Professional Documents
Culture Documents
net/publication/245425104
CITATIONS READS
159 409
3 authors:
Melike Nikbay
Istanbul Technical University
91 PUBLICATIONS 644 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Computational Aeroelasticity Investigations For NATO-RTO-AVT-203 and Aeroelastic Prediction Workshop View project
Turbulence And The Aerodynamic Optimisation Of Nonplanar Lifting Systems View project
All content following this page was uploaded by Melike Nikbay on 23 October 2014.
We consider the problem of optimizing for steady-state conditions a given nonlinear aeroelastic system, by
varying both aerodynamic and structural parameters. We model the structure by nite elements and predict
the aerodynamic loads by a three-dimensional nite volume approximation of the Euler equations. We present
a complete optimization methodology whose key components are a computer aided geometric design method for
representing the design, an analytical approach for sensitivity analysis, a gradient-based optimization algorithm
and two staggered schemes for evaluating the aeroelastic responses and solving the coupled sensitivity equations.
We illustrate our methodology and demonstrate its potential with various aeroelastic optimizations of idealized
and builtup wing structures.
T ODAY, the design of complex aeronautical systems remains a ever, such an approach can induce relative errors for the aero-
challenging task because it must account for several nonlinear elastic sensitivities that exceed 25% (Ref. 10) and, therefore, is not
uid, structure,and thermal couplingeffects. Although for such sys- a reliable methodology.
tems intuition, experience, and engineering skills still dominate the The shortcomings of nite differencing for sensitivity analysis
design process, numerical simulation offers an important potential are usually overcome by an analytical evaluation of the sought-after
for speeding up the design cycle. gradients. However, for multidisciplinary problems in general and
In general, a design problemis de ned in terms of physicaldesign aeroelasticity in particular, the analytical approach imposes special
parameters, objectives to be optimized, and constraints to be satis- requirements on the uid and structural analyzers and calls for the
ed. Typically, the objectives and constraints are functions of the solution of yet another complex system of coupled uid/structure
system’s response,which in turn is governedby a set of coupled par- equations. Ghattas and Li11 and MÀ ller and Lund12 have addressed
tial differential equations that describe the various physical effects this issue in the two-dimensionalcase, and Maute et al.,13 as well as
and their coupling. However, because of complexity and compu- Hou and Satyanarayana,14 have addressedit in the three-dimensional
tational cost issues, most often the coupling effects are neglected context.
when developing an optimization method or applying it to a practi- In this paper, we build on the preliminarywork exposed in Ref. 13
cal system. For example, in conventionalstructuraloptimization,the and present a methodology for optimizing realistic aeroelastic sys-
dependence of a uid-induced structural load on a given optimiza- tems thatis based on coupledanalyticalsensitivities.We considerthe
tion criterionis typicallyignored, and in conventionalaerodynamics case where the structure is represented by a detailed nite element
optimization,the structureis assumed to be rigid and the various op- model and the uid by the three-dimensional Euler equations. Our
timization criteria cover only the ow characteristics. For a survey overall methodologyis based on a modular framework for formulat-
of structural optimization, see, for example, Kirsch1 and BendsÀe2 ; ing and solving the target optimizationproblem. Its key components
and for a survey of aerodynamic optimization, see Newman et al.3 are a comprehensivethree- eld formulationof an aeroelasticsystem
and Jameson.4 Nevertheless, because the overall performance of an and fast parallel staggered algorithms for evaluating the aeroelastic
aeronautical system is governed in many cases by coupling effects, responses and computing the analytically derived gradients of the
multidisciplinary optimization methods in general, and aeroelas- optimization criteria.
tic optimization algorithms in particular, have ourished in recent The remainder of this paper is organized as follows. In Sec. II, we
years. formulate the aeroelastic optimization problem and present a three-
For practical reasons, simple structural models and linear ow level modular approach for solving it. In Sec. III, we overview the
theories have dominated the eld of aeroelastic optimization (for three- eld representationof the aeroelastic problem and summarize
example, see Refs. 5– 8). However, with the availability of ever the numerical procedure adopted for evaluating the design criteria.
faster computing platforms and the advent of parallel process- In Sec. IV, we describe the global sensitivity equations for the aero-
ing, detailed nite element structural models and nonlinear ow elastic problem described by the three- eld formulation and derive
theories have made their way into the aeroelastic optimization the corresponding direct and adjoint solution approaches. We also
process. For example, Giunta and Sobieszczanski-Sobieski9 have present and discuss a staggered procedure for computing the aeroe-
performed the aeroelastic optimization of a supersonic aircraft lastic sensitivities associated with the direct approach. In Sec. V,
design by coupling a linear nite element code for the struc- we investigate the accuracy and performance of the proposed sen-
ture and an Euler code for the uid. However, their approach is sitivity analysis and illustrate our methodology with the aeroelastic
computationally intensive because it relies on a nite difference optimization for steady-state conditions of at and builtup wing
scheme for evaluating the gradients of the optimization criteria. structures. We consider optimization criteria and optimization vari-
To lower the computational requirements, Giunta10 introduced a ables that span lift, drag, structural weight, structural displacements
and stresses, orientation of the bers in composite structural ele-
ments, as well as the shape of the dry and wet surfaces.
Received 30 November 2000; revision received 20 April 2001; accepted
for publication 31 May 2001. Copyright ° c 2001 by the American Institute
of Aeronautics and Astronautics, Inc. All rights reserved.
