Transonic Aerodynamic Loads Modeling of X-31

Aircraft
Mehdi Ghoreyshi

, Martiqua L. Post

, and Russell M. Cummings

Modeling and Simulation Research Center, U.S. Air Force Academy

USAF Academy, Colorado 80840-6400
Andrea Da Ronch

and Kenneth J. Badcock

CFD Lab, University of Liverpool

Liverpool, UK, L69 3GH
The generation of reduced order models for computing the unsteady and nonlinear aero-
dynamic loads on the X-31 aircraft from pitching motions in the transonic speed range is
described. The models considered are based on Duhamels superposition integral using
indicial (step) response functions, Volterra theory using nonlinear kernels, Radial Basis
functions, and a surrogate-based recurrence framework, both using time-history simula-
tions of a training maneuver(s). One of the biggest challenges in creating these reduced
order modeling techniques is accurate identication of unknowns. A large number of step
function calculation is required for any combination of angle of attack and free-stream Mach
number. A method to eciently reduce the number of step function calculations is pro-
posed. 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 from a
limited number of simulations (samples). Each sample itself is directly calculated from un-
steady Reynolds-Averaged Navier-Stokes simulations starting from an initial steady-state
condition with a prescribed grid motion. An indirect method is proposed to estimate the
nonlinear Volterra kernels from time-accurate computational uid dynamic simulations of
dierent training maneuvers. These maneuvering simulations were also used to estimate
the unknown parameters in a model based on Radial Basis functions. A Design of Exper-
iment method was used to generate several pitching motions at dierent amplitudes and
free-stream Mach numbers. The model based on a surrogate-based recurrence framework
then approximates the aerodynamic responses induced by pitching motions at new ight
conditions. Results are reported for the X-31 conguration with a sharp leading-edge ge-
ometry, including canard/wing vortices. The validity of models studied was assessed by
comparison of the model outputs with time-accurate computational uid dynamic simula-
tions of new maneuvers. Overall, the reduced order models were found to produce accurate
results, although a nonlinear model based on indicial functions yielded the best accuracy
among all models. This model, along with a developed time-dependent surrogate approach,
helped to produce accurate predictions for a wide range of motions in the transonic speed
range within a few seconds.
I. Introduction
Modern ghter aircraft often maneuver in extreme conditions of the ight envelope characterized by
vortical ows and shock eects. Such conditions are important for unsteady ow which can cause ight

NRC Research Fellow, AIAA Senior Member

Associate Professor, AIAA Senior Member

Professor of Aeronautics, AIAA Associate Fellow

Research Associate, AIAA Member

Professor of Computational Aerodynamics, AIAA Senior Member

1 of 29
American Institute of Aeronautics and Astronautics
30th AIAA Applied Aerodynamics Conference
25 - 28 June 2012, New Orleans, Louisiana
AIAA 2012-3127
This material is declared a work of the U.S. Government and is not subject to copyright protection in the United States.
dynamic and aeroelastic instabilities.
1
The linearized unsteady aerodynamic models no longer work at these
conditions. Currently, the use of computational uid dynamics (CFD) solutions is considered as the state
of the art in modeling unsteady nonlinear ow physics
2
and oers an early and improved understanding
and prediction of aircraft aerodynamic characteristics. It is well known that unsteady aerodynamic forces
and moments not only depend on the instantaneous states but also by all of the states at times prior to
the current state.
3
Unfortunately, this 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.
4
With the advances in computing techniques, one straightforward way to solve this problem
is to develop a full-order model based on direct solution of discretized Reynolds-Averaged Navier-Stokes
(RANS) equations coupled with the dynamic equations governing the aircraft motion.
57
Creating a full-order model for Stability & Control (S&C) analysis is a computationally very expensive
approach since such a model needs a large number of coupled computations for dierent values of motion
frequency and amplitude.
8
An alternative approach to creating the full-order model is to develop a Reduced
Order Model (ROM) that seeks to approximate CFD results by extracting information from a limited number
of full-order simulations.
9, 10
Ideally, the specied ROM can predict aircraft responses over a wide range of
amplitude and frequency within a few seconds without the need of running CFD simulations again.
Recent eorts on the development of ROMs can be classied into two types: time domain or frequency
domain approaches.
11
The frequency domain models are obtained from matching transfer functions com-
puted from the measured input-output data.
12
Examples of the frequency domain ROMs are the indicial
response method by Ballhaus and Goorjian
13
and Tobak et al.,
14, 15
and a frequency-domain approach based
on proper orthogonal decomposition by Hall.
16
Some examples of time domain ROMs include the unit
sample response by Gaitonde and Jones,
17
Volterra theory by Silva and Bartels,
18
Radial Basis Functions
(RBF)
19
and state-space modeling.
20
These ROM techniques have been used extensively for utter predic-
tion, limit cycle oscillation, and gust response modeling,
21
but their application to S&C is still new. Also,
only a few studies have been conducted for reduced order modeling of aircraft congurations, mostly limited
to the subsonic ow regime. The current paper aims to assess the accuracy of prediction and incurred com-
putational cost of reduced order models based on the indicial response method, Volterra theory, RBF, and
a Surrogate-Based Recurrence-Framework (SBRF) for X-31 aircraft pitching motions in the transonic speed
range. More details about these models are provided in section II.
The transient aerodynamic response due to a step change in a forcing parameter, such as angle of
attack or pitch rate is a so-called indicial function. Assuming that the indicial functions are known,
the aerodynamic forces and moments induced in any maneuver can be estimated by using the well-known
Duhamel superposition integral.
22
Tobak et al.
14, 15
and Reisenthel et al.
23, 24
detailed the superposition
process for the modeling of unsteady lift and pitch moment from angle of attack and pitch rate indicial
functions. Ghoreyshi and Cummings
25
extended this model to include lateral forces and moments as well.
The accuracy of ROMs based on indicial functions depends, to a large extent, on accurate identication of
indicial functions; these functions can be derived from analytical, CFD, or experimental methods.
26
Limited
analytical expressions of indicial functions exist for two-dimensional airfoils.
27
For incompressible ows,
Wagner
28
was the rst who detailed the analytical unsteady lift of a thin airfoil undergoing a plunging
motion using a single indicial function (the so-called Wagners function) with its exact values dened in
terms of Bessel functions. For unsteady, compressible ow past two-dimensional airfoils, Bisplingho et
al.
29
also described an exponential approximation to the exact solutions of the indicial functions at dierent
Mach numbers. However, these analytical expressions are not valid for aircraft congurations due to the
three-dimensional tip vortices.
Direct and indirect methods are used to estimate indicial functions. Leishman
30
has presented an indirect
technique for identifying indicial functions from aerodynamic responses due to harmonic motions. However,
the derived indicial functions using indirect methods depend largely on the quality of motion, e.g. amplitude
and frequency. Experimental tests are practically nonexistent for direct indicial function measurements due
to wind tunnel constraints. An alternative is to use CFD, but special considerations are required to simulate
step responses in CFD. Singh and Baeder
31
used a surface transpiration approach to directly calculate the
angle of attack indicial response using CFD. Ghoreyshi et al.
32
also described an approach based on a grid
motion technique for CFD-type calculation of linear and nonlinear inidical functions with respect to angle
of attack and pitch rate. Ghoreyshi and Cummings
25
later used this approach to generate longitudinal and
lateral indicial functions for a generic unmanned combat air vehicle and used these functions for predicting
the aerodynamic responses to aircraft six degrees of freedom maneuvers. In this paper, the transonic indicial
2 of 29
American Institute of Aeronautics and Astronautics
functions of the X-31 aircraft are calculated using CFD and the grid motion approach.
For motions at low angles of attack, and assuming incompressible ow, only a single indicial function
with respect to each forcing parameter needs to be generated.
33
For ows at transonic speed, however many
indicial functions need to be generated for each combination of angle of attack and free-stream Mach number.
The generation of all these functions using CFD is expensive and makes the creation of a ROM very time
consuming. Note 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. A cost-eective unsteady
aerodynamic model needs a mathematical description of indicial functions as a function of angle of attack
and free-stream Mach number. However, this model is often unavailable for three-dimensional congurations.
It is more common to use surrogate models which are mathematical approximations of the true response of
the system, built using some observed responses, and therefore the total cost of creating ROM is reduced.
In this paper, a surrogate model is proposed based on the Kriging technique
34
to model indicial functions as
a function of angle of attack and free-stream Mach number.
Volterras functional mathematics is considered as one of the most important tools for the representation
of non-linear systems where the system output depends on the current and past values of the input. The
approach was rst introduced by Vito Volterra in a book published in 1930.
35
This work and further
developments by others (for example N. Wiener,
36
Barrett
37
and Kalman
38
) have been extensively used in
electrical and biological systems engineering.
3943
Recently, there is an increasing interest in using Volterra
theory in the eld of aerodynamic loads modeling,
2, 44
where the aerodynamic forces and moments are
approximated by an innite series of multi-dimensional convolution integrals of the inputs and increasing
order kernels named the so-called Volterra kernels.
45
Generally the Volterra kernels are not known and
need to be determined. Silva
46
used the rst and second terms of the Volterra functions and proposed a
method based on the impulse functions to directly calculate rst- and second-order kernels using CFD. In
his approach the rst-order 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. Jirasek and Cummings
44
used
this approach to nd Volterra kernels for X-31 aircraft at subsonic speeds. However, the CFD simulation
of system impulse response at transonic speeds is very complicated as the impulse occurs over a very short
period of time. Da Ronch et al.
47
proposed an approach of determining these functions from unsteady time-
domain solutions. In this paper, the chirp and spiral training maneuvers are used to estimate the Volterra
kernels.
The complex CFD system (that is ow, ow equations, and boundary conditions) can be represented by
a simplied input-output relationship. The relationship can be learned using neural networks and surrogate
models.
45, 4850
Marques and Anderson
45
used a temporal neural network to approximate the unsteady lift
and pitch moment of a 2D airfoil for the changes in angle of attack in the transonic regime. Although, they
found a good match for the lift, the predictions of pitch moment did not match as well. The network outputs
in their work depended on time histories of angle of attack only and not previous values of lift and pitch
moment. Faller and Schreck
49
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 coecients at t
0
. 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 interval. This process continues with time until it reaches the nal time. However, this model does not
have the time histories of motion variables, therefore it might not have sucient information for modeling
highly unsteady ows. Glaz et al.
50
developed a mapping between aerodynamic loads and time histories of
both motion parameters and loads and then tried to learn this mapping using a surrogate model with the
aid of Design of Experiments. Ghoreyshi et al.
51
extended this mapping to include both RANS and Euler
calculations and used RBF neural networks trained from some special training maneuvers. Da Ronch et
al.
47
also tried to establish this mapping using a SBRF model for a pitching airfoil at transonic speed. A
key issue in this modeling is the eects of the training motion on identifying model parameters. This paper
discusses the eects of three training motions: chirp, spiral, and Schroeder.
19
All these reduced order models require dierent computational costs and are based on various simplifying
assumptions pertaining to the ow physics. For example, the generation of linear indicial functions are
relatively inexpensive, but the model cannot accurately capture nonlinear ow phenomena associated with
transonic ow and/or dynamic stall. The accuracy of models based on RBF and SBRF depends to a large
extend on the training maneuvers used. Therefore in this paper, these models are tested for predicting the
3 of 29
American Institute of Aeronautics and Astronautics
unsteady loads induced in pitching motion at transonic speeds. The conguration, the X-31 aircraft, has
highly swept slender wings as shown in Fig. 1, resulting in complex vortical ow under various conditions.
The unsteady aerodynamic modeling of this conguration at transonic regime is very complicated due to
strongly nonlinear aerodynamic phenomena associated with shock eects and vortical ow.
The test case, geometry denition, and ROMs will rst be detailed. The paper continues to describe the
generation of ROMs. Finally, the accuracy of the aerodynamic models will be compared with time-dependent
CFD computations of new maneuvers.
II. Formulation
A. CFD Solver
The ow solver used for this study is the Cobalt code
52
that solves the unsteady, three-dimensional and
compressible Navier-Stokes equations in an inertial reference frame. These equations in integral form are
53

