You are on page 1of 10

TH E AERONAUTICAL JOURNAL FEBRUARY 1997 77

Aeroelastic optimisation of composite wings


using the dynamic stiffness method
M. Lillico and R. Butler
School of Mechanical Engineering
University of Bath, UK.

S. Guo and J.R. Banerjee


Department of Mechanical Engineering and Aeronautics
City University, London, UK.

ABSTRACT
A computer program for use in the conceptual stage of aircraft Es geometric elastic axis
design has been developed. The program obtains minimum mass El bending rigidity
designs for high aspect ratio, composite wings, subject to constraints [F(V,a»] flutter matrix
on flutter speed, divergence speed and material stress. The wing is G„ value of constraint n
modelled as a series of composite beam elements and both flutter Gs mass axis
speed and divergence speed are calculated using a normal mode GJ torsional rigidity
approach. Modal analysis is carried out by applying the Wittrick- h vertical displacement
Williams algorithm to the dynamic stiffness method, whereas H, Hoffman failure index
unsteady aerodynamic loads are calculated from strip theory, 'a polar mass moment of inertia per unit length
although an option which uses lifting-surface theory is also K coupled bending-torsional rigidity
presented. A previously published example is given to validate the [K] stiffness matrix
analysis. Single level optimisation is carried out using a sequential [KD(co)] dynamic stiffness matrix
quadratic programming strategy combined with the modified meth- m mass per unit length
ods of feasible directions optimizer, for which flutter sensitivities are hi bending moment
obtained by an efficient determinant interpolation technique. Design M
BOX initial mass of element structural box
variables include topological variables such as spar and engine posi- [M] mass matrix
tions as well as layer thicknesses, which are modelled using n
* number of model elements
quadratic functions. The wing of a regional turboprop aircraft is opti- n
CON number of constraints
mized to illustrate the use of the program. The problem was "DV number of design variables
modelled using 10 elements and had 43 design variables, 162 Nu axial force due to bending and torsion
constraints and required just over 20 minutes of CPU time on a Nn shear flow due to bending and torsion
workstation. This, coupled with the fact that a full three-dimensional 1 shear flow due to shear force
FE model of the same wing would require over 1000 elements, {q} vector of generalised coordinates
illustrates the suitability of the dynamic stiffness method to the [Q] aerodynamic force matrix
conceptual design stage. 5 shear force
1 time
*FS»i thickness of 8° layer in front spar for element i
hs»j thickness of 9° layer in rear spar for element i
NOMENCLATURE l thickness of 8° layer in skin for element i
SK,B.i
ah geometric elastic axis position behind the mid-chord as T torque
a percentage of semi-chord V airspeed
A, cross sectional area enclosed by wing box W objective function
A(s) laminate reduced axial stiffness parameter X
J
value of theyth design variable
A| !-A33 laminate axial stiffness xa distance between £ 5 and Gs, negative if Gs is behind Es
b semi-chord {X} vector of design variables
B(s) laminate reduced coupling stiffness parameter y span-wise distance from wing root
c chord z(*) vertical distance from geometric elastic axis
cb wingbox chord
Cm, non dimensional zero lift pitching moment a design variable perturbation size
C(,f) laminate reduced shear stiffness parameter P allowable error in sensitivities
d wing box depth Y maximum vertical distance between parabola and chord
e error A velocity or frequency perturbation size

Paper No. 2161. Manuscript received 27 September 1995, revised version received 4 October 1996. accepted 9 January 1997.

9 layer angle two-spar wing model, which divides the wing into a series of box
p air density beam elements, is constructed and then analysed to obtain the nor-
a absolute error in flutter speed mal modes using the dynamic stiffness method. In this approach the
(a frequency one-dimensional exact elements used form an accurate representa-
y torsional displacement tion of the composite box-beam. During optimisation, the unsteady
aerodynamics of the wing are formulated based on strip theory, but
an alternative lifting-surface formulation has been implemented so
Subscripts that results can be checked using more refined analysis. Both the
flutter and divergence analyses were originally incorporated in a pro-
d divergence gram called Calfun^2\ which has been extended and used as the
/ flutter aeroelastic analysis module in the current program. Layer thickness
g generalised distribution along the span is modelled using a quadratic function,
L lower value thereby reducing die number of design variables. The paper is self-
min minimum value contained, with the inclusion of a brief discussion of die theory used
U upper value in the program for free vibration analysis, aeroelastic analysis and
optimisation. The analysis is validated against published results.
Optimisation results are presented for an example problem to
Superscripts illustrate the use and application of the program.
perturbed value

2.0 ANALYSIS FEATURES


1.0 INTRODUCTION
2.1 Wing structural model
It is well recognised that wing design is a complex process, involv-
ing die application of a number of interrelated disciplines, such as The wing is modelled as a series of uniform beam elements, can-
aerodynamics and structures. An optimum design procedure should tilevered at die root, as illustrated in Fig. 1(a). Each element consists
successfully balance diese various disciplines to produce an accept- of a rectangular composite box comprised of die two spars togedier
able compromise. In practice, however, designs are often frozen with the upper and lower skins between die spars (Fig. 1(b)). The
before all of die disciplines have been fully considered. In particular, cross-sectional properties EI, GJ, K, m, Ia, xa and ah art constant for
aeroelastic criteria are often considered once die wing has largely each element and the geometric elastic axis, Es, whose position is
been designed to satisfy other requirements and aeroelastic problems defined by ah, is die elastic axis of an equivalent box of identical
such as flutter or divergence speed are dealt with by making ad hoc
changes to rigidity and mass distributions. Clearly, a wing design
resulting from this process cannot be expected to be optimum.
Aeroelastic and multidisciplinary optimisation have been the 9 Engine e.g.
subject of an increasing amount of research over die last 30 years as
improved analysis and optimisation techniques have been developed
and as computer power has increased. During this period various
computer programs have been developed, of which the most widely
used seems to be Astro&v. Otiiers include FastopM, AdopW, StarsW, ,. DsL" 1 -, — " « _,
Catia-ElfiniW and Optsysf®. Usually diese programs are run using
detailed, three-dimensional finite element (FE) models, which at die
conceptual design stage, i.e. when die effect of major design changes
(e.g. to engine position or spar position) needs to be investigated
rapidly, can be botii difficult to construct and time consuming to
analyse. In contrast, for simple and quick conceptual calculations, a
computer program based on one-dimensional FE models which
accurately predict the behaviour of composite materials would be
better suited for conceptual design. To the audiors' knowledge, such (a) Wing planform.
a program does not exist.
Some work has been carried out in developing programs suitable
for conceptual design. An early, widely used program is TSO (wing
aeroelastic syndiesis procedure)'7'- This program which models die Skin
wing as an equivalent composite plate, takes into account aeroelastic
Rear Front
behaviour such as flutter and divergence. A program developed later
spar Gsx f *E,
by Wakayama and Kroo<8> uses a simple box beam to model spar.
me (metallic) structure, but only includes the effect of static aero- Skin
elasticity and cannot be used to optimise composite wings, which are
becoming more important in the design of new aircraft. This pro- 0.35M box 0.15M'box
gram'8' does however include wing planform geometry parameters as 12
variables. A composite wing program which also uses a box beam
I h > Cv/2