II. Aeroelastic Optimization
¤ Assistant Professor, Department for Aerospace Engineering Sciences A generic optimization problem associated with a given system
and Center for Aerospace Structures. can be formulated as
† Research Assistant, Department for Aerospace EngineeringSciences and
g.s/ ¸ 0; g 2 Rn g (3) (1– 4) is in fact the saddle-point of the preceding Lagrangian and,
therefore, is the solution of the following Kuhn– Tucker conditions:
s D fs 2 Rn s js L · s · sU g (4)
@L @z @h @g
D C ´t C °t D0 (8)
where s is a set of n s abstract parameters restricted by lower and @s @s @s @s
upper bounds s L and sU , z is a cost function of interest, h is a set of @L
n h equality constraints, and g is a set of n g inequality constraints. DhD0 (9)
@´
The physical design parameters can be de ned as functions of
the abstract optimization variables s. For aeroelastic optimization @L
problems, the structural and aerodynamic design parameters can ! ° t g D 0; ° ¸0 (10)
@°
cover, among others, the cross-sectional and thickness dimensions
of the structural elements, the shape of the dry and wet surfaces, and An SQP method for solving these equations can be described as an
the Mach number and angles of attack. extended Newton– Raphson algorithm that incurs at each step the
The optimizationcriteriaq, that is, the objectivefunction z and the solution of a system of equations of the form
constraintsh and g, can cover the aerodynamic performance factors 2 3
such as the lift and drag, as well as the structural behavior descrip- @2 L
.k/
@g .k/ @h .k/
tors such as the displacements, stresses, and strains. In general, the
2 3
6 7 @ L .k/
optimization criteria q depend on the aeroelastic response of the 6 @s2 @s @s 7 2 .k/
3
6 7 1s 6 @s 7
system characterized by the structural displacement vector u and 6 t @g .k/ 7 4 .k/ 5 6 7
6° .k/ g.k/ 0 7 1° D ¡6 .k/ t .k/ 7 (11)
the uid state vector w, which in turn are functions of the physical 6 @s 7 4° g 5
design parameters, or rather the abstract optimization variables s. 6 7 1´ .k/
4 .k/ 5 h.k/
Therefore, @h
0 0
@s
q D q.s; u; w/ (5)
where
u D u.s/; w D w.s/ (6)
1s.k/ D s.k C 1/ ¡ s.k/ ; 1° .k/ D ° .k C 1/ ¡ ° .k/
In this work, we follow the nested analysis and design approach and 1´ .k/ D ´ .k C 1/ ¡ ´ .k/ (12)
assume that u and w always satisfy the aeroelastic state equations.
For this reason, we do not include the aeroelastic state equations
and the superscript.k/ denotesthe iterationindex in the optimization
in the set of equality constraints (2), but determine the structural
loop. The solution of the system of linear equations (11) is also the
displacements u and uid state variables w at each iteration of the
solution of the quadratic minimization problem:
optimization process.
The aeroelasticoptimizationproblemjust describedcan be solved .k/
1 @2 L @ z .k/ .k/
by combining three different numerical models, namely, the op- min 1s.k/ T 2 1s.k/ C 1s (13)
timization, design, and analysis models (for example, see Maute s 2 @s @s
et al.15 ). @g .k/ .k/
1) In the optimization model, the generic problem (1 – 4) is solved 1s C g.k/ ¸ 0 (14)
@s
by a numerical optimization algorithm (for example, see Gill et al.16
and Vanderplaats17 ). Here, we choose for this purpose a sequential @h .k/ .k/
quadratic programming (SQP) algorithm that has demonstrated ro- 1s C h.k/ D 0 (15)
@s
bustnessand ef ciencyfor a broadrange of optimizationproblems.18
We overview this algorithm in Sec. II.A but refer the reader to and, therefore, can be ef ciently obtained by a quadratic program-
Schittkowski19 and Gill et al.,20 among others, for further details. ming method (for example, see Gill et al.20 ) Note that the preceding
2) In the design model, a shape, or structural, or material, or quadratic subproblem is formulated and solved in each iteration
aerodynamic design parameter is de ned as a function of a set of step.
abstract optimization variables. The rationale for such an approach The evaluation of the second derivatives of the optimization
is outlined in Sec. II.B. criteria being a complex operation, the Hessian of L is usually
3) In the analysis model, the optimization criteria are evaluated. approximated by a rst-order information, for example, by the
Because these depend in general on the structural and uid state Davidon– Fletcher– Powell (DFP) or Broyden– Fletcher– Goldfarb–
variables u and w, each design in the optimization process u and w Shanno (BFGS) method (see Ref. 21). However, such an approxi-
calls for the solution of the coupled aeroelastic state equations. The mation may lead to an incorrect increment 1. ¢ /. For this reason,
sensitivity analysis is also assigned to this model because it, too, re- the following correction is introduced:
quires the solution of coupled uid/structure equations. Aeroelastic 2 3 2 3 2 3
analysis and sensitivity analyses are discussed in Secs. III and IV, s.k C 1/ s.k/ 1s.k/
respectively. 6 .k C 1/ 7 6 .k/ 7 6 .k/ 7
4° 5 D 4 ° 5 C ® .k/ 4 1° 5 (16)
.k C 1/ .k/ .k/
A. Optimization Model ´ ´ 1´
It is well known, at least for single-discipline applications,
that the generic optimization problem (1 – 4) can be solved ef - and the step size ® .k/ is determined by a line search. Because the
ciently by gradient-based methods, which are divided into primal, Lagrangian cannot be used to measure an improvement due to
dual, penalty– barrier, and Lagrange approaches. In recent years, Eq. (16), a merit function is introduced and minimized by the line
Lagrange methods have been shown to be robust and ef cient for a search. Reaching a local optimum is guaranteed by following a se-
broad range of problems in structural optimization.18 These meth- quence of decreasing merit function values. The convergence of the
ods convert the constrained optimization problem (1 – 4) into that of optimization process is measured by the L 2 norm of the residual of
nding the extremum point of the following Lagrangian: the Kuhn– Tucker conditions (8 – 10). Note that by construction of
the solution procedure described herein, the constraintsare satis ed
L.s; ´; °/ D z.s/ C ´ t h.s/ C ° t g.s/ (7) only when the optimum point is reached.
where ´ and ° are Lagrange multipliers and, therefore, dual vari- B. Design Model
ables of s, and the t superscript designates the transpose operation. The design model provides an interface between the described
More speci cally, the solution of the original optimization problem optimization model and the analysis model. In general, the relation
MAUTE, NIKBAY, AND FARHAT 2053
where 8 j is a shape function, XO j the vector of control nodes, and » F D F2 .x; w/ (28)
the elemental coordinate system. Frequently used design elements
are Lagrange and Bézier curves as well as B-spline and NURBS for where F2 denotes the vector of arbitrary Lagrangian Eulerian Roe
edges; surfaces can be described by Lagrange, Bézier, NURBS, or uxes resulting from a second-order nite volume discretizationof
Coons patches. The variationof the coordinatesof the control nodes the Euler ow equations.
are used to vary the shape of the body, that is, si D xOk D 1 XO k , where The three- eld formulation outlined earlier leads to a computa-
xOk is a displacement of a control point in an x, y, or z direction. The tional strategy that is in general 25% more computationally expen-
variation x D 1X of the mesh nodes is given by sive than a comparable computational method based on a two- eld
X formulation (u; w) of the aeroelastic problem. However, most if not
xD 8 j .» /Ox j (21) all two- eld formulations of computational aeroelasticity assume,
j among other things, a very small displacement eld for the structure
and, therefore, are restricted to few applications.
