 2003.
[136] Morelli EA. Real time parameters estimation in the frequency domain. AIAA
paper 994043; August 1999.
[137] Strang WZ, Tomaro RF, Grismer MJ. The dening methods of cobalt: a
parallel, implicit, unstructured Euler/NavierStokes ow solver. AIAA paper
19990786; 1999.
[138] Gottlieb JJ, Groth CPT. Assessment of Riemann solvers for unsteady onedimensional inviscid ows of perfect gasses. J Fluids Struct 1998;78
(2):43758.
[139] Tomaro RF, Strang WZ, Sankar LN. An implicit algorithm for solving time
dependent ows on unstructured grids. AIAA paper 1997-0333; 1997.
[140] Spalart PR, Allmaras SR. A one equation turbulence model for aerodynamic
ows. AIAA paper 1992-0439; January 1992.
[141] Wilcox DC. Reassessment of the scale determining equation for advanced
turbulence models. AIAA J 26 November 1988; 1299310.
[142] Menter F. Eddy viscosity transport equations and their relation to the k
model. ASME J Fluids Eng 1997;119:87684.
[143] Tyssel L. Hybrid grid generation for complex 3D geometries. In: Proceedings
of the 7th international conference on numerical grid generation in computational eld simulation; 2000. p. 33746.
217
[144] Tyssel L. The TRITET grid generation system. International Society of Grid
Generation (ISGG). In: Proceedings of the 10the international conference on
numerical grid generation in computational eld simulations; 2000.
[145] Ladson CL. Effects of independent variation of Mach and Reynolds number on
the low-speed aerodynamic characteristics of NACA0012 airfoil section.
NASA technical paper 4074; January 1988.
[146] Landon RH. NACA 0012. Oscillating and transient pitching, in compendium of
unsteady aerodynamic measurements. AGARD-R-702, Data Set 3; August
1982.
[147] Beyers ME. Subsonic roll oscillation experiments on the standard dynamic
model. AIAA paper 83-2134; August 1983.
[148] Jermey C, Schiff LB. Wind tunnel investigation of the aerodynamic characteristics of the standard dynamic model in coning motion at Mach 0.6. In: 12th
AIAA atmospheric ight mechanics conference, AIAA Paper 1985-1828; 1985.
[149] Balakrishna S, Niranjana T. Wind tunnel dynamic ying study of the pitching
moment derivatives of the standard dynamics model in active control. AIAA
paper 87-2626; August 1987.
[150] Huang XZ. Wing and n buffet on the standard dynamic model. RTO
Technical report RTO-TR-26; October 2000.
[151] Cummings RM, Schtte A. Integrated computational/experimental approach
to unmanned combat air vehicle stability and control estimation. J Aircr
2012;49(6):154257.
[152] Leser T, Vicroy D, Schtte, A. SACCON static wind tunnel tests at DNW-NWB
and 140 220 NASA LaRC. AIAA paper 2010-4393; JuneJuly 2010.
[153] Spalart PR, Schur M. On the sensitisation of turbulence models to rotation
and curvature. Aerosp Sci Technol 1997;1:297302.
[154] Canter DE, Groves AW. X-31 post-stall envelope expansion and tactical utility
testing. AIAA paper 1994-2171; June 1994.
[155] Alcorn CW, Croom MA, Francis MS. The X-31 experience-aerodynamic
impediments to post-stall agility. AIAA paper 1995-362; January 1995.
[156] Williams DL, Nelson RC, Fisher D. An investigation of X-31 roll characteristics
at high angle-of-attack through subscale model testing. AIAA paper 1994806; January 1994.
[157] Rein M, Hhler G, Schtte A, Bergmann A, Leser T. Ground-based simulation
of complex maneuvers of a delta-wing aircraft. AIAA paper 2006-3149; June
2006.
[158] Schefter J. X-31: how they're inventing a radical new way to y. Pop. Sci.
1989;234(2):5864.
[159] Ghoreyshi M, Jirsek A, Cummings RM. Challenges in the aerodynamics
modeling of an oscillating and translating airfoil at large incidence angles.
Aerosp Sci Technol 2013;28(1):13344.
[160] Mackman TJ, Allen CB, Ghoreyshi M, Badcock KJ. Comparison of adaptive
sampling methods for generation of surrogate aerodynamic models. AIAA J
2013;51(4):797808.
[161] Da Ronch A, Vallespin D, Ghoreyshi M, Badcock KJ. Evaluation of dynamic
derivatives using computational uid dynamics. AIAA J 2012;50(2):47084.
[162] Betts JT. Survey of numerical methods for trajectory optimization. J Guid
Control Dyn 1997;20(1):5760.
[163] Hull DG. Conversion of optimal control problems into parameter optimization problems. J Guid Control Dyn 1998;21(2):193207.
[164] Love MC. Flight maneuvers. New York, USA: McGraw-Hill; 1999.
[165] Anderson F. Flying the mountains: a training manual for ying single-engine
aircraft. New York, USA: McGraw-Hill; 2003.
[166] Ross IM, Fahroo F. User's manual for DIDO 2002: a MATLAB application
package for dynamic optimization. NPS Technical report. Department of
Aeronautics and Astronautics, Naval Postgraduate School, AA-02-002; 2002.