model was developed by Austin et a/'9'. Their approach applies con-


0.8c
straints on aeroelastic behaviour in an indirect manner by limiting ic °- 2 c «<-
die rigidity properties of me wing to acceptable values.
The current paper describes a program called Calfunopt for die (b) Section A-A
optimum (minimum mass) design of high aspect ratio, composite
wings at the conceptual structural design stage. As before'10-'", opti- Figure 1. Wing model. In (a) 6 is the ply angle in the x-y plane (note
misation is constrained by aeroelastic requirements, but additional that positive ply angle in the y-z plane is measured downwards from
constraints have been added to limit die material stress. A simplified the y-axis).
• • H H H M H I
bILLICO, BUTLER, G U O AND BANERJEE AEROELASTIC OPTIMISATION OF COMPOSITE WINGS

cross-sectional geometry but made from isotropic material. The improve accuracy, particularly for higher modes, whereas the DSM
effect of material bending-torsion coupling of composite beams is uses an exact shape function to form a single (frequency dependent)
taken into account by assuming that the relationships between the dynamic stiffness matrix and so the solution is independent of the
forces and displacements are defined by the following differential number of elements used in the analysis. Therefore, when compared
equations. to an FE analysis of the same problem, the DSM has the advantage
of using a much simpler, but sufficiently accurate, model.

W=
"0-<f (1)
2.3 Aeroelastic analysis using the determinant and V-g
methods
The flutter speed of a wing modelled as above is found using the
.(2) normal mode method and generalised coordinates. Here, the
dy dy
equation of motion for an oscillating wing in an air flow can be
expressed, using the DSM, as follows,
where h, y and V|/ are defined in Fig. 2 and the El, GJ and K of each
element are calculated using the theory developed by Berdichevsky
et a/"3' and Armanios and Badir"4'. The bending-torsional coupling r„ , o PV2[QU
behaviour of the wing is accounted for by the use of both K and xa {q} = {o} (5)
terms. K is the material bending-torsion coupling and the ,vH term
takes into account the bending-torsion coupling arising from non-
coincident mass and geometric elastic axes. By splitting the coupling [F(V,co)]{q} = {0} (6)
into two separate terms, an insight is obtained into the relative
importance of each type of coupling whilst achieving the aeroelastic
requirements. The composite laminates used for the wing model where [KD(co)]g represents the generalised dynamic stiffness
consist of up to eight layers and are symmetrical, but may be matrix of the wing and [Q]g is the generalised unsteady aerodynamic
unbalanced. For simplicity it is assumed that the upper and lower matrix, calculated by using either strip theory, assuming two dimen-
skins are identical. sional flow using Theodorsen's method'"-20' or lifting-surface (L-S)
theory, based on Davies' method'2" which takes three dimensional
The box is considered to be the principal load carrying component flow effects and compressibility into account.
and any structure in front of the front spar and behind the rear spar is It is well known that strip theory is not very accurate for modern,
assumed not to contribute towards the rigidity of the wing, although high-speed, subsonic aircraft and that it is not suitable for the
some contribution towards the mass and inertia is allowed for. To detailed design of such aircraft. However, it is used here to calculate
account for this, as well as for the effect of systems within the wing, the aeroelastic constraints during optimisation, see Section 3, to pro-
masses of 35% and 15% of the initial element wing box mass, MBOX, vide a basic indication of the flutter and divergence boundaries at the
are assumed at 80% of the element chord and at the centre of the conceptual design stage. The L-S option then provides a means of
wing box, respectively, as shown in Fig. 1(b). An engine can also be checking the acceptability of the conceptual design produced by the
modelled through the use of a lumped mass, with both mass and program.
inertia properties attached at a defined distance from one of the
The flutter speed Vf and flutter frequency (u^ are obtained when
nodes (shown as dots in Fig. 1(a)).
Equation (5) is satisfied for the non-trivial case, i.e. when the deter-
minant of the coefficient matrix of Equation (6) is zero. Two inde-
pendent methods have been used to calculate V, and u>f. These can
2.2 Free vibration analysis using the dynamic stiffness
be obtained in an iterative manner by stepping through a range of
method
airspeeds and frequencies until the determinant of the coefficient
In this paper, the dynamic stiffness method (DSM) has been used to matrix is zero. This method, which is referred to here as the determi-
carry out free vibration analysis for a non-uniform beam. A more nant method, is computationally expensive when the search ranges
detailed treatment of the theory has been given elsewhere'15-17'. for airspeed and frequency are large. For this reason (and also
Briefly, the dynamic stiffness matrix for each element of a composite because a determinant interpolation technique has been developed),
beam is obtained from the exact analytical solution of the following an adapted form of the determinant method is used in conjunction
differential equations of motions for each element with the following, more efficient, method in the way described in
Section 3.2 below.
The second method is the V-g method'22' which converts Equation
El — j - + K —\ 3 + m —-j- - mx. —-f- =0 . . . (3) (5) into a complex eigenvalue equation. To use the V-g method,
dy* dy dt2 " dt2 generalised mass, [M]g, and stiffness matrices, [K]g, have to be
extracted from the generalised dynamic stiffness matrix, [KD(o))]t, as
follows
_ , a v a3/. d2h , a 2 v
=0 . . . (4)
GJ —-=- + A —-=• + ffu, dr dt2
dy dy f[K D (co4-[K D (cD,)]^
[M]g = .. (7)
Hence, within the assumptions of these equations, the dynamic (to?-co2)
stiffness matrix, which contains both rigidity and mass properties,
represents an exact (continuous) model of the element. Individual
element matrices within a structure are assembled using the standard [^([KDML+O^M],) . . . (8)
stiffness matrix assembly procedure to obtain the dynamic stiffness
matrix for the whole wing. The calculation of natural frequencies
and mode shapes follows from the well known algorithm of Wittrick
and Williams'18'. where (Oj and (O2 are two small, arbitrary values of frequency.
This approach differs from traditional FE methods, which approx- As a static instability problem, the divergence analysis is treated
imate displacements with shape functions to obtain separate mass as a simplified case of flutter analysis (i.e. (0 = 0). In this case a
and stiffness matrices. Thus FE methods require more elements to similar approach to the determinant method is used.
80 T H E AERONAUTICAL JOURNAL FEBRUARY 1997