The design element concept can also be applied to describe parame- The three-way coupled system of Eqs. (22– 24) can be solved
ter distributions and their variation in the optimization process.27;28 ef ciently by an iterative staggered procedure that allows the usage
The mechanical approach has the advantage that no additional of three different solvers, each tailored to each differentsubproblem
discretizationbesides the structure or the ow domain is necessary. (for example, see Farhat and Lesoinne31 ). In this work, we employ
However, it can be computationally expensive because, for each the second-orderstaggeredalgorithmdescribednext and graphically
optimization variable, the response of a ctitious structure must be shown in Fig. 1, where the superscript .n/ denotes the iteration
determinedand the variableshape can hardlybe controlled.Restrict- number.
ing shape variations is often necessary to guarantee smoothness or 1) For a given externalload P.n/ , determine the structuralresponse
to meet geometric constraints. On the other hand, shape variations u by solving Eq. (25). Then, for numerical stability purposes,
.n/
can be easily controlled in the geometrical approach by restricting perform the following underrelaxation:
the motion of the control nodes. However, in that case an additional
discretization model must be generated for the design elements. u.n/ D .1 ¡ µ /u.n ¡ 1/ C µ uQ (29)
2054 MAUTE, NIKBAY, AND FARHAT
where The partial derivatives @q j =@si , @q j =@u, @q j =@x, and @q j =@w can
@F1 be directly evaluated within the discretized structural and uid
H1 D (36) models by taking into account the relation between the struc-
@w
tural/aerodynamic design parameters and the abstract optimization
is the Jacobian of the rst-order Roe ux. Here, the usage of H1 variables de ned in the design model. To determine the deriva-
rather than H2 follows the rationale of a defect-correction method tives du=dsi , dx=dsi , and dw=dsi of the aeroelastic response with
(even if we advocate in this context a single Newton– Raphson subit- respect to the abstract variable si , the governing aeroelastic equa-
eration) and is motivated by CPU considerations.As usual, we solve tions (22– 24) have to be differentiated,which yields
the steady-state Eq. (24) by the homotopy approach
£¡ ¢¯ ¤ 2 3 2 32 3
A w ¡ w.n/ 1¿ .n/ C F2 .w/ D 0 (37) @S @S @S @S du
6 @si 7 6 @u @x @w 7 6 dsi 7
where A is the matrix of cell volumes and 1¿ .n/ is the pseudo time 6 7 6 76 7
6 @ D 7 6 @D @D 7 6 dx 7
step and, therefore, Eq. (35) is replaced in practice by 6 7C6 76 7
0 76 7 D 0 (44)
6 @si 7 6 @u
£¡ ¯ ¢ ¤¡ ¢ ¡ ¢ 6 7 6 @x 7 6 dsi 7
1 1¿ .n/
A C H1 w.n C 1/ ¡ w .n/
D ¡F2 w .n/
(38) 4 @F 5 4 5 4 dw 5
@F @F
0
5) Extract the uid pressure p.n C 1/ on the uid/structure interface @si @x @w dsi
| {z }
C 1/
0F=S , compute the aerodynamic load P.n F by integrating the uid A
pressure on 0F=S , and transfer the following load to the structure:
P.n C 1/
D T p P.n C 1/
(39) Again, @S =@si and @F =@ si can be directly evaluated within the dis-
T F
cretized structural and uid models by taking into account the re-
where T p is a transformation matrix that accounts for potentially lation between the structural/aerodynamic design parameters and
nonmatching uid and structure meshes. Here, T p is constructed the abstract optimization variables de ned in the design model.
according to the conservative load transfer algorithm described by In the sequel, we denote by A the Jacobian of the aeroelastic
Maman and Farhat32 and, therefore, satis es problem.
To determine the gradients of the design criterion q j with respect
T p D Tut (40) to the abstract optimization variable si , the governing equations
MAUTE, NIKBAY, AND FARHAT 2055
for the gradients of the aeroelastic response (44) are manipulated and @F2 =@x to reduce computational complexity, by replacing the
together with Eq. (43) to obtain mesh motion equation (23), as well as its effects by a transpira-
2 t 3 2 3 tion boundary condition, and reformulating the aeroelastic sensi-
@q j @S tivity problem as a standard two- eld .du=dsi ; dw=dsi / problem.36
6 @u 7 6 @si 7 The consequences of a related approach are investigated in Sec. V.
6 7 6 7
dq j @q j 6 @q 7 6 7 Alternatively, we propose next another block Gauss– Seidel-like
6 j7 ¡1 6 @ D 7
D ¡6 7 A 6 7 (45) staggered solution procedure for solving the complete GSE (44)
dsi @si 6 @x 7 6 @si 7 where @PT =@x and @F2 =@x need not be explicitly computed and
4 5 4 @F 5
@q j stored. Indeed, in the iterative staggered algorithm to be described,
@w @si these quantities appear only in a matrix– vector product of the form
A[du=dsi dx=dsi dw=dsi ]t . More speci cally, @P T =@x appearsand
Equation (45) suggests two alternatives for computing the sought- is transformed in such a matrix– vector product as follows:
after sensitivities.In the direct approach, the derivativesof the aero-
elastic response are rst computed for each abstract optimization du @P T dx @PT dw du @ PT
K ¡ ¡ DK ¡ (50)
variable si as follows: dsi @x dsi @w dsi dsi @si
2 3 2 3
du @S and @ F2 =@x appears in such a matrix– vector product as follows:
6 dsi 7 6 @si 7 @ F2 dx dw
6 7 6 7 C H2 (51)
6 dx 7 6 7
6 7 D ¡A¡1 6 @D 7 (46) @x dsi dsi
6 ds 7 6 @si 7
6 i7 6 7 with the advantage that @P T =@x £ dx=dsi , @P T =@w £ dw=dsi , and
4 dw 5 4 @F 5
@F2 =@x £ dx=dsi can be computed analytically in the discrete uid
dsi @si model.