t
___
QdV +
__
(f

i +g

j +h

k). ndS =
__
(r

i +s

j +t

k). ndS (1)

where V is the uid element volume; S is the uid element surface area; n is the unit normal to S;

i,

j,
and

k are the Cartesian unit vectors; Q = (, u, v, w, e)
T
is the vector of conserved variables, where
represents air density, u, v, w are velocity components and e is the specic energy per unit volume. The
vectors of f, g, and h represent the inviscid components and are detailed below
f = (u, u
2
+p, uv, uw, u(e +p))
T
g = (v, v
2
+p, vu, vw, v(e +p))
T
(2)
h = (w, w
2
+p, wv, wv, w(e +p))
T
where the superscript T denotes the transpose operation. The vectors of r, s, and t represent the viscous
components are described as
r = (0,
xx
,
xy
,
xz
, u
xx
+v
xy
+w
xz
+kT
x
)
T
s = (0,
xy
,
yy
,
yz
, u
xy
+v
yy
+w
yz
+kT
y
)
T
(3)
t = (0,
xz
,
zy
,
zz
, u
xz
+v
zy
+w
zz
+kT
z
)
T
where
ij
are the viscous stress tensor components; T is the temperature; and k is the thermal conductivity.
The ideal gas law closes the system of equations and the entire equation set is nondimensionalized by free-
stream density and speed of sound.
52
The Navier-Stokes equations are discretised 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,
54
and least squares gradient calculations using QR factorization. To
accelerate the solution of discretized system, a point-implicit method using analytic rst-order inviscid and
viscous Jacobians. A Newtonian sub-iteration method is used to improve time accuracy of the point-implicit
method. Tomaro et al.
55
converted the code from explicit to implicit, enabling Courant-Friedrichs-Lewy
numbers as high as 10
6
. The Cobalt solver has been used at the Air Force Seek Eagle Oce (AFSEO) and
the United States Air Force Academy (USAFA) for a variety of unsteady nonlinear aerodynamic problems
of maneuvering aircraft.
5660
B. ROMs Based on Volterra Theory
The X-31 aircraft pitch moment modeling is considered in this paper, which is highly nonlinear in the
transonic regime. The output function y is dened as y = C
m
(t). Using Volterra theory,
35
the output
of a continuous-time, casual, time-invariant, fading memory system in response to an input, x(t), can be
modeled using the p-th order Volterra series:
y (t) = (x(t)) =
p

i =1
H
i
(x(t)) (4)
4 of 29
American Institute of Aeronautics and Astronautics
where, H
i
represents the i-th order Volterra operator, which is dened as an i-fold convolution between the
input, x(t), and the i-th order Volterra kernel, H
i
, i.e.
H
i
(x(t)) =
_
t

. . .
_
t

H
i
(t
1
, t
2
, . . . , t
i
)
i

n=1
x(
n
) d
n
(5)
where, for forced oscillations about the pitch axis, the input vector includes:
x(t) =
_
(t) , (t) , (t)
_
(6)
A multi-input Volterra series is then formulated as
y (t) =
_
x
1
(t) , x
2
(t) , . . . , x
m
(t)
_
=
p

i =1
H
m
i
(7)
The term H
m
i
is the multi-input Volterra operator dened as a m
p
-fold summation of p-fold convolution
integrals between the inputs and the p-th order multi-input Volterra kernels.
61
The output response up to
second order is rewritten as
y (t) =
m

j =1
_
t

H
xj
1
(t ) x
j
() d +
m

j1 =1
m

j2 =1
_
t

_
t

H
xj
1
, xj
2
2
(t
1
, t
2
) x
j1
(
1
) x
j2
(
2
) d
1
d
2
+ O
_
|x|
3
_
(8)
Note that the superscripts in Eq. (8) identify to which inputs the kernel corresponds, for example, the
second-order kernel H
xj
1
, xj
3
2
correlates the inputs x
j1
and x
j3
. Note that the second and higher-order kernels
are symmetric with respect to the arguments, H
xj
1
, xj
3
2
= H
xj
3
, xj
1
2
. The determination of Volterra kernels
using CFD simulations is discussed in the section on System Identication.
C. ROMs Based on Indicial Functions
Linear and nonlinear models based on indicial functions are detailed here. Assuming a linear relationship
between the input function and the output, a linear ROM is dened as the convolution (or Duhamels
superposition
30
) of the indicial response with the derivative of the forcing function
62
y(t) =
d
dt
[
_
t
0
A(t )x()d] (9)
where A represents the unit response, or indicial response, of the system. Note that a linear Volterra system
and the linear step-type ROM given in Eq. (9) are identical. For a linear system, H
1
is the impulse response
function and H
1
(t) = dA(t)/dt applies. The indicial response functions are used as a fundamental approach
to represent the unsteady aerodynamic loads. The mathematical models are detailed by Tobak et al.
14, 15
and Reisenthel et al.
23, 24
The pitch moment is only considered in this paper as it is the more dicult
parameter to predict. If the time responses in aerodynamic coecients due to the step changes in angle
of attack, , and angular velocity, q, are known, then the total produced pitch moment at time t can be
obtained using Eq. (10):
C
m
(t) = C
m0
+
d
dt
[
_
t
0
C
m
(t )()d] +
d
dt
[
_
t
0
C
mq
(t )q()d] (10)
where C
m0
denotes the zero angle of attack pitch moment coecient and is found from static calculations. For
non-linear aerodynamic responses due to motions starting from dierent Mach numbers, the dependencies
on the angle of attack and Mach number are added to the indicial functions, i.e. :
C
m
(t) = C
m0
(M) +
d
dt
[
_
t
0
C
m
(t , , M)()d] +
d
dt
[
_
t
0
C
mq
(t , , M)q()d] (11)
5 of 29
American Institute of Aeronautics and Astronautics
where M denotes the free-stream Mach number. The response function due to pitch rate, i.e. C
mq
(, M) can
be estimated using a time-dependent interpolation scheme from the observed responses. This value is next
used to estimate the second integral in Eq. 11, however, the estimation of the integral with respect to the angle
of attack needs more explanation. Assuming a set of angle of attack samples of = [
1
,
2
, ...,
n
] at free-
stream Mach numbers of M = [M
1
, M
2
, ..., M
m
], the pitch moment response to each angle of
i
, i = 1, 2, ..., n
at Mach number of M
j
, j = 1, 2, ..., m is denoted as A