2.4 Material strength calculation flow in the spars is given by the sum of q and A/,2. Once the laminate
loads, Nu, Nu and q are obtained, laminate analysis is carried out to
Strength constraints have been included in the optimisation so that obtain the axial, transverse and shear stresses in the spar and skin
the effect of loads on skins and spars are accounted for, improving laminate layers.
the realism of an optimum design. It has been assumed that the bend- To reduce the number of strength constraints imposed on the
ing loads are carried by the skin laminates and that the shear loads design, the Hoffman Failure Index is used to determine if the
are carried by spar laminates but both skins and spars carry the shear stresses in a layer will cause it to fail or not. For a wing of neS box
due to torsion. The wing bending, torsion and shear loads are found elements, each consisting of four symmetric laminates, with each
at the root of each element in the wing model, assuming that the lift laminate composed of eight layers, failure occurs when the following
per unit (wetted) area over the wing is constant. inequality is not satisfied.
The stresses due to bending and torsion at the root of each beam
element are found using the theory developed by. Berdichevsky et Hk<\ J=l,2,...(l6xn„) ..(15)
a/d3-i4)i a ft e r determining the curvature and rate of change of twist
for each element using Equations (1) and (2). Assuming that the where for each layer
loading produces no lateral or extensional displacement, the axial
force, Nu, in each box laminate at the root of each element due to
...(16)
bending and torsion (via coupling) is obtained from
and
Ae^B(s) 1 1 1 1 1
Nu=-A(s)z{s)-. XT Xc YT Yc 1|- {XTXC)
dy2 (17)
[*U?) . . . (9)
7*79 = T
(YTYC)
r ; Fi-i — —x- and F^ •
$ {2XTXC)
•1

B(s)2z(s) d2h
<m*» B(s) d2h XT is the longitudinal tensile failure stress, Xc is the longitudinal
C(s) dy2 C(s) dy2 compressive failure stress, YT is the transverse tensile failure stress,
f-L-d* Yc is the transverse compressive failure stress and S is the shear fail-
J C s) ure stress. (Note that both Yc and Xc are absolute values). The layer
longitudinal stress is denoted by f,, layer transverse stress by f2 and
The shear flow, Nn, in each laminate at the root of each element due layer shear stress by f[2.
to torsion and bending (via coupling) is obtained using Since the Hoffman Failure Index includes the effect of all three
layer stresses, its use reduces the number of strength constraints that

N.,
1
4-LdA
C(s
J C x\
-24
'ft
dy
-2
m^
.B2htB{s)
V J GM
...(10)
have to be imposed on the design for a given loading condition by
66% compared with constraints which consider individual layer
stresses.

In the above
3.0 OPTIMISATION
A(s) = Au M ...(11) To find a minimum mass wing design which satisfies constraints on
flutter speed, divergence speed and material strength, the following
optimisation problem is solved,

B(s) = 2 ..(12) minimise W • •(18)

subject to G„ < 0 for n = 1,2 nc (19)

C(s) = 4 M • •(13)
andlX,.}^}^} (20)

where W, the objective function, is the mass of the wing, G„ are con-
and z(s) is the vertical distance between the origin (the centre of the straints and {XL} and {XuJ are lower and upper bounds, respectively,
wing box) and the skin for which Nu is being calculated. (Note that on the design variables, {X} (note that W and Gn are functions of the
all the integrals in Equations (9) and (10) are closed integrals along design variables). The constraints are calculated from
the circumference of the cross-section of the beam element and the
variable V is measured around the circumference of the cross-
section). As can be seen A/]2 is constant around the circumference of
the cross-section whilst N, t is dependent upon A(s), B(s), C(s) and G,=l- ...(21)
" / min
z(s) of the laminate.
The shear flow in the spars due to shear load is found in a simple
manner using the following equation, G, =1 vd • • • (22)
S (14)
q=
Yd • • •(23)
; 2+ . = H,-1 for i = 1,2,....(16xne[)
which assumes that the vertical shear load is reacted only by the web
of the spars, that the load is equally distributed between the front and V/m,„ and Vdmin are the design flutter and divergence speeds, respec-
rear spars and that the shear load in the two spars is uniform. Shear tively. During optimisation and sensitivity analysis Vs and Vd are
•ILLICO, BUTLER, G U O AND BANERJEE ROELASTIC OPTIMISATION OF COMPOSITE WINGS