Let u denote the displacement of the structure at equilibrium,
and then substituted in Eq. (43). However, if the number of opti- and as before, the superscript .n/ denote the nth iteration step. The
mization criteria is smaller than the number of abstract optimization staggeredalgorithm we propose for solving the three- eld GSE (44)
variables, it is more computationally ef cient to adopt the adjoint with A given by Eq. (49) is as follows:
method where the adjoint solutions 1) For a given derivative @P.n/
T =@si , compute the corresponding
a j D fau ; ax ; aw g j derivativeof the structural displacementby differentiatingEqs. (22)
and (25) and applyingan underrelaxationto the obtainedresult. This
are computed rst for each optimization criterion q j as follows: gives
2 3 du.n/ du.n ¡ 1/ duN
@q j D .1 ¡ µ / Cµ (52)
2 3 6 @u 7 dsi dsi dsi
au 6 7
4 ax 5 D A¡T 6 @q j 7 where uN is obtained from the solution of
6 7 (47)
6 @x 7
aw j 4 5 duN @P0 @P.n/
T @K
@q j K D C ¡ u (53)
dsi @si @si @ si
@w
2) Compute
Then, these adjoint solutions are injected in Eq. (45) to obtain
2 3 du.n/ du.n/
@S T
D Tu (54)
3t 6 2 @si 7 dsi dsi
au 6 7
@q j @q j 4 5 6 @D 7 3) When by use of Eq. (30). Equation (31) is used and it is noted
D ¡ ax 6 7 (48) that, by construction, the ctitious stiffness matrix KN does not de-
@si @si 6 @si 7
aw j 4 5 pend on the abstract optimization variables si , compute the deriva-
@F
tive of the uid mesh motion dx.n/Ä =dsi by solving
@si
.n/ .n/
Hence, whether the direct or adjoint approach is selected, a coupled dx dx
KN ÄÄ Ä D ¡KN Ä0 0 (55)
system of linear equations has to be solved. Because of space lim- dsi dsi
itations, we discuss next only the direct approach. For the adjoint with
approach, we refer the reader to the companion paper34 for some
preliminary results. dx.n/ du.n/
T
0
D ; x.n/ D u.n/
T on 0F =S (56)
dsi dsi
A. Direct Approach
From Eqs. (25– 27) and (32– 35), it follows that The details of this step are similar to those of step 3 of the staggered
2 3 algorithm described in Sec. III.
@ PT @P T 4) By the use of Eq. (38), compute the derivatives of the uid
6 K ¡ ¡ state variables dw.n/ =dsi by solving
6µ @x @w 7 7
6 N ¶ µ ¶ 7
6 K Ä0 Tu KN ÄÄ 0 7 dw.n C 1/ @F2 @F2 dx .n/
(57)
AD6 0 7 (49) H2 D¡ ¡
6 ¡Tu 0 I 7 dsi @si @x dsi
6 7
4 5 C 1/
0
@F2
H2
5) Compute @P.n F =@x and @P.nF C 1/ =@w analytically, evaluate
.n C 1/
@x @P F =@si using the chain rule
C 1/ C 1/ C 1/
where H2 is the Jacobianof the second-orderRoe ux. Constructing @P.n
F @P.n
F d x .n/ @P.n
F dw .n/
a second-order Jacobian of the uxes is not a trivial matter, particu- D C (58)
@si @x dsi @w dsi
larly for nite volume upwind schemes.35 However, the importance
of a second-orderspace-accurateJacobian for sensitivity analysis is and compute
highlighted in Sec. V. C 1/
For sensitivity analysis purposes, it is conceivable to avoid the @ P.n
T @P.n C 1/
D Tp F (59)
computation of the two mesh motion related matrices @P T =@x @si @ si
2056 MAUTE, NIKBAY, AND FARHAT
As in the case of the staggered algorithm proposed for solving the number to M1 D 0:6, and the remaining freestream ow condi-
governing aeroelastic state equations, we monitor the convergence tions to those corresponding to a ight at an altitude of 10,000 m
of the staggered procedure just described by requiring simultane- (½1 D 4:135 £ 10¡1 kg/m3 and p1 D 2:650 £ 104 N/m2 ). We dis-
ously cretize the at plate by 768 three-nodedshell elements and the uid
® ® computational domain by 44,598 grid points (Fig. 2). We label this
® duN .n C 1/ @P0 @P.n C 1/ @K ® speci c uid mesh as mesh A and note that the structure mesh and
®K ¡ ¡ T
C u ®
® dsi @si @si @si ®2 the uid mesh on the wet surface do not match geometrically and
topologically.
® ®
® N .0/ @P0 @P.0/ @K ® In all computations aimed at investigating the effect of mesh
SA ® du
· ² ®K ¡ ¡ T
C u® (60) re nement, we set ² AA D ² SA D 10¡8 . For these computations, we
dsi @ si @si @si ® 2 choose this rather small value for the tolerancesto guard our conclu-
® ® sions against any effect of algebraic misconvergence. On the other
® dw.n C 1/ @F2 @F2 dx .n C 1/ ®
®H 2 C C ® hand, in all computations aimed at assessing the effect on perfor-
® dsi @si @x dsi ® mance of the order of accuracy of the Jacobian H and/or the degree
2
® ® to which the mesh motion is accountedfor in the sensitivityanalysis,
® dw .0/ @F2 @F2 dx .0/ ® we set ² AA D ² SA D 10¡5 .
·² SA ®
®H2 dsi C @ si C @x dsi ®
® (61) First, we compute the sensitivity of each of the drag, lift, and von
2
Mises stress ¾ E on the upper surface of the plate at point E, and of
where ² SA is a speci ed tolerance and SA is sensitivity analysis. vertical displacement of the plate u C at point C, with respect to the
Again, note that only the rst inequality (60) assesses the conver- structuralthickness,s1 , the backsweepof the wing, s2 , the freestream
gence of the proposed staggered procedure. The second inequality Mach number, s3 , and the ® angle of attack, s4 . For controlling the
(61) emphasizes the importance of requiring that the ow solver variation of the backsweep, we adopt the design element concept
converges to the same precision as that imposed on the staggered outlined in Sec. II.B. More speci cally, we describe the geometry
algorithm. of the structural model by a single Coons element A – B – C – D with
Finally, we note that the convergence of the preceding iterative linear shape functions along the edges. We identify varying the
staggered procedure proposed herein for solving the GSE (44) can backsweep with varying the horizontalposition of the control nodes
be accelerated by storing at a given optimization step the gradients C and D and, therefore, de ne the abstract optimization variable s2
of the uid state, and using these to initialize the sensitivity analysis as s2 D xO1C D xO1D . Each value of s2 de nes a unique structural shape
in the following optimization step. Such a technique increases the variation that must be superposed to the elastic displacement eld
memory requirements but accelerates the solution time. u before performing Eq. (33) and solving Eq. (34) to update the
motion x of the uid mesh.