(t,
i
, M
j
). In these response simulations, (t) = 0
at t = 0 and is held constant at
i
for all t > 0. For a new angle of

> 0 at a new free-stream Mach

number of M

, the responses of A

(t,
k
, M

) are being interpolated at

k
= [
1
,
2
, ...,
s
], such that
0 <
1
<
2
< ... <
s
and
s
=

. These angles can have a uniform or non-uniform spacing. The indicial

functions of C
m
k
for k = 1, ..., s at each interval of [
k1
,
k
] are dened as
C
m1
=
A

(t,
1
, M

) C
m0

1
(12)
C
m
k
=
A

(t,
k
, M

) A

(t,
k1
, M

k

k1
(13)
where C
m0
denotes the zero angle of attack pitch moment coecient. The interval indicial functions are then
used to estimate the values of rst integral in Eq. 11. These steps can easily be followed for a negative angle
of attack, i.e.

< 0. Once this model is created, the aerodynamic response to a wide range of motions can
be predicted on the order of few seconds, however, the above model still requires a special time-dependent
surrogate model to predict response functions at each

and M

from some available samples.

D. Surrogate-Based Modeling of Indicial Functions
Having a ROM to predict the aerodynamic responses to any arbitrary motion over a wide ight regime
could become a very expensive approach because a large number of indicial functions needs to be computed.
In order to achieve a reasonable computational cost, a special time-dependent surrogate-based modeling
approach is adapted to predict indicial responses for a new point from available (observed) responses. These
observed responses are viewed as a set of time-correlated spatial processes where the output is considered a
time-dependent function. Romero et al.
63
developed a framework for multi-stage Bayesian surrogate models
for the design of time dependent systems and tested their model for free vibrations of a mass-spring-damper
system assuming the input parameters of stiness and damping factor at dierent initial conditions. This
framework is examined for reduced order modeling of nonlinear and unsteady aerodynamic loads. Assume
an input vector of x(t) =
_
x
1
(t) , x
2
(t) , ..., x
n
(t)
_
where n represents the dimensionality of the input
vector. To construct a surrogate model for tting the input-output 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 D(mn) is then dened as:
D =
_

_
x
11
x
12
x
1n
x
21
x
22
x
2n
.
.
.
.
.
.
.
.
.
.
.
.
x
m1
x
m2
x
mn
_

_
(14)
where rows correspond to dierent 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 Z(mp) as:
Z =
_

_
y
11
y
12
y
1p
y
21
y
22
y
2p
.
.
.
.
.
.
.
.
.
.
.
.
y
m1
y
m2
y
mp
_

_
(15)
6 of 29
American Institute of Aeronautics and Astronautics
where for aerodynamic loads modeling, p equals the number of iterations used in time-marching CFD calcu-
lations. The objective of surrogate modeling is to develop a model that allows predicting the aerodynamic
response of y(x
0
) =
_
y
01
, y
02
, ..., y
0p
_
at a new combination of input parameter of x
0
. 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 Z
i
(D) for i = 1, 2, ..., p. A universal-type Kriging function
34
is used to approximate
these models. Each Z
i
(D) function can be approximated as the sum of a deterministic mean (trend),
and a zero-mean spatial random process of with a given covariance structure of
2
, therefore each function
value at the new sample of x
0
is

Z
i
(x
0
) = + (16)
where, the tilde accent shows that surrogate model is an approximation of the actual function. Universal
Kriging, which is used in this paper, assumes that the mean value is a linear combination of known
regression functions of f
0
(x), f
1
(x), .., f
n
(x). In this paper, the linear functions are used, therefore, f
0
(x) = 1
and f
j
(x) = x
j
for j = 1, 2, ..., n. This changes Eq. (16) as:

Z
i
(x
0
) =
n

j =0

ij
f
j
(x
0
) + (17)
where
ij
represent the regression coecient for the j-th regression function of response function at time
step i, i = 1, 2, ..., p. To estimate the spatial random process of , a spatially weighted distance formula is
dened between samples given in matrix D such that for sample x
i
and x
j
, the distance is written as:
d (x
i
, x
j
) =
n

h=1

x
(i)
h
x
(j)
h

p
h
(
h
0 and p
h
[0, 1]) (18)
where |.| shows the Euclidean distance; the parameter
h
expresses the importance of the h-th component
of the input vector, and the exponent p
h
is related to the smoothness of the function in coordinate direction
h. A correlation matrix R(mm) with a Gaussian spatial random process is then dened as:
R =
_

_
exp
_

d(x1,x1)

2
_
exp
_

d(x1,x2)

2
_
exp
_

d(x1,xm)

2
_
.
.
.
.
.
.
.
.
.
.
.
.
exp
_

d(xm,x1)

2
_
exp
_

d(xm,x2)

2
_
exp
_

d(xm,xm)

2
_
_

_
(19)
To compute the Kriging model, values must be estimated for
ij
, ,
h
, and p
h
. These parameters can
be quantied using the maximum likelihood estimator, as described by Jones et al.
64
Next the vector of
R(m1) is dened from correlations between the new design parameter x
0
and the m sample points, based
on the distance formula in Eq. (18), i.e.
r =
_

_
exp
_

d(x1,x0)

2
_
exp
_

d(x2,x0)

2
_
.
.
.
exp
_

d(xm,x0)

2
_
_

_
(20)
and now

Z
i
(x
0
) can be estimated as

Z
i
(x
0
) =
n

j =0

ij
f
j
(x
0
) +r
T
R
1
(Z
i
(D) F) (21)
7 of 29
American Institute of Aeronautics and Astronautics
where, is the n + 1 dimensional vector of regression coecients; Z
i
(D) is the observed responses at time
step i, i = 1, 2, ..., p and matrix F is
F =
_

_
f
0
(x
1
) f
1
(x
1
) f
n
(x
1
)
.
.
.
.
.
.
.
.
.
.
.
.
f
0
(x
m
) f
1
(x
m
) f
n
(x
m
)
_

_
(22)
The total response at x
0
is then combination of predicted values of each surrogate model, i.e.

Z(x
0
) =
_

Z
1
(x
0
) ,

Z
2
(x
0
) , ...,

Z
p
(x
0
)
_
(23)
E. ROMs Based on Surrogate-Based Recurrence-Framework
The set of non-linear equations describing the CFD system can be interpreted as a general representation of a
non-linear, time-invariant, and discrete-time dynamical system. The state vector consists of the conservative
variables, and its size is proportional to the number of grid points. In this study, the aerodynamic loads
form the vector of outputs, which are not only a function of the instantaneous values of the inputs, but also
a function of the time history of the inputs.
To generate a computationally ecient approximation of the unsteady aerodynamic loads without solving
the expensive CFD equations, the form of a dynamical system is assumed.
50
When the state vector of the
full-order system is nite in dimension, the following non-linear system is equivalent to the unsteady CFD
equations
y (t) =
_
x(t) , x(t t) , . . . , x(t mt) , y (t t) , . . . , y (t nt)
_
(24)
where x takes the form of Eq. (6), and the function maps the inputs to the outputs. The terms m and
n represent the number of previous values of the external inputs and outputs, respectively, inuencing the
output at the current time instant. These parameters account for time-history eects and phase-lag in the
ow development.
Central to the generation of the reduced-order model is the computation of the function . Without a
closed-form analytical expression, a numerical approximation of is constructed using a number of CFD
solutions. For the pitching aircraft, any motion can be expressed as function of three parameters, e.g.,

0
,
A
, and k that represent 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. A description of the Kriging-based framework used in this study is detailed
in Ref.
65
Let N
T
be the number of training cases at which CFD solutions are available. Each training case
consists of dierent combinations of the independent parameters,
x
i
=
_

i
(t) ,
i
(t) ,
i
(t) , . . .
_
for i = 1, . . . , N
T
(25)
and the corresponding aerodynamic loads are indicated by y
i
(t). The approximation of the function is
obtained by interpolating the sampled data in the form of an input/output relationship. Several interpolation
methods are available in the literature, and two of these have been used in the present study. Kriging
interpolation is a common choice, but for increasing number of independent parameters the problem can
result to be ill-conditioned. An alternative approach is the multi-linear interpolation technique, which is in
general faster than the Kriging interpolation.
F. ROMs based on RBFNN
The nonlinear and unsteady aerodynamics can be viewed as a multi-input/multi-output dynamic system
with a reconstructed state space model given by Eq. (24). The function is approximated through RBF
8 of 29
American Institute of Aeronautics and Astronautics
neural networks from a set of training data. RBF network provides an approximation of the functions
based on the location of data points, and is generally much faster than multi-layer feed-forward neural
networks.
66
Given an input vector of {X
c
j
: j = 1, ...., p}, X
c
j
R and a corresponding output vector of
{Y
c
j
: j = 1, ...., p}, Y
c
j
R, the RBF approximates the output at a new given point as:

Y (X

) =
P

k=1

i
(X) (26)
such that

Y (X
c
j
) = Y
c
j
, for j = 1, 2, ..., p (27)
where
k
are the weights of the linear combiners. The functions
i
are named Radial Basis Functions and
are often described by a Gaussian basis function as:

i
(X) = exp(
X

X
c
j

2

2
) (28)
where, is a real variable to be chosen by the user, and . denotes the Euclidean norm such that the
functions
i
will vanish at suciently large values of X

X
c
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.
G. System Identication
The identication of the Volterra kernels is performed using an unsteady time-domain simulation as the
source of the data. The CFD solution is discrete in time, and the time-step is indicated by t

. Denote the
input vector at each time step as x[n] = x(nt

) = x(t). The discrete-time representation of Eq. (8) is

y [n] =
m

j =1

k =0
H
xj
1
[n k] x
j
[k] +
m

j1 =1
m

j2 =1

k1 =0
n

k2 =0
H
xj
1
, xj
2
2
[n k
1
, n k
2
] x
j1
[k
1
] x
j2
[k
2
] + O
_
|x|
3
_
(29)
The computational determination of system impulse response may be complicated as it occurs over a
very short period of time,
67
thus, it is more common to use indirect methods. The indirect estimation of
Volterra kernels involves the resolution of an overdetermined system. Values of aerodynamic coecients
and the time-history of the motion variables are known from the CFD simulation used as training input.
Let y = (y[0], y[1], . . . , y[n])
T
denote each aerodynamic load computed using CFD, and let A contain the
permutations of input parameters relevant to the unsteady motion. Equation (29) can be recast in the form
y = Ab (30)
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 A
T
A, Gauss elimination, Moore-Penrose generalized inverse approach and the QR factorization.
However, the Moore-Penrose approach and the QR factorization are more accurate than the Gaussian elimi-
nation and the direct inversion solutions. The cost of the QR factorization is O
_
n
2
_
, and the Moore-Penrose
inversion involves O
_
n
3
_
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 manoeuvre of sucient duration. A remedy to this is the use of a simplied form
of the kernel parametric structure. For example, Maciej et al
68
proposed to set all o-diagonal terms of the
kernel to zero, i.e.
9 of 29
American Institute of Aeronautics and Astronautics
H
xj
1
, xj
2
, ..., xjp
p
[n k
1
, n k
2
, . . . , n k
p
] = 0 (31)
for k
1
= k
2
= . . . = k
p
In this work, the Volterra kernels are identied from Eq. (30) 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. (30) is used to predict the resulting unsteady aerodynamic loads in place of the full-order system.
An approach based on grid motion is used to calculate the indicial response directly from CFD. This
approach allows the uncoupling of eects of angle of attack and pitch rate for the indicial functions. Cobalt
uses an arbitrary Lagrangian-Eulerian formulation and hence allows all translational and rotational degrees
of freedom.
32
The code can simulate both free and specied six degree of freedom (6DoF) motions. The
rigid motion is specied from a motion input le. For the rigid motion 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 also dened using the rotation angles of yaw, pitch and bank. The aircraft reference point velocity, v
a
, 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 location. The initial aircraft velocity, v
0
, is specied in terms of
Mach number, angle of attack and side-slip angle in the main le. The instantaneous aircraft location for
the motion le is then dened from the relative velocity vector, v
a
v
0
. For CFD-type calculation of a step
change in angle of attack, the grid immediately starts to move at t = 0 to the right and downward. The
translation continues over time with a constant velocity vector. Since there is no rotation, all the eects 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.
A training maneuver(s) is needed to provide enough information to learn the mapping between input
and output of ROMs given by Eq. (24). Previous studies to generate training maneuvers for aerodynamic
characteristics
6973
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 suciently 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, special training maneuvers are considered: a linear chirp, spiral, and Schroeder. These
maneuvers are dened in Table 1:
Table 1. Special Training Maneuvers
Maneuver Denition
Linear chirp (t) =
0
+
A
sin(t
2
)
Spiral (t) =
0
+
A
tsin(t)
Schroeder (t) =
0
+
A