Optimisation procedure MMFD SQP/MMFD by interpolation using a quadratic function. This approach, which
was also used by Berkes<24>, gives the advantage that all the design
Optimum mass (kg) 186 38 18101 variables are global (i.e. they affect the whole wing) and therefore
optimisation can take place within a single level. Also, because of
Iterations/cycles 12 8
the resulting smooth variation in thickness, the manufacture of the
DOT time (s) 0-44 22-67
wing is likely to be easier than if wing thicknesses were allowed to
Sensitivity time (min) 11-24 818
vary independently for each element. The method gives three design
Constraint time (min) 53-29 12-63 variable per laminate ply, resulting in a total of 36 thickness design
Total CPU time (min) 64-54 21 19
variables for the composite wing and nine thickness design variables
for the metallic wing.
Table 1
Comparison of optimum mass and CPU time for different During optimisation, each spar is constrained to remain straight.
optimisation procedures (note that an iteration is a program loop Hence there are two position variables per spar
at the optlmlser level and a cycle is a loop at the strategy level) 1. the slope of the spar relative to the quarter chord line
2. the chord-wise distance between the spar and the quarter chord
calculated to an accuracy of ±0-0005 ms-1 (note that (Of is calculated line in the tip element.
to an accuracy of ±0-0005 rads-')- The engine position variables are
Optimisation is carried out using a sequential quadratic program- 1. position in front of the wing quarter chord line
ming strategy (SQP) combined with the modified methods of feasi- 2. distance from the wing root
ble directions (MMFD) optimizer implemented in the mathematical 3. distance below the geometric elastic axis.
programming optimizer DOT'23'. This is a gradient search technique The bounds on the spar position variables are that the slope of the
which uses a search direction, found by considering the sensitivity to front (rear) spar is allowed to vary between a position parallel to the
design variable changes of the active constraints, the violated con- quarter chord and a position parallel to the wing leading (trailing)
straints and the objective function, to solve Equations (18) - (20) by edge. Also movement of up to ±5% of the tip chord is allowed for
iteratively applying design variable changes to the design. A both spars in the tip element. The engine eg is allowed to move
constraint is considered active if it is satisfied but within 5% of its spanwise by up to 49% of the width of the elements either side of the
limiting value and is considered violated if the constraint exceeds its node it is attached to. The rearward limit for engine eg movement is
limiting value by more than 1%. The convergence criterion used to the leading edge of the wing whereas the forward limit was arbitrarily
terminate the optimisation is activated when a change of less than set to 5 m in front of the quarter chord. However, the vertical move-
0-25% in W between two consecutive iterations occurs. Sensitivity ment is restricted to between 1-5 m above and below the
analysis is explained in Sections 3.1 and 3.2. geometric elastic axis.
The use of a strategy improves efficiency since objective function Design variables may cover a range of orders of magnitudes
and constraints values are obtained using an approximation. This is depending on the differing variable types. For example, variables
important, because at the conceptual design stage the designer must such as layer thicknesses are of the order of 10~3 m to \0~* m and
be able to explore many wing configurations as quickly as possible. variables such as engine position are of the order of 1 m to 10 m.
Table 1 compares the final optimum designs obtained with and This may result in numerical problems, since the optimisation soft-
without using a strategy for the optimisation problem described in ware can have difficulty in optimising problems with sensitivities
Section 4.3 (note that CPU times were obtained whilst running the and design variables which can vary by several orders of magni-
problem on a Silicon Graphic Indy 150 MHz workstation). As can tudes. Hence, all the variables are normalised with respect to their
be seen, the use of a strategy allows a significant reductions in CPU initial value so that they all have a starting value of 1. The upper and
time whilst a lower optimum mass is obtained. The CPU time saved lower bounds on the design variables are also modified to account
is largely due to the reduction in constraint evaluation time for the for this large difference in numerical values.
reason previously stated and it more than compensates for the
increase in time used by DOT due to SQP.
The design variables used during optimisation are split into three 3.1 Sensitivity analysis
groups: front and rear spar position (four variables), engine position Since the normal modes used in the flutter analysis are found by
(three variables) and laminate layer thickness. The total number of solving a transcendental eigenvalue problem, flutter sensitivities
design variables can vary between 16 and 43, depending on whether cannot be easily determined analytically. Hence these, along with
the wing is of metal or composite construction. Laminate thickness sensitivities for divergence speed, material strength and objective
distribution is modelled using a quadratic function to define the function, are found numerically using forward finite differencing, i.e.
thickness of each laminate layer along the span. The layer thickness
design variables are taken as the thickness of layers in three ele-
ments, and the thickness of layers for other elements is determined 3G,. _ Gn'-G„
dXj ax j ••• (24)

forj = l,2,...,nDV andn = l,2,...nCON

where G„' is the constraint value with the normalised variable Xj per-
turbed by an amount cu^, nDV is the total number of design variables and
nCON is the number of design constraints (the sensitivities of the objec-
tive function are found with the terms G„' and G„ in
Equation (24) replaced by W ' and W, respectively). For the problems
presented later, a value of a equal to 0-005 is used for all design
variables and the appendix gives the basis for the selection of this value.
The calculation of the divergence speed, Hoffman Failure Index
and objective function is computationally inexpensive and so the
perturbed values of these are found by performing complete
re-analysis of the perturbed problem. However, the calculation of
perturbed flutter speed in this way would be time consuming.
Figure 2. Coordinate system and notation for beam element. Therefore, a quick method has been developed as described below.
82 T H E AERONAUTICAL JOURNAL FEBRUARY 1997

Aco, = 0 - 25 x 0 • 7 ( ' _1) rads"1 . . . (27)