V. Applications and Numerical Investigations We report in Table 1 the obtained aeroelastic sensitivity results.
In thissection,we rst illustratethe aeroelasticsensitivityanalysis Note that all of the sensitivities with respect to the angle of attack
presented in this paper and investigate some interesting computa- ® are given in per radian. The sensitivities in Table 1 re ect well-
tional issues. Then, we highlightthe potentialof the proposedaeroe- known effects, such as increasing the backsweep decreasesthe drag,
lastic optimizationmethodology by applying it to one academic and increasing the Mach number increases the lift, and increasing the
one realistic wing problem. In all cases, we solve all uid subprob- structural thickness decreases the stress level.
lems such as Eqs. (38) and (57) by a parallel generalized minimum
residual (GMRES) algorithm preconditioned by a restricted addi- 1. Effect of Mesh Re nement
tive Schwarz method (see Cai et al.37 ) and all structuralsubproblems For a at plate structure,the nite element model describedearlier
such as (1) and (53) by a direct method. We perform all computa- is suf ciently well resolved.Therefore,we considerhere re ning the
tions in double-precisionarithmetic on a 12-processor Origin 2000 uid mesh only, and construct mesh C with 87,827 grid points. For
system. comparison purposes, we also generate a coarser mesh than mesh
A, namely, mesh B with 19,384 grid points. We report in Table 2
A. Aeroelastic Sensitivity Analysis of an Idealized Wing Structure
the relative variations of the earlier computed sensitivities with the
Besides illustrating the proposed aeroelastic sensitivity method, size of the uid mesh.
our objective here is to investigatethe in uence of the following pa- The results reported in Table 2 show that only the accurate predic-
rameters on accuracy and computational ef ciency, that is, on per- tion of the drag requires a ner mesh than mesh A. The aerodynamic
formance: 1) mesh re nement, 2) order of accuracy of the Jacobian lift, the structuraldisplacement u C , as well as their sensitivitieswith
H used in the sensitivity analysis, and 3) degree to which the effect respect to the thickness, backsweep, and angle of attack ®, are well
of the mesh motion is included in the sensitivity analysis. convergedfor mesh A, and the sensitivitiesof the structuraldisplace-
For this purpose, we consider a cantilevered wing with a ment u C and the von Mises stress ¾ E with respect to the backsweep
NACA0012 airfoil. We idealize the structure as a at plate and and the Mach number are reasonably well converged.
assume that it is made of an isotropic material characterized by
Young’s modulus E D 1:45 £ 1011 N/m2 and Poisson ratio º D 0:2 2. Approximate Sensitivities
and that its thickness is t h D 0:025 m. We set the angles of attack There are two tempting approaches for reducing the computa-
to ® D 2:5 deg (pitch), and ¯ D 0 deg (yaw), the freestream Mach tional complexityof the proposedsensitivityanalysis and, therefore,
Table 2 Relative variations of the aerodynamic and structural criteria as well as their sensitivities
with the size of the uid mesha
Approach 1 drag, % 1 lift, % 1 stress ¾ E , % 1 displacement, u C %
Aeroelastic analysis 51:50 j¡26:49 ¡1:32 j0:43 ¡0:87 j0:50 ¡0:80 j0:40
Sensitivity analysis
optimization variables
Thickness 4:89 j ¡3:95 ¡0:29 j0:54 ¡1:35 j0:48 ¡1:14 j0:39
Backsweep 29:90 j¡18:22 ¡0:00 j0:19 ¡4:60 j1:95 ¡2:26 j1:11
Mach number 40:27 j¡20:54 ¡1:47 j0:34 ¡2:30 j1:66 ¡7:07 j0:57
Angle of attack ® 0:60 j ¡0:91 ¡0:93 j0:77 ¡0:24 j0:91 0:01 j0:83
a Format is mesh Bj mesh C, and reference values are those tabulated in Table 1 for mesh A.
Table 3 Relative errors in the sensitivities associated with the Table 4 Performance of the aeroelastic sensitivity analysis
simpli ed computational strategies using four different computational strategies
Relative errors Approach
in the sensitivities, %
Computational cost SH2 SH1 SH2SM SH1SM
Sensitivity variable SH1 SH2SM SH1SM
Total no. of iterations
Lift with respect to thickness 4:94 14:66 13:29 Staggered procedure (52 – 59) 15 14 14 15
Lift with respect to backsweep 3:05 0:93 67:62 GMRES (57) 317 271 255 262
Displacement with respect to thickness 3:29 6:07 2:48 PCG (55) 246 220 0 0
Displacement with respect to backsweep 12:25 8:80 69:80 CPU time, s
Overall 235 220 143 148
Mesh motion 61 57 0 0
increasing the speed of the optimization process. The rst one con-
sists in employing the rst-order ux Jacobian H1 instead of the Observe that, at least for this problem, using a rst-order rather
second-order one H2 in Eq. (49) because H1 is computationally than second-order ux Jacobianin the sensitivity analysis,or setting
cheaper to evaluate than H2 . The second one consists in setting dxÄ =dsi D 0, results in relative errors in the sensitivitiesthat are less
dxÄ =dsi D 0 on the basis that the mesh motion of the uid grid than 15%. However, combining both approximations leads to an
points that are not located on the uid/structure interface is a math- unacceptable maximum error that exceeds 60%.