n
k =1
_
1
2N
cos(
2kt
T

k
2
N
)
where
0
is the mean angle of attack;
A
is amplitude and is angular velocity. The chip maneuver used
has a constant amplitude and linearly increasing frequency in time. In the spiral maneuver the amplitude
increases linearly in time, so 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. These are maneuver amplitude,
A
, the maneuver length, T,
and the number of frequencies in the maneuver, N.
III. Test Case
The X-31 aircraft is considered in this paper. The aircraft geometry and wind tunnel data were provided
to the participants in NATO RTO Task Group AVT-161 (Assessment of Stability and Control prediction
Methods for NATO Air and Sea Vehicles).
74
The objective of this task group is to evaluate CFD codes
10 of 29
American Institute of Aeronautics and Astronautics
against the wind tunnel results. 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 Canter and Groves,
75
Alcorn et al.,
76
Williams et al.
77
and Rein
78
) and CFD
simulations (example is the work of Sch utte et al.
5
). A three-view drawing of the vehicle is shown in Fig. 1.
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 57
o
and the outer sweep is 45
o
. The inner wing sweep places
the wing behind the supersonic shock wave, while the outer one improves the vehicle stability and control.
79
The canard is a cropped delta wing with a sweep angle of 45
o
. Additional characteristics of the model are
the inner and outer leading edge aps, the trailing edge aps, the front wing and rear fuselage strakes.
The computational mesh was 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
80, 81
which builds prism layers using a frontal technique. TRITET rebuilds the inviscid mesh while respecting
the size of the original inviscid mesh from ICEMCFD. The mesh overview is shown in Fig. 2. 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 far-eld, symmetry, and solid wall. The low-speed experiments are available
from the DLR, German Aerospace Center.
78
The wind tunnel model has a closed inlet and tted with moving
lift and control surfaces. The experiments are composed of two 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 of second setup are used to validate CFD predictions. CFD simulations were run on
the Cray XE6 (open system) machine at the Engineering Research Development Center (ERDC) [Machine
name is Chugach with 2.3GHz core speed and 11,000 cores]. Four turbulence models were tested: the
Spalart-Allmars (SA),
82
the SA with Rotation Correction (SARC),
83
the Menters Shear Stress Transport
(SST)
84
and SARC with Detached Eddy Simulations (SARC-DES). Figure 3 compares the lift, drag, and
pitch moment coecients obtained from each turbulence model with the available measurements. All these
models yield similar predictions at low angles of attack, but they result in a wide spread of predictions at
moderate to high angles. For angles between, = 15
o
-23
o
, SARC-DDES and SST models perform quite well
compared to SA and SARC. DES and SST accurately predict unsteady separated ows occurring at these
angles, but for angles above 23
o
, all models fail to predict accurately the massively separated ows.
Some features of aerodynamic characteristics from the SARC-DES 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 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. 4(a) for 10 degrees angle of
attack. Around an angle of 14 degrees, the canard vortex starts to breakdown and the wing vortex is formed
as shown in Fig. 4(b). The wing vortex helps to further forward movement of aerodynamic center and
increase of pitch moment. At 18 degrees 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. 4(c). This results in an
aftwards movement of aerodynamic center and a change in the pitch moment slope sign. The wing vortex
breakdown point moves towards the leading edge by increasing angle of attack (Fig. 4(d)- (e)). The canard
vortex is fully separated at these angles. As the vortex breakdown point becomes close to the wing leading
edge (Fig. 4(f)), the pitch moment starts to rise again. The SARC-DES turbulence model was used for all
ROM computations.
IV. Results and Discussion
All ROMs are rst evaluated to predict the unsteady pitch moment resulting from a sinusoidal pitch
oscillation at free-stream Mach number of 0.9 and reduced frequency of k = 0.01. The amplitude of oscillation
is held constant at seven degrees and the mean angle of attack is zero degrees. Figure 5(a) shows the computed
pitch moment coecient, C
m
, by solving the RANS and SARC-DES turbulence model equations in a time-
accurate fashion. This solution and other time-accurate CFD simulations are labeled as Time-Marching
in the ROM plots. The cost of simulating three pitch cycles is approximatly 52 wall-clock hours using 256
processors (2.3 GHz). Figure 5(a) shows that the pitch moment curve makes a nonlinear loop on the gure
of moment versus angle of attack due to occurrence of shock waves and vortices. This gure also shows that
11 of 29
American Institute of Aeronautics and Astronautics
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 during the pitch oscillation
are shown in Figs. 5(b)-(g) and briey discussed here. In Fig. 5(b) the angle of attack is -2.1
o
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 Figs. 5(b)-(d). Figure 5(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. = -7
o
, the leading edge vortex becomes much stronger
and the wing surface pressure close to the leading edge drops further as shown in Fig. 5(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.
=7
o
, as shown in Fig. 5(f). Figures 5(e)-(g) show 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.
For the identication of the Volterra kernels, the chirp and spiral training maneuvers were generated using
CFD as the source of the data. The variation of angle of attack with time for these maneuvers is shown
in Figs. 6(a) and (c). Both maneuvers ran for 2.4 seconds of physical time and started from a steady-state
solution. The chirp maneuver has an amplitude of seven degrees starting from zero degrees angle of attack
and pitching with a frequency of 1Hz at t = 0. Note that the chirp motion frequency increases linearly with
time. The spiral maneuver has an initial amplitude of 3.5
o
, 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 approximatly 84 wall-clock hours using 256 processors. 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 unsteady pitch moment data
shown in Fig. 5(a). The ROM predictions based on spiral and chirp training maneuvers are compared with
the time-accurate solution data in Figs. 6(b) and (d). The comparisons show a good agreement with CFD
data for a ROM identied from spiral data, but the ROM identied from chirp data does not match well, 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 sucient information to identify the Volterra kernels
corresponding to a swept amplitude motion at constant frequency. However, the ROM based on chirp data
could be used for predicting aerodynamics responses from pitch oscillations at many other frequencies (those
covered by the simulation of chirp training maneuver), but the ROM based on spiral is possibly 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 coecient 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 4.95
o
. The number of frequencies in the maneuver, N,
was set to 20 with an initial amplitude of seven degrees. This maneuver ran for 2.4 seconds of physical time
as well and is shown in Fig. 7(e). The aircraft responses to these three maneuvers were generated using
URANS equations. The training data were next normalized using the mean and standard deviation of each
input. The data are then rearranged according to Eq. (24) and the RBF network performance is tested
for dierent 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 sucient for modeling the
studied motions. The trained networks were then tested against the target motion; the ROM predictions
are shown in Figs. 7(b),(d) and (f). These gures show that the predicted ROM values agree well with
the time-marching solution, although the ROM based on Schroeder maneuver showed better accuracy than
models based on the chirp and spiral maneuvers.
The indical pitch moment responses of the X-31 aircraft with a unit step change in angle of attack and
pitch rate are shown in Fig. 8(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 C
Lq
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
12 of 29
American Institute of Aeronautics and Astronautics
computations started from a steady-state solution and then advanced in time using second-order accuracy
with ve Newton subiterations. As shown in Fig. 8(a), the pitch moment responses have a negative peak
at t = 0 followed by an increasing trend. 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 surface of the vehicle
that cause a sharp negative pitch moment peak in the responses.
32
As the response time progresses, the
waves begin to move away from the vehicle and the pitch moment responses start to increase and then
asymptotically reach the steady-state values. The cost of generating each indicial function is around 1.5
wall-clock hours using 256 processors. A linear ROM was created using Eq. (10) and used for prediction of
target maneuver. The results are compared with full-order model in Fig. 8(b). The gure shows that linear
ROM fails to accurately predict the pitch moment values at all angle of attack. The functions of C
L
vary
largely with angle of attack at transonic speed range and thus a linear ROM cannot predict these eects.
Next, the X-31 C
L
functions were simulated at dierent angles of attack and at free-stream Mach number
of 0.9 and shown in Fig. 8(c). Note that the pitch moment slope is not symmetric with zero 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 approximatly 21 wall-clock hours. Fig. 8(c) shows that
the responses at initial time are invariant with angle of attack, but the intermediate and nal values change
depending on the angle of attack. Fig. 8(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 target maneuver was evaluated.
Figure 8(d) shows that the nonlinear ROM predictions agree very well with full-order simulation values.
Note that such a nonlinear ROM could be used for computing the pitch moment responses from many other
motions with dierent amplitudes and frequencies. For example, the ROM was used to predict two pitch
oscillations with four and six degrees amplitude at M = 0.9 and k = 0.01. The predictions are compared
with time-marching solutions in Figs. 9(a) and (b). Again a very good match was found. Also, the ROM
predictions were evaluated for the chirp, spiral and Schroeder maneuvers used in RBF work. Figs. 9(c)-(e)
show that even for this varying amplitude and frequency motions, the created ROM matches very well with
CFD data.
For the generation of the SBRF model, time-accurate simulations were pre-computed for various combi-
nations of pitch amplitude and Mach number at a xed reduced frequency. The two-dimensional parameter