where £ is the iteration number. Hence, as the solution is refined, the


points selected for use in the interpolation technique are closer to V,
and (Of, which means that the parabolas give better approximations
to Vf and (af. The process is repeated until, for three consecutive
iterations, the difference in the current Vf and (fly, compared to those
given by the preceding iteration, is less than the user specified
accuracy on both Vf and tfy. If this termination criterion is not satis-
fied within 20 iterations then the V-g method is used to calculate
flutter speed to the user specified accuracy level. Typically, the
process converges within 10 (interpolation) iterations. In addition the
checks described below are imposed on the flutter speed and flutter
frequency refinement and if any of these are violated, the V-g
method is used to calculate the flutter speed to any user-specified
accuracy. If the flutter speed cannot be found using the V-g method
Unperturbed Parabolas then as a final resort the determinant method is used to calculate the
— — Perturbed Parabolas flutter speed.
The determinant interpolation technique assumes that the parabo-
las constructed to obtain Vf and V/ permit accurate extrapolation to
give a good approximation of both Vf and Vf. Hence to check that
Figure 3. Illustration of the determinant interpolation technique.
this is the case, two tests, similar to those outlined by Butler and
Williams'25', are carried out. The first test ensures that the gradient at
3.2. Calculation of flutter speed using a determinant inter- the parabola intersection point, e.g. slope 5) of Fig. 3 is within 10%
polation technique of the gradient at the mid-point, e.g. slope 5 2 . The second test checks
The method, which is illustrated in Fig. 3, re-uses both real and that the maximum vertical distance between a parabola and a chord
imaginary (flutter) determinant data obtained for the unperturbed through the two outermost points used in its construction, e.g. the y
design, to analyse the perturbed design. The condition for flutter, of Fig. 3, is less than 1% of the vertical distance between the two
from Equation (6) can be expressed as outermost points. If either test is violated whilst calculating the
unperturbed parabolas then the V-g method is used to calculate the
|[F(V,a))]| = 0 ...(25) perturbed flutter speed for all design variables. Alternatively, if
either test is violated when placing the parabola through points 1', 2'
Here I[F(V,Q))]I is a complex function of the two eigenvalues, V and and 3', the V-g method is used to find the perturbed flutter speed for
CO.
only the current design variable.
To obtain the perturbed flutter speed Vf the value of the real and
imaginary parts of this determinant, (l[F(V,co)]lR and l[F(V,co)]l,,
respectively) are obtained for the unperturbed design at a single
value of frequency (U3f - A(af) and for three values of airspeed (Vf,
4.0 RESULTS
Vf + AVf, Vf - AVf). Quadratic interpolation is then carried out to 4.1 Sensitivities
find point 1 on Fig. 3, i.e. the point at which the magnitude of the
real and imaginary determinants are approximately equal. This Table 2 compares sensitivity results using the determinant interpola-
process is then repeated at two other values of frequency (coy, (af + tion technique with those using the V-g method for the initial design
Acty) to obtain points 2 and 3. The curve through these points repre- of the example described in Section 4.3. Typically, the CPU time
sents the locus of intersection points of the real and imaginary deter- taken to calculate a perturbed flutter speed on a Sun Sparc 10 (Model
minants of the unperturbed design and where this curve crosses the 41) micro-computer using the determinant interpolation method was
airspeed axis is an estimate of the unperturbed flutter speed V}. A is 2-5 s whilst for the V-g method the time was 96 s. These timings
set to a small value of 1CH, so that the points selected for use during were obtained using an airspeed step size of 2 ms-', an airspeed
interpolation are close to both Vf and Vf. This ensures that the
quadratics produced are good approximations to the behaviour of the Table 2
Flutter sensitivity results for determinant Interpolation technique
determinant near to the flutter solution. Points 1', 2', and 3' of Fig. 3
and V-g method using quadratlcally linked thicknesses
are the corresponding real and imaginary intersection points for the
perturbed design. These are obtained by vertically displacing the
unperturbed real and imaginary parabolas to go through the
perturbed values of l[F(V,co)]lR and l[F(V,co)]l,, respectively, noting Design variable Determinant V-g
that these are calculated at V} for the 3 values of frequency ((of, u>{ + interpolation
AcOy, coy - Acoy). The perturbed flutter speed is then assumed to be x
i 3G,/ax,(xl0-3) dG,/dXf(x 1(H)
where a parabola through points 1', 2' and 3' cuts the airspeed axis.
To reduce the time taken in the calculation of the unperturbed flut- Front spar slope -34 -34
ter speeds and unperturbed flutter frequencies, a modified form of Rear spar intersection -214 -210
the above technique is used. The flutter speed is first found to an '/.V.Jd.l 10 10
accuracy of ±0-1 ms-' and the flutter frequency to an accuracy of 're.o.5 -25 -25
±01 rads-' using the V-g method described in Section 2.3. Then the 'flS.-.lO.IO -3 -3
determinant interpolation technique is used to refine the solution. 'R.S',30.5 -26 -26
(Flutter frequency is estimated by calculating the real and imaginary '.S7C.90.5 -65 -64
parabolas at constant velocity, over a range of three frequencies.) '.VK.-30.I 89 90
Here AVf and Atos are reduced using, 'SK..10.10 -18 -18
'.SK.d.l 34 33
AV}= 0 - 2 5 x 0 . ^ ' - ' W ...(26) engine y position 79 7S
LILLICO, BUTLER, G U O AND BANERJEE AEROELASTIC OPTIMISATION OF COMPOSITE WINGS

Striz and Venkayya*27' Balis-Crema, Mastroddi Calfun


and Coppotelli'28'
membrane ribs shear ribs membrane ribs membrane ribs without with
membrane spars membrane spars shear spars membrane spars caps caps

10-50B 10-50B 6-26B 6-13B 6-66B 666B


26-60T 26-60T 24-75T 24-94T 24-04T 27-58T
(0 55-86B 55-85B 27-57B 37-60B 41-77B 41-77B
(Hz) 7912T 7912T 71-77T 73-27T 7112T 82-74T
125-83B 125-82B 110-35B 103-87B 116-60B 116-60B
134-42T 134-42T 136-26T 136-26T 120-20T 137-89T

mesh 24x9 24x9 24x9 9x6x45 9x15x45 10x4 10x4

"/ 276 276 264 241 233 240 276


(ms->) (210) (245)

CD, - - - 11-8 12-4 10-7 10 6


(Hz) (12-8) (150)

vd 302 - - 263 257 236 261


(ms-i) (209) (240)

Table 3
Comparison of results for the Rudislll and Bhatia wing. B denotes bending, T denotes torsion. Calfun results shown without brackets were
obtained using lifting surface theory and those in brackets were obtained using strip theory