ematical artifact of the proposed sensitivity analysis methodology Also observe that, in all cases, the staggered procedure (52– 59)
and perhaps does not affect its performance. Such an assumption converges in 14– 15 outer iterations. Throughout these iterations,
eliminates the need to solve problem (55), which arises at each it- the accumulated number of GMRES iterations for solving the
eration of the staggered procedure proposed for solving the GSE. uid subproblems (57) is slightly higher for the mathematically
Both approaches,which could also be combined,amount to comput- consistent strategy SH2, because this strategy generates more
ing approximate sensitivities rather than sensitivities that are con- ill-conditioned systems. The simpli ed computational strategies
sistent with the three- eld formulation (22– 24) of the aeroelastic SH2SM and SH1SM reduce signi cantly the CPU time, but de-
problem. To investigate the merit of these approaches, we consider liver unacceptable accuracies. Both Tables 3 and 4 show that SH2
again the wing problem described in Sec. V.A and de ne the fol- delivers the best performance/accuracy ratio. The performance re-
lowing four different computational strategies: 1) SH2, using H2 sults summarized in Table 4 also suggestthat, for the mathematically
in Eq. (49) and taking into account dxÄ =dsi ; 2) SH1, using H1 in consistent strategies for computing the gradients, handling the uid
Eq. (49) and taking into account dxÄ =dsi ; 3) SH2SM, using H2 mesh motion accounts for about 26% of the total CPU time, which
in Eq. (49) and ignoring dxÄ =dsi (simpli ed mesh motion); and 4) is a reasonable CPU cost.
SH1SM, using H1 in Eq. (49) and ignoringdxÄ =dsi (simpli ed mesh
motion). B. Aeroelastic Optimization of a Composite Wing
Note that the computational strategy SH2, which is the only one Next, we consideragain the same wing as in Sec. V.A but idealize
among the four that is mathematically consistent with our formula- here the structure as a two-ply composite plate with the rst layer
tion of the aeroelastic problem, is the one used for generating the characterized by a uniform Young modulus E 1 D 1:45 £ 1011 N/m2
results reported in the two preceding subsections. and the second layer by E 2 D 3:28 £ 109 N/m2 . We also assume that
We adopt mesh A and recompute the same aeroelastic sensitivi- both layers have the same shear modulus G D 3:38 £ 109 N/m 2 ,
ties as before using the four different computational strategies just Poisson ratio º12 D 0:2, and effective thickness t h D 0:0275 m.
outlined. Setting the results of the computational strategy SH2 as We discretize the idealized structure by 432 three-noded composite
reference results, we report in Table 3 the relative errors in the sen- shell elements and the uid computational domain by 23,298 grid
sitivities predicted by the other computational strategies. We also points.
summarize in Table 4 the performance results obtained when the We consider an initial design where the ber orientations are
proposed aeroelastic sensitivity procedure is equipped with either Á1 D 0 deg and Á2 D 45 deg with respect to the axis of symmetry
of the four computational strategies. of the plane and clamp again the wing at its root A – B (Fig. 2). We
2058 MAUTE, NIKBAY, AND FARHAT
seek to minimize the drag over lift ratio D = L of this wing for three local angle of attack ®, which in turn leads to a higher induced drag
different freestream Mach numbers M1 D [0:4; 0:8; 1:2] by opti- and, for this problem, an even higher lift.
mizing the orientation of its composite bers. The other freestream
ow conditions are the same as given in Sec. V.A. C. Shape and Thickness Optimization of the ARW2 Wing
We reportin Table 5, for each Mach number, the predictedoptimal Here, we illustrate the computational methodology presented in
ber orientation and the achieved reduction in D =L as a percentage this paper with the aeroelastic optimization of the ARW2 wing,38
of the drag over lift ratio of the initial design. whose geometry and material propertiesare summarized in Table 6.
For each considered Mach number, the optimal ber orientations Unlike in the earlier examples, we consider here a detailed nite
appear to be the same for both layers, which can be expected when element model of the structure (Fig. 4) that includes, among others,
the layers have the same geometry and thickness. In all three cases, the spars, ribs, hinges, and control surfaces of this wing, and that
the drag over lift ratio is reduced by increasing the lift rather than contains a total of 11,838 degrees of freedom. We generate a three-
reducing the drag. Furthermore, Fig. 3 reveals that the optimal ber dimensional unstructured uid mesh around the wet surface with
orientations increase the wing bending and torsional exibilities. 43,847 grid points.
This illustrates the indirect role that composite bers can play in We consider a ight con guration at 12,200 m with ½1 D
optimizing the deformed shape. A larger twisting leads to a higher 2:928 £ 10¡1 kg/m3 , p1 D 1:882 £ 104 N/m2 , and set M 1 D 0:8
Fig. 3 Vertical displacement and local twist angle along the one-quarter chordlength line for M1 = 0:8.
Fig. 4 ARW2 structural, uid, and design models; initial con guration.
MAUTE, NIKBAY, AND FARHAT 2059
and the angles of attack to ® D 2:5 deg and ¯ D 0 deg. Again, we von Mises stress of 8:90 £ 107 N/m2 associated with the initial de-
seek to minimize the drag over lift ratio. However, for this wing sign. Because the stresses are controlled at each nite element node
problem, we introduce the following lift, stress, and displacement for the upper, middle, and lower surfaces of a plate or shell element,
optimization constraints: this leads to 2412 stress constraints.
1) The lift L can change only by an amount greater or equal to We perform the constrained aeroelastic optimization by varying
the variation of the weight induced by the variation of the thickness the backsweep and twisting of the wing (Fig. 5), as well as the
of the stiffeners. We write this as follows: thicknesses of its stiffeners as follows:
1) We controlthe shape of the wing by the designelement concept.
1L ¸ 21W (62) More speci cally, we introduce a Coons element A – B – C – D in the
ground plan x 1 – x2 of the wing (Fig. 4) and introduce the following
where W is the weight of the wing and the factor 2 is to account for bounded optimization variables:
the weight of half of the fuselage and other supporting structures.
2) The vertical displacementsat the pointsA and B (Fig. 4) cannot s1 D xO1A D xO1B ; ¡0:635 · s1 · 0:635 m
exceed u max D 0:381 m. This maximum displacementconstraint can
also be interpreted as a maximum bending exibility constraint.
3) The von Mises stress in the stiffeners cannot exceed ¾max D s2 D xO 2B D ¡ xO2A ; ¡0:0254 · s2 · 0:01016 m
9:05 £ 107 N/m2 . This value of ¾max is suggested by the maximum
2) We subdivide the stiffeners into two groups. In the rst group,
we choose to vary the thickness as a linear function of two op-
timization variables s3 and s4 to account for the initial variations
in the thickness of the stiffeners in this rst group. In the second
group, we identify the thickness variation with a single abstract
optimization variable s5 , that is, we force all of the stiffeners to
have the same thickness variation. To avoid a zero thickness stiff-
ener, we limit s3 , s4 , and s5 by ¡0:002565 · s3 · 0:0127 m and
¡0:000762 · s4 ; s5 · 0:0381 m.