space was lled using Design of Experiement methods and is shown in Fig. 10(a). Also, the pitch motion
simulations of all samples are shown in Fig. 10(b). The SBRF model was used for the prediction of the pitch
moment coecient time history for sinusoidal forced motions about zero degrees angle of attack, amplitude
of seven degrees and values of Mach number of 0.78, 0.825, and 0.88. Model predictions are compared to
time-accurate results in Fig. 11. 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 unaected 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. 11 for all ight conditions, with the model predictions being generated in few seconds.
An issue regarding SBRF model is that cost of simulating three pitch cycles for each sample shown in
Fig. 10(a) is around 52 wall-clock hours 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. Note that the ROM based on these functions is still cheaper than the full-order model and
SBRF model because the indicial functions eliminate the need of repeating calculation for each frequency.
The samples could be generated using methods of factorial design, Latin hypercube sampling, low discrepency
sequences, and designs based on statistical optimality criteria.
85
Factorial designs are extremely simple to
construct and have been used in this work. The X-31 motions considered encompass and M values in the
range of [7
o
, 7
o
] 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. 12(a). The indicial functions with respect to
angle of attack are calculated using the CFD and grid motion approach for each sample conditions. The pitch
rate indicial functions are calculated for a unit step change in the pitch rate for each Mach number shown in
13 of 29
American Institute of Aeronautics and Astronautics
Fig. 12(a). The total cost of generating all functions is now approximatly 90 hours. The calculated indicial
functions due to a unit step change in angle of attack are shown in Fig. 12(b) for each Mach number in the
sample design. This gure shows that the pitch moment initial, intermediate, and nal loadings are dierent
at each Mach number. The initial peak in the pitch moment becomes smaller for higher Mach number. An
explanation is given by Leishman;
86
this is due to the propagation of pressure disturbances at the speed of
sound. Note that the pitch moment responses suddenly change for Mach numbers above 0.85 due to shock
formation. Fig. 12(c) also shows that pitch rate indicial functions decrease as Mach number increases. A
new ROM is now created along with a time-dependent surrogate model to determine the terms in Eq. 11 at
each time step. The validity of ROM is tested for several motions in the angle of attack/frequency/Mach
number space and compared with time-accurate CFD simulations in Fig. 13. This gure shows that the
ROM predictions agree well with the CFD data. 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 pitch moment changes
suddenly at high speeds. More samples at high speeds could improve the model predictions. The future
work extends the results to include dierent sampling strategies. Finally, the created ROM could predict
many motions in the angle of attack/frequency/Mach number space within a few seconds.
V. Conclusions
The aircraft stability and control analysis requires a very large number of CFD simulations to determine
appropriate forcing parameters within the frequency/amplitude/Mach number space. 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 solution for several oscillations.
Also, the conguration used in this work, the X-31 aircraft, has highly swept slender wings resulting in
complex vortical ow under various conditions. A highly rened mesh, small time step, and the use of
hybrid turbulence models such as Detached-Eddy Simulation and Delayed Detached-Eddy simulation are
required to accurately resolve the unsteady ow around the aircraft in space and time. Because of the
combination of large grids, small time steps, hybrid turbulence models, and a large number of simulations,
the full-order modeling approach is too expensive for stability and control analysis of aircraft. This paper
investigates the use of reduced order models that signicantly reduce 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 Duhamels superposition integral using indicial (step) response
functions, Volterra theory using nonlinear kernels, Radial Basis functions, and a surrogate-based recurrence
framework, both using time-history simulations of a training maneuver(s). The indicial functions were
directly calculated from unsteady RANS simulations starting from an initial steady-state condition with a
prescribed grid motion. An important feature of this approach is uncoupling the eects of angle of attack
and pitch rate into responses from pitching motions. A method to eciently 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 limited number of samples. An indirect method was proposed
to estimate the nonlinear Volterra kernels from time-accurate computational uid dynamic simulations of
chirp and spiral training maneuvers. These maneuvering simulations were also used to estimate the unknown
parameters in a model based on Radial Basis functions. A Design of Experiment method was used to generate
several pitching motions at dierent amplitudes and free-stream Mach numbers. The model based on a
surrogate-based recurrence framework then approximated the aerodynamic responses induced by pitching
motions at new amplitides and Mach numbers. Overall, the reduced order models were found to produce
accurate results, although a nonlinear model based on indicial functions yielded the best accuracy among all
models. This model, along with a developed time-dependent surrogate approach, helped to produce accurate
predictions for a wide range of motions in the transonic speed range. The cost of generating each motion
using time-accurate CFD simulations was approximatly 52 wall-clock hours using 256 processors (2.3 GHz)
but the reduced order model predictions were within a few seconds. The future work extends the results to
include dierent sampling strategies and will include aerodynamics modeling of lateral forces and moments
as well.
14 of 29
American Institute of Aeronautics and Astronautics
VI. Acknowledgements
Mehdi Ghoreyshi was supported by the National Research Council/US Air Force Oce of Scientic
Research. Computer time was provided by the Department of Defence Engineering Research Development
Center (ERDC). The USAFA authors appreciate the support provided by the Modeling and Simulation
Research Center.
References
1
Nelson, R. C. and Pelletier, A., The Unsteady Aerodynamics of Slender Wings and Aircraft Undergoing Large Amplitude
Maneuvers, Progress in Aerospace Sciences, Vol. 39, No. 2, 2003, pp. 185284.
2
Silva, W. A., Application of Nonlinear Systems Theory to Transonic Unsteady Aerodynamics Responses, Journal of
Aircraft, Vol. 30, No. 5, 1993, pp. 660668.
3
Tobak, M. and Schi, L. B., On the Formulation of the Aerodynamic Characteristics in Aircraft Dynamics, NASA TR
R456, 1976.
4
Liebe, R., Flow Phenomena in Nature: A Challenge to Engineering Design, WIT Press, Southampton, Great Britain,
2007.
5
Sch utte, A., Einarsson, G., Raichle, A., Schoning, B., M onnich, W., and Forkert, T., Numerical Simulation of Manoeu-
vreing Aircraft by Aerodynamic, Flight Mechanics, and Structural Mechanics Coupling, Journal of Aircraft, Vol. 46, No. 1,
2009, pp. 5364.
6
Ghoreyshi, M., Vallespin, D., Da Ronch, A., Badcock, K. J., Vos, J., and Hitzel, S., Simulation of Aircraft Manoeuvres
based on Computational Fluid Dynamics, AIAA Paper 20108239, August 2010.
7
Ghoreyshi, M., Jir asek, A., and Cummings, R. M., CFD Modeling for Trajectory Predictions of a Generic Fighter
Conguration, AIAA Paper 20116523, August 2011.
8
Chin, S. and Lan, C. E., Fourier Functional Analysis for Unsteady Aerodynamic Modeling, AIAA Journal , Vol. 30,
No. 9, 1992, pp. 22592266.
9
Lisandrin, P., Carpentieri, G., and van Tooren, M., Investigation over CFD-based Models for the Identication of
Nonlinear Unsteady Aerodyanmics Responses, AIAA Journal , Vol. 44, No. 9, 2006, pp. 20432050.
10
McDaniel, D. R., Cummings, R. M., Bergeron, K., Morton, S. A., and Dean, J. P., Comparisons of CFD Solutions of
Static and Maneuvering Fighter Aircraft with Flight Test Data, Journal of Aerospace Engineering, Vol. 223, No. 4, 2009,
pp. 323340.
11
Lucia, D. J., Beran, P. S., and Silva, W. A., Reduced-Order Modeling: New Approaches For Computational Physics,
Progress in Aerospace Sciences, Vol. 40, 2004, pp. 51117.
12
Jategaonkar, R. V., Flight Vehicle System Identication, 2006, AIAA Educational series, Vol. 216, Reston, VA.
13
Ballhaus, W. F. and Goorjian, P. M., Computation of Unsteady Transonic Flow by the Indicial Method, AIAA Journal ,
Vol. 16, No. 2, 1978, pp. 117124.
14
Tobak, M., Chapman, G. T., and Schi, L. B., Mathematical Modeling of the Aerodynamic Characteristics in Flight
Dynamics, NASA TN85880, 1984.
15
Tobak, M. and Chapman, G. T., Nonlinear Problems in Flight Dynamics Involving Aerodynamic Bifurcations, NASA
TN86706, 1985.
16
Hall, K. C., Thomas, J. P., and Dowell, E. H., Reduced-order Modelling of Unsteady Small-Disturbance Using a
Frequency-Domain Proper Orthogonal Decomposition Technique, AIAA Paper 1999655, Jan 1999.
17
Gaitonde, A. and Jones, D. P., Reduced Order State-Space Models from the Pulse Responses of a Linearized CFD
Scheme, International Journal of Numerical Methods in Fluids, Vol. 42, No. 6, 2003, pp. 581606.
18
Silva, W. A. and Bartels, R. E., Development of Reduced-Order Models for Aeroelastic Analysis and Flutter Prediction
Using the CFL3Dv6.0 Code, Journal of Fluids and Structure, Vol. 19, No. 6, 2004, pp. 729745.
19
Jir asek, A., Jeans, T. L., Martenson, M., Cummings, R. M., and Bergeron, K., Improved Methodologies For Maneuver
Design of Aircraft Stability and Control Simulations, AIAA Paper 2010515, Jan 2010.
20
Goman, M. G. and Khrabov, A. N., State-Space Representation of Aerodynamic Characteristics of An Aircraft at High
Angles of Attack, AIAA Paper 924651, August 1992.
21
Juang, J., System Identication of a Vortex Lattice Aerodynamic Model, NASA TM-2001-211229, October 2001.
22
Leishman, J. G. and Nguyen, K. Q., State-Space Representation of Unsteady Airfoil Behavior, AIAA Journal , Vol. 28,
No. 5, 1989, pp. 836844.
23
Reisenthel, P. H., Development of a Nonlinear Indicial Model Using Response Functions Generated by a Neural Network,
AIAA Paper 970337, Jan 1997.
24
Reisenthel, P. H. and Bettencourt, M. T., Data-based Aerodynamic Modeling Using Nonlinear Indicial Theory, AIAA
Paper 970337, Jan 1999.
25
Ghoreyshi, M. and Cummings, R. M., Aerodynamics Modeling of a Maneuvering Aircraft Using Indicial Functions,
AIAA Paper 2012689, Jan 2012.
26
Librescu, L. and Song, O., Thin-Walled Composit Beams: Theory and Application, Springer, Dordrecht, Netherland,
2006.
27
Manglano-Villamarine, C. E. and Shaw, S. T., Three-dimensional Indicial Response of Finite Aspect Ratio Yawed
Wings, The Aeronautical Journal , Vol. 111, No. 1120, 2007, pp. 359371.
28
Wagner, H.,