range of ±20 ms-1 and a frequency range of ±20 rads-1 centred on the increase and subsequently both flutter and divergence speeds also
unperturbed values of flutter speed and divergence speed. Note that increase, as expected. The calculated flutter speed for this case using
it is difficult to quantify the exact gain in CPU time because the time L-S theory compares quite well with the result from Ref. 27. The
taken using the latter is dependent upon the range of airspeeds and Calfun flutter speeds for both cases lie within the range of results
frequencies considered and also upon the step size. It can be seen given in Refs 27-28. Comparing the results using strip theory with
that up to an accuracy of three decimal places there is very little dif- those given by L-S theory in Calfun, the strip theory gives lower
ference between the sensitivities calculated by the two methods. flutter and divergence speeds. Since the strip theory is based on 2-D
incompressible flow, it should be less accurate than the L-S theory
which takes into account the effect of 3-D compressible flow.
4.2 Flutter and divergence example However, the lower divergence speed and flutter speed given by
Strip theory would give a greater margin of safety, i.e. a conservative
In order to compare the results of this paper with published results, a
estimate, compared with the L-S theory results and hence its use in
wing model which was presented by Rudisill and Bhatia*26' has been
this paper for the optimisation of high aspect ratio wings would seem
taken as a test example. Results from Calfun are compared with
to be justified.
those obtained by Striz and Venkayya<27' and Balis-Crema et a/<28'
for this model. In the former*27', the effect of different FE modelling
on the flutter speed was investigated whilst in the latter*28' (which
4.3 Optimisation results
also used an FE model) the effect of using different aerodynamic
parameters in the unsteady aerodynamic calculation was studied. To illustrate the use of Calfunopt, a wing based on that of a 40 seat,
The current wing model was as an assembly of 10 beam elements twin-engined regional aircraft, with a maximum take off mass of
and the modes corresponding to the first six natural frequencies were 14 500kg was optimised. It should be noted that this example does
used in the aeroelastic analysis. In order to investigate the effect of not contain all of the practical features of actual wing design. Instead
the spar caps on the flutter and divergence speeds, two cases are con- it has been chosen to prove the methodology and to illustrate the
sidered. In the first case, for simplicity, the spar caps are ignored effects that aeroelastic constraints have on design. In order to
when calculating the torsional rigidity. In the second case the contri- achieve the latter, the design flutter and divergence speeds were
bution of the spar caps to the torsional rigidity is considered by given unrealistically high values for this type of wing.
smearing half the total spar cap cross-sectional area around the wing The wing was unswept and had a semi-span of 12-8 m, a tip chord
box as a constant thickness layer. The first six natural frequencies, of 1-458 m, a root chord of 2-917 m and a thickness to chord ratio
flutter and divergence speeds from Calfun are shown in Table 3 and varying linearly between 12% at the tip and 18-29% at the root. It
compared with' the results given in Refs 27-28. The Calfun natural was modelled using 10 elements with the front spar initially at 20%
frequency results which do not take spar caps into account compare chord and the rear spar initially at 60% chord for each of the ele-
quite well with the results of Ref. 28 and the membrane rib-shear ments. The wing box was of carbon fibre composite construction, the
spar model results of Ref. 27 (Striz and Venkayya suggested that this material properties for which are given in Table 4, with a lay-up in
model is the most realistic). Also, for this model the flutter speed both the spars and skins of [907-3073070°] s . An engine of mass
using L-S theory is very close to that of Ref. 28. When the spar caps 360 kg and polar mass moments of inertia about the x and v axes of
are considered in the calculation of GJ, the torsional frequencies 24-44 kgm2 and 53-148 kgm2, respectively, was mounted on the
T H E AERONAUTICAL JOURNAL FEBRUARY 1997

Property Failure stress (10°Nnr2) Element number

Density kgnr 3 1490 Xr 1337 1 5 10


Longitudinal Young's 151x10°
modulus Nnr 2 Xc 1200 90° 0125 0125 0123
Transverse Young's 10x10° Front -30° 0125 0163 0-203
modulus Nm-2 r> 66
spar 30° 0125 0161 0-230
Shear modulus Nm-2 5-9 x 10° 0° 0-125 0125 0127
Yc 250
Major axis Poisson's 0-3 S 70
ratio
90° 0125 0125 0125
Table 4
Material properties and failure stresses Rear -30° 0-310 0175 0125
spar 30° 0166 017 0125
wing. The initial x, v and z position of the engine were, respectively: 0° 0125 0-125 0125
1-767 m in front of the quarter chord, 3-5 m from the wing root and
0-49 m below the geometric elastic axis. Using quadratic linking the
thickness distribution of each spar and skin layer was determined
from the layer thicknesses in elements 1, 5 and 10 and the initial 90° 0125 0125 0125
thickness of each layer within these elements was 0-8 mm, 0-5 mm Skin -30° 0125 0125 0125
and 0-3 mm respectively. A lower bound of 0-125 mm was placed on
all layer thicknesses. The normal modes corresponding to the first 30° 0125 0150 0125
six natural frequencies were used in the analysis. 0° 1-961 0-880 0125
The design flutter speed, Vfmin, and design divergence speed, Vimin,
were both set to 250-0 ms-1 at an air density of 1-225 kgnr 3 , and the
Table 5
maximum material stresses are given in Table 4. The aerodynamic Values of layer thickness design variables for optimum
pitching moment was calculated using a Cmo value of -0-04 and an design In mm
airspeed of 210 ms-'. Strength constraints are calculated for a 2-5g
manoeuvre case. The initial mass of the wing was 300-9 kg with an near their the lower bound of 0125 mm in all elements. This is bene-
initial flutter speed of 368-0 ms-' and an initial divergence speed of ficial to the aeroelastic behaviour of the wing as it will result in the
330-2 ms-1. The problem had 43 design variables, 162 constraints geometric elastic axis being moved forward at the tip, thus reducing
and optimisation started with all constraints satisfied, aerodynamic coupling. Rear spar layer thickness increases towards
although the strength constraints in some elements were significantly the root, which is expected given the fact that shear flow due to
more critical than the aeroelastic constraints.
torsion and shear force is additive in the rear spar. In Fig. 5(a) the
Figure 4 shows that optimisation took 8 iterations to converge effect of the reduction in shear in the front spar due to the engine
requiring 21 min 11 s, whilst running the problem on a Silicon position is clearly seen in elements 1-3. From Fig. 5(b) it can be
Graphic Indy 150 MHz workstation, during which the overall wing
mass was reduced by 40% to 181-01 kg. The flutter speed and
divergence speed of the optimised wing using strip theory were 1.25
250-2 ms-' and 250-0 ms-', respectively. It is worth noting that L-S
theory gave results above 306-2 ms-' (Mach 0-9) for the optimized
wing but since linear theory is inappropriate at these speeds, the L-S
results are not quoted here.
Table 5 shows that, with the exception of the 0° skin layers, the
thickness of the layers in the spars and skin was reduced from initial 10.25
values. The front spar layers of the optimum design, are at their 0.00
lower bound at the wing root but the ±30° layers increase in thick- -0.25
-0.25
ness towards the wing tip whilst the 90° and 0° layers remain at or
• • • -0.50 j—i—1_ _j—i—i—1_
-0.50
1 2 3 4 5 6 7 8 9 10 1 2 3 4 3 6 7 8 9 10
Element Element
I 1 1 1 I—n—r -
! ' '(b;
.. . . . . .