When the SH2 computational strategy de ned in Sec. V.A for
the sensitivity analysis is used, the proposed aeroelastic optimiza-
tion method converges after 20 iterations to the optimal shape
and thickness corrections graphically shown in Fig. 5 and reduces
the drag over lift ratio, that is, the objective function, by a fac-
tor equal to 1.24 (19:44%). Table 7 contrasts some performance
characteristics of the initial and optimized designs of the ARW2
wing. From these results, we conclude that, for this problem, our
constrained aeroelastic optimization reduces the D =L ratio essen-
tially by reducing the drag. It increases the backsweep, which re-
duces wave drag, and adjusts only slightly the twist of the wing.
Figure 6, which compares the Mach contour plots of the initial
and optimized con gurations, also reveals that the optimized con-
Fig. 5 Optimization variables and optimal con guration: backsweep guration features a weaker shock, which is consistent with drag
s1 , 0:49772 m; twist s2 , 0:00347 m; D thickness s3 , ¡ 0:00247 m; D
thickness s4 , 0:00212 m; and D thickness s5 , ¡ 0:00076 m.
Table 7 Aeroelastic characteristics of the initial and optimized
designs of the ARW2 wing
Con guration
Parameter Initial.0/ Optimized
Lift L, N 9:63798 £ 103 8:06121 £ 103
Drag D, N 0:48684 £ 103 0:32764 £ 103
D= L 0:505 0:406
Weight of wing, N 1:78100 £ 103 0:99256 £ 103
Vertical displacement 0:23141 0:25284
at point B m
Maximum von Mises 8:90134 £ 107 9:05000 £ 107
Fig. 6 Steady-state Mach contours for initial and optimized con gu- stress, N/m2 (reached at point E) (reached at point F)
rations.
Fig. 7 Convergence of the objective function using different computational strategies for the sensitivity analysis.
2060 MAUTE, NIKBAY, AND FARHAT
reduction. The reduction of the thicknesses of some stiffeners re- 9 Giunta, A., and Sobieszczanski-Sobieski, J., “Progress Toward Using
sults in a signi cant reduction of the weight of the wing. A com- Sensitivity Derivatives in a High-Fidelity Aaeroelastic Analysis of a Super-
parison of the displacements and maximum stresses of the initial sonic Transport,” Proceedings of the 7th AIAA/USAF/NASA/ISSMO Sympo-
and optimized con gurations also reveals that the optimized wing sium on Multidisciplinary Analysis and Optimization, AIAA, Reston, VA,
1998, pp. 441– 453.
is more exible in bending but is subjected to a slightly higher stress 10 Giunta, A., “Novel Sensitivity Analysis Method for High Fidelity Mul-
level. tidisciplinary Optimization of Aerostructural Systems,” AIAA Paper 2000-
Finally, we show in Fig. 7 the performanceresultsobtainedfor our 0683, Jan. 2000.
optimization methodology and this ARW2 problem, when equipped 11 Ghattas, O., and Li, X., “Domain Decomposition Methods for Sen-
with the four different computationalstrategies de ned in Sec. V.A, sitivity Analysis of a Nonlinear Aeroelastic Problem,” International Jour-
as well as a central nite difference scheme for performing the sen- nal of Computational Fluid Dynamics, Vol. 11, No. 1 – 2, 1998, pp. 113–
sitivity analysis. These results con rm again that using the second- 130.
12 À
order Jacobian of the uxes and taking into account the sensitivity M ller, H., and Lund, E., “Shape Sensitivity Analysis of Strongly Cou-
of the mesh motion with respect to the abstract optimization vari- pled Fluid– Structure Interaction Problems,” AIAA Paper 2000-4823, Sept.
2000.
ables in the computation of the gradients of the optimization criteria 13 Maute, K., Lesoinne, M., and Farhat, C., “Optimization of Aeroelastic
(strategy SH2) is the winning strategy. They also highlight the supe- Systems Using Coupled Analytical Sensitivities,” AIAA Paper 2000-0560,
riority of the analytical approach for computing the gradients over Jan. 2000.
nite differencing. 14 Hou, G.-W., and Satyanarayana, A., “Analytical Sensitivity Analysis of
three- eld formulation of an aeroelastic problem, as well as the Design: With Applications, McGraw – Hill, New York, 1984.
18 Schittkowski,K., Zillober, C., and Zotemantel, R., “Numerical Compar-
fast evaluation, also by a staggered algorithm, of the aeroelastic
response of each intermediate con guration. We have illustrated ison on Nonlinear Programming Algorithms for Structural Optimization,”
Structural Optimization, Vol. 7, 1994, pp. 1 – 28.
this methodology and highlighted its potential with the aeroelastic 19 Schittkowski, K., “NLPQL: A FORTRAN Subroutinefor Solving Con-
optimization of various idealized and builtup wing structures. All strained NonlinearProgramming Problems,” Annalsof OperationsResearch,
obtained performance results highlight the superior performance Vol. 5, No. 6, 1985, pp. 485– 500.
of a mathematically consistent evaluation of the gradients in the 20 Gill, P., Saunders, M., and Murray, W., “SNOPT: An SQP Algorithm
sensitivity analysis, that is, a second-order evaluation of the ux for Large-Scale Constrained Optimization,” Dept. of Mathematics, Rept. NA
Jacobian when the ow is discretized by a second-order space- 97-2, Univ. of California, San Diego, CA, 1997.
21 Luenberger, D., Linear and Nonlinear Programming, Addison Wesley
accurate scheme, and a proper evaluation of the sensitivities with
respect to the uid mesh motion because this mesh motion is an Longman, Reading, MA, 1984.
22 Zienkiewicz, O., and Campbell, J., “Shape Optimization and Sequen-
integral component of the three- eld formulation of an aeroelas-
tial Linear Programming,” Optimum Structural Design, edited by R. H.
tic problem. These results also point to the superior performance
Gallagher and O. C. Zienkiewicz, Wiley, New York, 1973, pp. 109– 126.
of the optimization loop when the gradients are based on the ana- 23 Kikuchi, N., Chung, K., Torigaki, T., and Taylor, J., “Adaptive Finite
lytically derived global sensitivity equations rather than on nite Element Methods for Shape Optimization of Linearly Elastic Structures,”
differences. ComputationalMethods in Applied Mechanical Engineering, Vol. 57, No. 1,
1986, pp. 67– 89.