Progress in Aerospace Sciences: Mehdi Ghoreyshi, Adam Jirásek, Russell M. Cummings

Uploaded by

Copyright:

Available Formats

Progress in Aerospace Sciences: Mehdi Ghoreyshi, Adam Jirásek, Russell M. Cummings

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Progress in Aerospace Sciences: Mehdi Ghoreyshi, Adam Jirásek, Russell M. Cummings

Uploaded by

Copyright:

Available Formats

Progress in Aerospace Sciences 71 (2014) 167217

Contents lists available at ScienceDirect

Progress in Aerospace Sciences

Reduced order unsteady aerodynamic modeling for stability

M. Ghoreyshi et al. / Progress in Aerospace Sciences 71 (2014) 167217

1.3. Current aerodynamic modeling efforts

M. Ghoreyshi et al. / Progress in Aerospace Sciences 71 (2014) 167217

unit step function

For xed wing aircraft, aerodynamic damping derivatives are

Reynolds number, Vc=

normalized time rate of angle of attack, 1/rad

The reduced-order model uses Kriging surrogates to account for

M. Ghoreyshi et al. / Progress in Aerospace Sciences 71 (2014) 167217

2. Review of unsteady aerodynamic prediction methods

where A and are the pitch amplitude and rotational velocity,

vertical displacements of the airfoil; this angle is named the

where _ e de =dt; CL is the lift force coefcient and t is Wagner

where s 2Vt=c is a non-dimensional time. Wagner's approximation

where a is the pitch axis location with respect to the half-chord as

Fig. 1. Airfoil pitch and plunge maneuvers.

M. Ghoreyshi et al. / Progress in Aerospace Sciences 71 (2014) 167217

Also, Kssner [85] dened a Wake Vortex Function, later named

where _ g dg =dt. The Kssner function can be related to Wagner

Fig. 2(a) compares Kssner and Wagner functions; it shows that

2.2. Volterra theory

where F is a continues functional of the input. Volterra used a p-th

where H0 is the steady-state term and is constant; Hi represents

and using the commutative law, it can also be written as

Korenberg and Hunter [82] mentioned some Volterra kernel

M. Ghoreyshi et al. / Progress in Aerospace Sciences 71 (2014) 167217

and multi-output system. For a forced oscillation about the pitch

where the term Hm

Note that the superscripts in Eq. (20) identify to which inputs

2.3. Indicial theory

where x_ dx=dt; A represents the unit response, or indicial

Duhamel's integrals with respect to and q, i.e.

M. Ghoreyshi et al. / Progress in Aerospace Sciences 71 (2014) 167217

to the Volterra series, the indicial theory is typically accurate for

with given initial condition u u0 for t t 0 and

where u is a r-dimensional vector of internal state variables over

where k is an integer value showing discrete time values. The state

In this system, the outputs can be determined from the states at

also, the output vector is dened as y C L;D;m t. In this system, the

are assumed to be cheaper to measure than primary data and

The mapping in Eq. (31) is then used to predict secondary output

and an output equation as

In Eq. (30), is a vector-valued nonlinear function that maps the

such a model brings the information of secondary data into the

and the corresponding aerodynamic loads are indicated by yi t.

where k are the weights of the linear combiners. The functions i

M. Ghoreyshi et al. / Progress in Aerospace Sciences 71 (2014) 167217

Gaussian basis function as

where is a real variable to be chosen by the user, and J  J

estimated kernels are very sensitive to the inaccuracies in dening

3. Review of system identication methods

System identication (SID) refers to the process of constructing

The identication of discrete-time Volterra kernels involves the

where the vector b contains the unknown Volterra kernels. The

where is a real variable to be chosen by the user, and J J