Uber die Entstehung des Dynamischen Auftriebes von Trag ugeln, Zeitschrift f ur Angewandte Mathematik
und Mechanik, Vol. 1, 1925, pp. 1735.
15 of 29
American Institute of Aeronautics and Astronautics
29
Bisplingho, R. L. Ashley, H. and Halfman, R. L., Aeroelasticity, Dover Publications, Mineola, N.Y., 1996.
30
Leishman, J. G., Indicial Lift Approximations for Two-Dimensional Subsonic Flow as Obtained from Oscillatory Mea-
surements, Journal of Aircraft, Vol. 30, No. 3, 1993, pp. 340351.
31
Singh, R. and Baeder, J., Direct Calculation of Three-Dimensional Indicial Lift Response Using Computational Fluid
Dynamics, Journal of Aircraft, Vol. 34, No. 4, 1997, pp. 465471.
32
Ghoryeshi, M., Jir asek, A., and Cummings, R. M., Computational Investigation into the Use of Response Functions for
Aerodynamic Loads Modeling, AIAA Journal , Vol. 50, No. 6, 2012, pp. 13141327.
33
Leishman, J. G. and Crouse, G. L., State-Space Model for Unsteady Airfoil Behavior and Dynamic Stall, AIAA Paper
891319, April 1989.
34
Ghoreyshi, M., Badcock, K. J., and Woodgate, M. A., Accelerating the Numerical Generation of Aerodynamic Models
for Flight Simulation, Journal of Aircraft, Vol. 46, No. 3, 2009, pp. 972980.
35
Volterra, V., Theory of Functionals, Blackie, London, 1930.
36
Wiener, N., Nonlinear Problems in Random Theory, MIT Press, Cambridge, MA, 1958.
37
Barrett, J. F., The Use of Functionals in the Analysis of Non-linear Physical Systems, Journal of Electr. Control ,
Vol. 15, 1963, pp. 567615.
38
Kalman, R. E., Pattern Recognition Properties of Multilinear Response Functions, Journal of Control Cybern., Vol. 9,
1980, pp. 531.
39
Stark, L., Neurological Control System: Studies in Bioengineering, Plenum Press, New York, 1968.
40
Weiner, D. D. and Spina, J. F., Sinusoidal Analysis and Modeling of Weakly Nonlinear Circuits, New York: Van Nostrand
Reinhold, 1980.
41
Hunter, I. W. and Korenberg, M. J., The Identication of Nonlinear Biological Systems: Wiener and Hammerstein
Cascade Models, Journal of Biol. Cybern., Vol. 55, 1986, pp. 135144.
42
Korenberg, M. J., A Fast Orthogonal Search Method For Biological Time-Series Analysis and System Identication,
Proceeding of IEEE International Conference Sys. Man. Cybernet., 1989, pp. 459465.
43
Korenberg, M. J., A Robust Orthogonal For System Identication and Time-Series Analysis, Journal of Biological
Cybernet, Vol. 60, 1989, pp. 267276.
44
Jir asek, A. and Cummings, R. M., Application of Volterra Functions to X-31 Aircraft Model Motion, AIAA Paper
20093629, June 2009.
45
Marques, F. D. and Anderson, J., Identication and Prediction of Unsteady Transonic Aerodynamic Loads by Multi-
Layer Functionals, Journal of Fluids and Structure, Vol. 1, No. 6, 2011, pp. 83106.
46
Silva, W. A., Discrete-Time Linear and Nonlinear Aerodynamic Impulse Responses for Ecient CFD Analysis, PhD
Dissertation, Faculty of the Department of Applied Science, The College of William and Mary, Williamsburg, VA, September
1997.
47
Da Ronch, A., McCracken, A., Badcock, K. J., Ghoreyshi, M., and Cummings, R. M., Modeling of Unsteady Aerody-
namic Loads, AIAA Paper 20112376, Aug. 2011.
48
Faller, W. E. and Schreck, S. J., Neural Networks: Applications and Opportunities in Aeronautics, Progress in
Aerospace Sciences, Vol. 32, No. 5, 1996, pp. 433456.
49
Faller, W. E. and Schreck, S. J., Real-Time Prediction of Unsteady Aerodynamics: Application for Aircraft Control and
Maneuverability Enhancement, IEEE Trans. Neural Networks, Vol. 6, No. 6, 1995, pp. 14611468.
50
Glaz, B., Liu, L., and Friedmann, P. P., Reduced-Order Nonlinear Unsteady Aerodynamic Modeling Using a Surrogate-
Based Recurrence Framework, AIAA Journal , Vol. 48, No. 10, 2010, pp. 24182429.
51
Ghoryeshi, M., Jir asek, A., Post, M. L., and Cummings, R. M., Computational Approximation of Nonlinear Aerody-
namics Using an Aerodynamic Model Hierachy, AIAA Paper 20113667, June 2011.
52
Strang, W. Z., Tomaro, R. F., and Grismer, M. J., The Dening Methods of Cobalt: A Parallel, Implicit, Unstructured
Euler/Navier-Stokes Flow Solver, AIAA Paper 19990786, 1999.
53
Da Ronch, A., Vallespin, D., Ghoreyshi, M., and Badcock, K. J., Evaluation of Dynamic Derivatives Using Computa-
tional Fluid Dynamics, AIAA Journal , Vol. 50, No. 2, 2012, pp. 470484.
54
Gottlieb, J. J. and Groth, C. P. T., Assessment of Riemann Solvers For Unsteady One-dimensional Inviscid Flows of
Perfect Gasses, Journal of Fluids and Structure, Vol. 78, No. 2, 1998, pp. 437458.
55
Tomaro, R. F., Strang, W. Z., and Sankar, L. N., An Implicit Algorithm For Solving Time Dependent Flows on
Unstructured Grids, AIAA Paper 19970333, 1997.
56
Forsythe, J. R., Homann, K. A., Cummings, R. M., and Squires, K. D., Detached-Eddy Simulation With Compressibil-
ity Corrections Applied to a Supersonic Axisymmetric Base, Journal of Fluids Engineering, Vol. 124, No. 4, 2002, pp. 911923.
57
Forsythe, J. R. and Woodson, S. H., Unsteady Computations of Abrupt Wing Stall Using Detached-Eddy Simulations,
Journal of Aircraft, Vol. 42, No. 3, 2005, pp. 606616.
58
Morton, S. A., Forsythe, J. R., Mitchell, A. M., and Hajek, D., Detached-Eddy Simulations and Reynolds-Averaged
Navier-Stokes Simulations of Delta Wing Vortical Flowelds, Journal of Fluids Engineering, Vol. 124, No. 4, 2002, pp. 924932.
59
Forsythe, R., Squires, K. D., Wurtzler, K. E., and Spalart, P. R., Detached-Eddy Simulation of Fighter Aircraft at High
Alpha, Journal of Aircraft, Vol. 41, No. 2, 2004, pp. 193200.
60
Jeans, T., McDaniel, D., Cummings, R. M., and Mason, W., Aerodynamic Analysis of a Generic Fighter Using Delayed
Detached-Eddy Simulations, Journal of Aircraft, Vol. 46, No. 4, 2009, pp. 13261339.
61
Balajewicz, M., Nitzsche, F., and Feszty, D., Application of Multi-Input Volterra Theory to Nonlinear Multi-Degree-of-
Freedom Aerodynamic Systems, AIAA Journal , Vol. 48, No. 1, 2010, pp. 5662.
62
Duy, D. G., Advanced Engineering Mathematics With MATLAB, second edition, Chapman and Hall/CRC, Florida,
2003.
16 of 29
American Institute of Aeronautics and Astronautics
63
Romero, D. A., Amon, C., Finger, S., and Verdinelli, I., Multi-Stage Bayesian Surrogates For the Design of Time-
Dependent System, Proceedings of ASME 2004 Design Engineering Technical Conference and Computers and Information
in Engineering, DETC 2004-57510, 2004.
64
Jones, D. R., Schonlau, M., and Welch, W. J., Ecient Global Optimization of Expensive Black-Box Functions, Kluwer
Academic Publications, Vol. 13, 1998, pp. 455492.
65
Da Ronch, A., Ghoreyshi, M., and Badcock, K. J., On the Generation of Flight Dynamics Aerodynamic Tables by
Computational Fluid Dynamics, Progress in Aerospace Sciences, Vol. 47, 2011, pp. 597620.
66
Arbib, M. A., The Handbook of Brain Theory and Neural Networks, Second Edition, The MIT Press, 2003.
67
Dasgupta, D. and Michalewicz, Z., Evolutionary Algorithms in Engineering Applications, Springer-Verlag, Berlin, Hei-
delberg, 1997.
68
Balajewicz, M., Nitzsche, F., and Feszty, D., Reduced Order Modeling of Nonlinear Transonic Aerodynamics Using a
Pruned Volterra Series, 50th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference,
Palm Springs, CA 2009.
69
Morelli, E. A., System IDentication Programs for AirCraft (SIDPAC), AIAA Paper 20024704, August 2002.
70
G ortz, S., McDaniel, D., and Morton, S. A., Towards an Ecient Aircraft Stability and Control Analysis Capability
Using High-Fidelity CFD, AIAA Paper 2007-1053, Jan 2007.
71
ONeill, C. and Arena, A., Time Domain Training Signals Comparison for Computational Fluid Dynamics Based Aero-
dynamic Identication, Journal of Aircraft, Vol. 42, No. 2, 2005, pp. 421.428.
72
Murpy, P. C. and Klein, V., Validation of Methodology for Estimating Aircraft Unsteady Aerodynamics Parameters
from Dynamic Wind Tunnel Tests, AIAA Paper 20035397, August 2003.
73
Morelli, E. A., Real Time Parameters Estimation in the Frequency Domain, AIAA Paper 994043, August 1999.
74
Cummings, R. M. and Sch utte, A., An Integrated Computational/Experimental Approach to UCAV Stability & Control
Estimation: Overview of NATO RTO AVT-161, AIAA Paper 20104392, June-July 2010.
75
Canter, D. E. and Groves, A. W., X-31 Post-Stall Envelope Expansion and Tactical Utility Testing, AIAA Paper
19942171, June 1994.
76
Alcorn, C. W., Croom, M. A., and Francis, M. S., The X-31 Experience - Aerodynamic Impediments to Post-Stall
Agility, AIAA Paper 1995362, Jan 1995.
77
Williams, D. L., Nelson, R. C., and Fisher, D., An Investigation of X-31 Roll Characteristics at High Angle-of-Attack
Through Subscale Model Testing, AIAA Paper 1994806, Jan 1994.
78
Rein, M., H ohler, G., Sch utte, A., Bergmann, A., and L oser, T., Ground-Based Simulation of Complex Maneuvers of a
Delta-Wing Aircraft, AIAA Paper 20063149, June 2006.
79
Schefter, J., X-31: How Theyre Inventing a Radical New Way to Fly, Popular Science, Vol. 234, No. 2, 1989, pp. 5864.
80
Tyssel, L., Hybrid Grid Generation for Complex 3D Geometries, Proceedings of the 7th International Conference on
Numerical Grid Generation in Computational Field Simulation, 2000, pp. 337346.
81
Tyssel, L., The TRITET Grid Generation System, International Society of Grid Generation (ISGG), Proceedings of
the 10th International Conference on Numerical Grid Generation in Computational Field Simulations, Forth, Crete, Greece,
2000.
82
Spalart, P. R. and Allmaras, S. R., A One Equation Turbulence Model for Aerodynamic Flows, AIAA Paper 19920439,
January 1992.
83
Spalart, P. R. and Schur, M., On the Sensitisation of Turbulence Models to Rotation and Curvature, Aerospace Science
and Technology, Vol. 1, 1997, pp. 297302.
84
Menter, F., Two-Equation Eddy-Viscosity Turbulence Models for Engineering Applications, AIAA Journal , Vol. 32,
No. 8, 1994, pp. 15981605.
85
Mackman, T. J., Allen, C. B., Ghoreyshi, M., and Badcock, K. J., Comparison of Adaptive Sampling Methods for
Generation of Aerodynamic Models for Flight Simulations, AIAA Paper 20111171, January 2011.
86
Leishman, J. G., Subsonic Unsteady Aerodynamics Caused by Gusts Using the Indicial Method, Journal of Aircraft,
Vol. 33, No. 5, 1996, pp. 869879.
17 of 29
American Institute of Aeronautics and Astronautics
2.63
45 deg
57 deg
45 deg
50 deg
13.21
Length unit in meter
7.26
Figure 1. The X-31 aircraft geometry
(a) Half-Model Mesh (b) Surface mesh around LEX and canard
Figure 2. The X-31 aircraft mesh model
18 of 29
American Institute of Aeronautics and Astronautics
Angle of Attack, (DEG)
C
L
-10 0 10 20 30 40 50 60
-0.5
0.0
0.5
1.0
1.5
Exp.
SA
SARC
SST
SARC-DDES
Angle of Attack, (DEG)
C
D
-10 0 10 20 30 40 50 60
0.0
0.5
1.0
1.5
Exp.
SA
SARC
SST
SARC-DDES
(a) lift coecient (b) drag coecient
Angle of Attack, (DEG)
C
m
-10 0 10 20 30 40 50 60
-0.05
0.00
0.05
0.10
Exp.
SA
SARC
SST
SARC-DDES
(c) pitching moment coecient
Figure 3. The X-31 static loads validations. The static conditions are : M=0.18 and Re = 2 10
6
.
19 of 29
American Institute of Aeronautics and Astronautics
(a) = 10
0
(b) = 14
0
(c) = 18
0
(d) = 20
0
(e) = 22
0
(f) = 25
0
Figure 4. The X-31 vortical ows using DDES-SARC turbulence model. The conditions are : M=0.18 and
Re = 2 10
6
.
20 of 29
American Institute of Aeronautics and Astronautics
Angle of attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Pt. A
Pt. B
Pt. C
Pt. D
Pt. E
Pt. F
(a) Target motion.
(b) Point A (c) Point B (d) Point C
(e) Point D (f) Point E (g) Point F
Figure 5. Target motion is a pitch harmonic motion as: = 7
o
sin(t), M=0.9, k=0.01, and Re = 2 10
6
. In
Fig. (a) static data are shown with a solid line.
21 of 29
American Institute of Aeronautics and Astronautics
Time (s)
A
n
g
l
e
o
f
a
t
t
a
c
k
,