24 Belegundu, A., and Rajan, S., “A Shape Optimization Approach Based
Acknowledgments
on Natural Design Variables and Shape Functions,” Computational Meth-
The authors acknowledge the partial support by the National ods in Applied Mechanical Engineering, Vol. 66, No. 1, 1988, pp. 87 –
Science Foundation under Grant ECS-9725504 and the partial sup- 106.
port by the Air Force Of ce of Scienti c Research under Grant 25 B öhm, W., Farin, G., and Kahmann, J., “A Survey of Curve and Surface
F49620-99-1-007. They thank K. Schittkowski for providing them Methods in CAGD,” Computer Aided Design, Vol. 1, No. 1, 1984, pp. 1 –
with the source code of the sequential quadratic programming 60.
26 Farin, G., Curves and Surfaces for Computer Aided Geometric Design,
algorithm.
Academic Press, London, 1988.
27 Bennett, J., and Botkin, M., “Structural Shape Optimization with Geo-
References metric Description and Adaptive Mesh Re nement,” AIAA Journal, Vol. 23,
1 Kirsch, U., Structural Optimization: Fundamentals and Applications, No. 3, 1985, pp. 458– 464.
Springer, Berlin, 1993. 28 Bletzinger, K.-U., Kimmich, S., and Ramm, E., “Ef cient Modeling in
2 Bendsøe, M., Optimization of Structural Topology, Shape, and Material,
Shape Optimal Design,” Computing Systems in Engineering, Vol. 2, No. 5 – 6,
Springer, Berlin, 1995. 1991, pp. 483– 495.
3 Newman, P., Hou, G.-W., and Taylor, A., “Observations Regard- 29 Farhat, C., Lesoinne, M., and Maman, N., “Mixed Explicit/Implicit
ing Use of Advanced CFD Analysis, Sensitivity Analysis, and De- Time Integration of Coupled Aeroelastic Problems: Three-Field Formula-
sign Codes in MDO,” Multidisciplinary Design Optimization—State tion, Geometric Conservation and Distributed Solution,”International Jour-
of the Art, edited by N. Alexandrov and Y. Hussaini, Society for nal for Numerical Methods in Fluids, Vol. 21, No. 10, 1995, pp. 807–
Industrial and Applied Mathematics, Philadelphia, 1997, pp. 263– 835.
279. 30
Farhat, C., Degand, C., Koobus, B., and Lesoinne, M., “Torsional
4 Jameson, A., “Re – Engineering the Design Process through Computa-
Springs for Two-Dimensional Dynamic Unstructured Fluid Meshes,” Com-
tion,” AIAA Paper 97-0641, 1997. putational Methods in Applied Mechanical Engineering, Vol. 163, No. 1 – 4,
5 Haftka, R., “Structural Optimization with Aeroelastic Constraints—A
1998, pp. 231– 245.
Survey of U.S. Applications,” International Journal of Vehicle Design, 31
Farhat, C., and Lesoinne, M., “Two Ef cient Staggered Procedures for
Vol. 7, No. 3– 4, 1986, pp. 381– 392. the Serial and Parallel Solution of Three-Dimensional Nonlinear Transient
6 Bowman, K., Grandhi, R., and Eastep, F., “Structural Optimization of
Aeroelastic Problems,” Computational Methods in Applied Mechanical En-
Lifting Surfaces with Divergence and Control Reversal Constraints,” Struc- gineering, Vol. 182, No. 3– 4, 2000, pp. 499– 515.
tural Optimization, Vol. 1, No. 1, 1989, pp. 153– 161. 32 Maman, N., and Farhat, C., “Matching Fluid and Structure Meshes for
7 Friedmann, P., “Helicopter Vibration Reduction Using Structural Opti-
Aeroelastic Computations:A Parallel Approach,” Computers and Structures,
mization with Aeroelastic/MultidisciplinaryConstraints: A Survey,” Journal Vol. 54, No. 4, 1995, pp. 779– 785.
of Aircraft, Vol. 28, No. 1, 1991, pp. 8– 21. 33 Sobieszczanski-Sobieski, J., “Sensitivity of Complex, Internally Cou-
8 Barthelemy, J.-F., Wrenn, G., Dovi, A., and Hall, L., “Super-
pled Systems,” AIAA Journal, Vol. 28, No. 1, 1990, pp. 153– 160.
sonic Transport Wing Minimum Design Integrating Aerodynamics and 34 Maute, K., Nikbay, M., and Farhat, C., “Analytically Based Sensitivity
Structures,” Journal of Aircraft, Vol. 31, No. 2, 1994, pp. 330– Analysis and Optimization of Nonlinear Aeroelastic Systems,” AIAA Paper
338. 2000-4825, Sept. 2000.
MAUTE, NIKBAY, AND FARHAT 2061
35 Lesoinne, M., Sarkis, M., Hetmaniuk, U., and Farhat, C., “A Linearized Tenth InternationalConference on Domain Decomposition Methods for Par-
Method for the Frequency Analysis of Three-Dimensional Fluid/Structure tial Differential Equations, edited by J. Mandel, C. Farhat, and X.-C. Cai,
Interaction Problems in All Flow Regimes,” Computer Methods in Applied American Mathematical Society, Providence, RI, 1998, pp. 479– 485.
Mechanics and Engineering, Vol. 190, No. 24 – 25, 2001, pp. 3121– 3146. 38 Sandford, M., Seidel, D., Eckstrom, C., and Spain, C., “Geometrical and
36 Piperno, S., Farhat, C., Larrouturou, B., “Partitioned Procedures for the Structural Properties of an Aeroelastic Research Wing (ARW-2),” NASA
Transient Solutionof Coupled Aeroelastic Problems,” ComputationalMeth- TM-4110, 1989.
ods in Applied Mechanical Engineering, Vol. 124, No. 1 – 2, 1995, pp. 79 –
112.
37 Cai, X.-C., Farhat, C., and Sarkis, M., “A Minimum Overlap Restricted E. Livne
Additive Schwarz Preconditionerand Applications in 3D Flow Simulations,” Associate Editor