\\

§|^LHg^
1 I i

0 12 3 4 5 6 7 8 1 2 3 4 3 6 7 8 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Element Element
Design iteration Design iteration
Figure 5. Hoffman Failure Index for optimum design. The symbols
Figure 4. Objective function and aeroelastic constraint histories. have the following meaning: O = 90° layer, 3 =-30° layer,
l n ( b ) 0 = V,andU = V& A = 30° layer and V = 0° layer.
LILLICO, BUTLER, G U O AND BANERJEE AEROELASTIC OPTIMISATION OF COMPOSITE WINGS

seen that one of the critical constraints governing the design is stress locations. In addition the wing model should be improved to allow
in the +30° rear spar layer. This is due to this layer being in com- for different upper and lower skin laminate lay-ups and to include
pression along the fibre axis and in tension perpendicular to the fibre skin-stringers.
axis and therefore nearer to failure because of the lower stress limits
in these directions, see Table 4.
The biggest change in layer thickness occurred for the 0° skin
layer, which is significantly thicker than any other layer. However
ACKNOWLEDGEMENTS
the most critical layer in the skins is not the 0° layer but the 90° The authors gratefully acknowledge the support from the Engineer-
layer in the lower skin (see Fig. 5(c)). Here the tensile failure stress ing and Physical Sciences Research Council (EPSRC) under grant
for the 90° layer in the spanwise direction is low, see Table 4, and numbers GR/H82334 and GR/H78467, British Aerospace Airbus
therefore a thick 0° layer is required to limit the transverse tensile Ltd and the School of Mechanical Engineering, University of Bath.
stress in the lower skin 90° layer.
The difference in thickness between the -30° and 30° layers in the
skin results in a small amount of positive bending-torsional
coupling, which produces a wash-out effect when the wing is loaded,
REFERENCES
i.e. the trailing edge of the wing tip twists up (negative \j/ in Fig. 2) 1. NEILL, D.J., JOHNSON, E.H. and CANFIELD, R. Astros — a multidiscipli-
when the wing tip is bent up (negative h). This has been shown by nary automated structural design tool. J Aircr, 1990, 27, (12),
Shirk et aP9\ as well as others, to be beneficial to the divergence pp 1021-1027.
performance of the wing but to degrade the flutter performance. 2. WILKINSON, K., MARKOWITZ, J., LERNER, E., GEORGE, D. and
During optimisation the internal structure is modified significantly BATILL, S.M. Fastop: aflutterand strength optimization program for
lifting-surface structures, J Airer. 1977, 14, (6), pp 581-587.
due to changes in spar position. The front spar moved forward in all 3. DODD, A.J., KADRINKA, K.E., LOIKKANEN, M.J., ROMMEL, B.A.,
elements, to finish at 15% of chord in the tip element and at 1717% SIKES, G.D., STRONG. R.C. and TZONG, T.J. Aeroelastic design
of chord in the root element. The rear spar also moved forward in all optimization program, J Aircr, 1990, 27, (12), pp 1028-1036.
elements so that its final position was 55% of chord in the tip and 4. BARTHOLOMEW, P. and WELLEN, H.K. Computer aided optimization of
41-71% of chord at the wing root. This is a major factor in the reduc- aircraft structures, J Aircr, 1990, 27, (12), pp 1079-1086.
tion of the overall wing mass, since it moves the geometric 5. CORNUAULT, C, PETIAU, C, COIFFIER, B. and PARET, A. Structural
elastic axis, Gs, of the wing forward and also reduces the aerody- optimization of aircraft practice and trends, 72nd Meeting of the Agard
namic coupling. The forward movement of Gs increases the distance Structures and Materials Panel, Bath, 1991, Agard Report 784,
between £ 5 and Gs (i.e. it increases the absolute value of x„). This pp 14.1-14.11.
6. BRAMA, T. The structural optimization system Optsys — current status
trend, which has an adverse effect on flutter, arises as a result of the and applications, 72nd Meeting of the Agard Structures and Materials
mass of 0-35MBOX at 80% chord, which restricts the forward move- Panel, Bath, 1991, Agard Report 784, pp 6.1-6.9.
ment of the mass axis. During optimisation, the position of the 7. MCCULLERS, L.A. and LYNCH, R.W. Dynamic characteristics of
engine changed significantly. The final engine position was 0-391 m advanced filamentary composite structures, Volume II Aeroelastic
in front of the quarter chord, 3-444 m from the wing root and synthesis procedure development, and Volume III Demonstration Com-
0117 m below the elastic axis. The large change in the x coordinate ponent Program, AFDL-TR-111, Sept 1974.
reduces the coupling between bending and torsion modes of the 8. WAKAYAMA, S. and KROO, I. Subsonic wing design using multidiscipli-
wing and reduces the inertia of the engine about the geometric nary optimization, 5th AIAA/NASA/USAF/ISSMO Symposium on
elastic axis. Multidisciplinary Analysis and Optimization, Panama City, Florida,
USA, 1994, AIAA Paper No. 94-4409-CP, pp 1358-1368.
9. AUSTIN, F., HADCOCK, R., HUTCHINGS, D., SHARP, D., TANG, S. and
WATERS, S. Aeroelastic tailoring of advanced composite lifting surfaces
5.0 CONCLUSIONS in preliminary design, 17th AIAA/ASME/SAE Structures, Structural
Dynamics and Materials Conference, 1976, pp 69-79.
This paper describes the theory and application of the multidiscipli- 10. LILLICO, M., BUTLER, R., BANERJEE, J.R. and Guo, S. Aeroelastic
nary optimisation program Calfunopt. Comparing the modal and optimization of high aspect wings using an exact dynamic matrix
flutter analysis results with published results shows that accurate nat- method, 5th AIAA/NASA/USAF/ISSMO Symposium on Multidiscipli-
nary Analysis and Optimization, Panama City, Florida, 1994, AIAA
ural frequencies can be obtained using the dynamic stiffness method. Paper 94-4401-CP, pp 1301-1309.
It also shows that strip theory, which, because it requires less CPU 11. BUTLER, R., LILLICO, M., BANERJEE, J.R., and Guo, S. Optimum design of
time, is used during optimisation, gives flutter speeds which are high aspect ratio wings subject to aeroelastic constraints, 36th
about 15% lower than those found using L-S theory. AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and
The program has successfully been used to optimise a high aspect Materials Conference, New Orleans, Louisiana, 1995, AIAA-95-1223-CP,
ratio, composite wing with constraints on the stress in the wing pp 558-566.
structure due to a 2-5g manoeuvre and on the aeroelastic perfor- 12. BANERJEE, J.R. Use and capability of Calfun — A program for calcula-
mance of the wing. The problem illustrates the suitability of the tion offlutterspeed using normal modes, Proceedings of the Interna-
dynamic stiffness method at the conceptual design stage. tional AMSE Conference on Modelling and Simulation, Athens,
Greece, June 1984, pp 121 -131.
13. BERDICHEVSKY, V„ ARMANIOS, E.A. and BADIR, A. Theory of
anisotropic thin-walled closed-section beams, Comp Eng, 1992, 2,
6.0 FUTURE WORK (5-7), pp 411-432.
14. ARMANIOS, E.A. and BADIR, A.M. Free vibration analysis of anisotropic
For the program to be useful to the designer, a number of short-com- thin-walled closed-section beams, 35th AIAA/ASME/ASCE/AHS/ASC
ings need to be addressed. Whilst constraints on stress in the wing Structures, Structural Dynamics and Materials Conference, Hilton
have been added, this is only for one loading which is not necessari- Head, South Carolina, 1994, AIAA Paper 94-1327-CP, pp 164-171.
ly the worst case. Additional load cases such as those arising from 15. BANERJEE, J.R. Coupled bending-torsional dynamic stiffness matrix for
gusts, turbulence and ground cases should be considered. Also, for beam elements, Int J Numer Meth Eng, 1989, 28, pp 1283-1298.
structures subject to shear or compressive loading, buckling is often 16. BANERJEE, J.R. and WILLIAMS, F.W. Coupled bending-torsional dynam-
ic stiffness matrix for Timoshenko beam elements, Comput Struct,
a governing criterion and therefore needs to be added to the program 1992, 42, (3), pp 301-310.
as an additional set of constraints. Other constraints which should be 17. BANERJEE, J.R. and WILLIAMS, F.W. Free vibration of composite beams
considered include laminate stability, design for ease of manufacture — an exact method using symbolic computation, J Aircr, 1995, 32, (3),
and practical limitation on the movement of the front and rear spar pp 636-642.
86 T H E AERONAUTICAL JOURNAL