(
D
E
G
)
0.0 0.5 1.0 1.5 2.0 2.5
-10
-5
0
5
10
Angle of Attack, (DEG)
C
m
-8 -6 -4 -2 0 2 4 6 8
-0.04
-0.02
0.00
0.02
0.04
(a) Spiral maneuver (b) ROM prediction
Time (s)
A
n
g
l
e
o
f
a
t
t
a
c
k
,

(
D
E
G
)
0.0 0.5 1.0 1.5 2.0 2.5
-10
-5
0
5
10
Angle of Attack, (DEG)
C
m
-8 -6 -4 -2 0 2 4 6 8
-0.04
-0.02
0.00
0.02
0.04
(c) Chirp maneuver (d) ROM prediction
Figure 6. Volterra reduced order modeling using Spiral and Chirp training maneuvers. The ow conditions
of training maneuvers are : M=0.9 and Re = 2 10
6
.
22 of 29
American Institute of Aeronautics and Astronautics
Time (s)
A
n
g
l
e
o
f
a
t
t
a
c
k
,

(
D
E
G
)
0.0 0.5 1.0 1.5 2.0 2.5
-10
-5
0
5
10
Angle of attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
(a) Spiral maneuver (b) ROM prediction
Time (s)
A
n
g
l
e
o
f
a
t
t
a
c
k
,

(
D
E
G
)
0.0 0.5 1.0 1.5 2.0 2.5
-10
-5
0
5
10
Angle of attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
(c) Chirp maneuver (d) ROM prediction
Time (s)
A
n
g
l
e
o
f
a
t
t
a
c
k
,

(
D
E
G
)
0.0 0.5 1.0 1.5 2.0 2.5
-10
-5
0
5
10
Angle of attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
(e) Schroeder maneuver (f) ROM prediction
Figure 7. RBF reduced order modeling- Training maneuver are a Spiral, a Chirp, and a Schroeder motion.
The ow conditions are : M=0.9 and Re = 2 10
6
.
23 of 29
American Institute of Aeronautics and Astronautics
s = 2Vt/c
I
n
d
i
c
i
a
l
f
u
n
c
t
i
o
n
(
1
/
r
a
d
)
0 20 40 60 80 100 120
-4
-3
-2
-1
0
C
mq
C
m
Angle of attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
(a) linear responses (b) ROM based on linear responses
s= 2Vt/c
C
m

(
1
/
r
a
d
)
0 20 40 60 80 100 120
-1
-0.8
-0.6
-0.4
-0.2
0
= 0
o
= 3
o
= -3
o
= -5
o
Angle of attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
(c) non-linear responses (d) ROM based on non-linear responses
Figure 8. ROM using indicial functions. The ow conditions are : M=0.9 and Re = 2 10
6
.
24 of 29
American Institute of Aeronautics and Astronautics
Angle of attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
Angle of attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
(a) = 4
o
sin(t), k =0.01 (b) = 6
o
sin(t), k =0.01
Time (s)
C
m
0 0.5 1 1.5 2 2.5
-0.04
-0.02
0.00
0.02
0.04
0.06
Time-Marching
Model
Time (s)
C
m
0 0.5 1 1.5 2 2.5
-0.04
-0.02
0.00
0.02
0.04
0.06
Time-Marching
Model
(c) chirp maneuver (d) spiral maneuver
Time (s)
C
m
0 0.5 1 1.5 2 2.5
-0.04
-0.02
0.00
0.02
0.04
0.06
Time-Marching
Model
(e) schroder maneuver
Figure 9. ROM using non-linear indicial functions for target motions at dierent angles of attack and frequency.
The ow conditions are : M=0.9 and Re = 2 10
6
.
25 of 29
American Institute of Aeronautics and Astronautics
Mach
P
i
t
c
h
a
m
p
l
i
t
u
d
e
,

(
D
E
G
)
0.7 0.75 0.8 0.85 0.9 0.95
0
2
4
6
8
(a) Samples Design for SBRF model.
(a) Training pitch motions for SBRF model.
Figure 10. Samples and training pitch motions for a SBRF model.
26 of 29
American Institute of Aeronautics and Astronautics
Angle of Attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
Angle of Attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
(a) = 7
o
sin(t), k =0.01, M=0.78 (b) = 7
o
sin(t), k =0.01, M=0.825
Angle of Attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
(e) = 7
o
sin(t), k =0.01, M=0.88
Figure 11. 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.
27 of 29
American Institute of Aeronautics and Astronautics
Mach
A
n
g
l
e
o
f
a
t
t
a
c
k
,

(
D
E
G
)
0.7 0.75 0.8 0.85 0.9 0.95
-8
-6
-4
-2
0
2
4
6
8
(a) Samples Design for indicial functions.
s = 2Vt/c
C
m

(
1
/
r
a
d
)
0 10 20 30 40
-1
-0.8
-0.6
-0.4
-0.2
0
Mach = 0.75
Mach = 0.80
Mach = 0.85
Mach = 0.90
s = 2Vt/c
C
m

(
1
/
r
a
d
)
0 10 20 30 40
-5
-4
-3
-2
-1
0
1
M = 0.75
M = 0.80
M = 0.85
M = 0.90
(a) Angle of attack indicial functions (a) Pitch rate indicial functions
Figure 12. Time-dependent surrogate modeling of indicial functions.
28 of 29
American Institute of Aeronautics and Astronautics
Angle of Attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
Angle of Attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
(a) = 7
o
sin(t), k =0.01, M=0.78 (b) = 5
o
sin(t), k =0.025, M=0.78
Angle of Attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
Angle of Attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
(c) = 7
o
sin(t), k =0.01, M=0.825 (d) = 5
o
sin(t), k =0.03, M=0.825
Angle of Attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
Angle of Attack, (DEG)
C
m
-8 -4 0 4 8
-0.04
-0.02
0.00
0.02
0.04
Time-Marching
Model
(e) = 7
o
sin(t), k =0.01, M=0.88 (f) = 5.5
o
sin(t), k =0.025, M=0.88
Figure 13. Transonic loads modeling in Mach number/angle of attack/frequency space. The ROM is based
on a time-dependent surrogate model that approximates the non-linear indicial functions at dierent ight
conditions. In above is angular velocity and k = c/2V is reduced frequency.
29 of 29
American Institute of Aeronautics and Astronautics