18. WITTRICK, W. H. and WILLIAMS, F.W. A general algorithm for comput- balance these
ing natural frequencies of elastic structures, Q J Mech Appl Math, 1971, which requires
24. pp 263-284. used
19. THEODORSEN, T. General Theory of Instability and Mechanisms of
Flutter, Naca Tech Report 496, 1934.
20. BlSPLINGHOFF, B.L., ASHLEY, H. and HALFMAN, R.L. Aeroelasilicity,
Addison-Wesley, 1955.
21. DAVIES, D.E. Theoretical Determination of Subsonic Oscillatory Air-
force Coefficients, Aeronautical Research Council, R&M 3804, 1976.
22. DOWELL, E.H., CURTISS, H.C., SCANLAN, R.H. and SISTO, F. A Modern
Course in Aeroelasticity, Sijthoff and Noordhoff, 1980.
23. Vanderplaats Research & Development, Dot User manual Version 4.20, 1995.
24. BERKES, U.L. Efficient optimization of aircraft structures with a large
number of design variables, J Aircr, 1990, 7, (12), pp 1073-1078.
25. BUTLER, R. and WILLIAMS, F.W. Optimum design using Viconopt, a
buckling and strength constraint program for prismatic assemblies of
anisotropic plates, Comput Struct, 1992, 43, (4), pp 699-708.
26. RUDISILL, C.S. and BHATIA, K.G. Second derivatives of dieflutterve-
locity and die optimization of aircraft structures, AIAA J, 1972, 10, (2),
pp 1569-1572.
27. STRIZ, A.G. and VENKAYYA, V.B. Multidisciplinary optimization stud-
ies using Astros, 72nd Meeting of die Agard Structures and Materials
Panel, Baui, 1991, Agard Report 784, pp 8.1- 8.29.
28. BALIS-CREMA, L. MASTRODDI, F. and COPPOTELLI, G. Structural
modelling effects on aeroelastic analysis. International Forum on Aero-
elasticity and Structural Dynamics, Manchester, 1995, pp 40.1- 40.11.
29. SHIRK, M.H., HERTZ, T.J. and WEISSHAAR, T.A. Aeroelastic tailoring —
theory, practice and promise, J Aircr, 1986, 23, (1), pp 6-18.
30. HAFTKA, R.T and GURDAL, Z. Elements of Structural Optimization, 3rd
Edition, Kluwer, The Netherlands, 1992, pp 256-263.

APPENDIX: ACCURACY AND STEP SIZE


SELECTION
Using Equation (24), the absolute error P in the flutter and
divergence speed sensitivities, can be expressed as

G„ -G,
£±P ...(Al)
ax,
v i J
Substituting for G„' and Gn using Equations (21) and (22) (note that
only the expressions for flutter speed are given, but those for diver-
gence speed may be easily obtained by replacing Vf, Vf and Vlmjn
with Vd, Vf and Vdmin , respectively)

v v
f- f <±P ...(A2)

If the flutter speed and divergence speed are found to an absolute


accuracy of a then the maximum error in the calculation of (Vf - Vf)
is ±2o, hence considering only the positive error, Equation (A2) may
be re-written as

V
fminaxj

Considering the sensitivities for the quadratically linked thickness


design variables, see Table 1, the maximum permissible error (i was
limited to 1 x 10- 3 . Substituting values of Vimin = 250 ms-' and Xj = 1
into Equation (A3) gives,

o < - ...(A4)
8

Care must be taken in the selection of a and a . If a is large then the


truncation error in Equation (24) becomes unacceptably large, see
Ref. 30. But if a is small then to maintain the sensitivity accuracy o
will need to be small. This will increase the computation time and a
very low o may be beyond the precision of the computer. Hence to

You might also like