You are on page 1of 11

Airfoil/Wing Optimization

Thomas A. Zang
Systems Analysis and Concepts Directorate, NASA Langley Research Center, Hampton, VA, USA

1 Introduction 2 Optimization Formulation 3 Shape Denition 4 Mesh Generation 5 Gradient Computation 6 Optimization Using CFD 7 Disclaimer Acknowledgments Notes References

1 2 4 7 7 8 10 10 10 10

1 INTRODUCTION
This chapter focuses on techniques for improving the aerodynamic characteristics of an aerospace vehicle through renement of the vehicles shape. Methods for aerodynamic shape optimization have progressed through increasingly complex aerodynamic analysis tools roughly speaking, for linear aerodynamics (e.g., panel methods), transonic small disturbance equations, and full potential equations in the 1970s; for linear aerodynamics with boundary-layer corrections in the 1980s; for Euler equations from the 1980s through the mid1990s; for NavierStokes equations on structured meshes in the 1990s; for Euler and NavierStokes equations on unstructured meshes from the late 1990s onward; and for

time-dependent ows, most recently. Although the discussion in this chapter is conned to external ows, most of the discussion is equally applicable to internal ows, such as those through aircraft engines. Readers interested in summaries of the important developments (and contributors) in this eld should consult review articles such as those by Labruj` ere and Slooff (1993), Newman et al. (1999), Reuther et al. (1999), and Mohammadi and Pironneau (2004), as well as the hundreds of references therein. Figure 1 illustrates the key processes and variables for aerodynamic shape optimization. (The gradient variables are shown in parentheses, as they are only relevant for gradientbased optimization methods.) The Optimization process feeds a set of design variables x (a vector of length d ) to the Shape Denition component, which provides a mathematical description s of the surface. The surface description is used by the Mesh Generation process, which produces the computational mesh y used by the Aerodynamic Analysis component. The analysis output is the state variable q (and its gradient q). The Objective and Constraint Evaluation process supplies the objective function f , the inequality and
x
Shape definition

Optimization

Mesh generation

f, g j, hk (f, g j ,hk ) Objective and constraint evaluation

q
(q )

Aerodynamic analysis

Figure 1. Key processes and variables for aerodynamic shape optimization.

Encyclopedia of Aerospace Engineering, Online 2010 John Wiley & Sons, Ltd. This article is 2010 US Government in the US and 2010 John Wiley & Sons, Ltd in the rest of the world. This article was published in the Encyclopedia of Aerospace Engineering in 2010 by John Wiley & Sons, Ltd. DOI: 10.1002/9780470686652.eae500

2 Aerospace System Optimization equality constraints g and h (perhaps along with their gradients) back to the Optimization process. Those components of the optimization process that are of special importance to aerodynamic shape optimization are the optimization formulation (objectives and constraints), shape denition, mesh generation, and gradient computation (for those methods that utilize gradient information). These topics are covered in that order. The initial expository material on the optimization formulation is given in the context of linear aerodynamics. Thereafter, the emphasis is on the important considerations for aerodynamic shape optimization using the Euler or NavierStokes equations, that is, what are commonly called computational uid dynamics (CFD) tools. The chapter concludes with some considerations for the use of CFD in aerodynamic shape optimization. The discussion is restricted to use of optimization for improving an existing design, that is, optimization that starts from an existing baseline design and is constrained to maintain the same topology for the shape. bounds on the i-th design variable. The design variables x have been added to the argument list of the state variable to emphasize its parametric dependence upon them. Likewise, we now write the state equation (1) as ; x ), x ; x) = 0 r (q (x (6)

2 OPTIMIZATION FORMULATION
The state variable q is a function of the spatial coordinate, . (The spatial variable has which is denoted by the vector x the unusual hat because in this volume the symbol x without the hat is reserved for the vector of design variables.) The state variable is a solution of the governing equations for an appropriate conceptual model for the ow (compressible or incompressible, viscous or inviscid, turbulent or laminar, nonlinear or linear). We write the governing equations generically as ), x ) = 0 r (q (x (1)

We necessarily have hk = rk for k = 1, . . . , dimension of r , as each component of r yields a constraint. Because the state vector is dened implicitly, as in equation (6), the objective and constraint functions in equations (2)(3) have a depen . For aerodynamic shape optimization, the dence upon q and x design variables characterize the shape of the airfoil, wing or aircraft. To illustrate the basic principles, we consider optimization of airfoils in the four-digit NACA series (see, e.g., Abbott and von Doenhoff, 1959, Chapter 6). These airfoils are described by the maximum camber (m), the distance of the location of the maximum camber (p) from the leading edge, and the maximum thickness (t ); all of these are given as fractions of the chord c. These parameters are illustrated in Figure 2a. For a four-digit NACA airfoil, say, a NACA d1 d2 d3 d4 airfoil, the rst digit d1 is m (in hundredths), the second digit is p (in tenths), and the last two digits are t (in hundredths). Thus, a NACA 2416 airfoil has a maximum camber of 0.20c located at x = 0.10c, and a maximum thickness of 0.16c, where c is the airfoil chord. Figure 2b is indicative of internal structures that constrain the geometry; such important constraints are ignored in the present illustration. The conceptual model used for the airfoil analysis in this example is inviscid, irrotational, incompressible, constantdensity ow. The velocity potential can be used as the (scalar) state variable. The mathematical model is given by the Laplace equation r= (7)

The optimization problems that are considered in this chapter have the form ; x), x ) min f (x; q(x
x

(2)

subject to ; x ), x ) 0 , gj (x; q(x ; x ), x ) = 0 , hk (x; q(x xiL xi xiU , j = 1, . . . , m k = 1, . . . , p i = 1, . . . , n (3) (4) (5)

where x is the vector of design variables, f is the objective function, gj and hk are inequality and equality constraint functions, respectively, and xiL and xiU are upper and lower

denotes the gradient operator (with respect to x ). where The Laplace equation is subject to a homogeneous Neumann n = 0, where n is unit vector norboundary condition ( mal to the surface) on the airfoil surface, an homogeneous Dirichlet condition ( = 0) at innity, and a Kutta condition (upper and lower surface pressures match) at the trailing edge The computational model is based upon a panel (boundaryelement) method1 . This very simple model ignores many important effects such as viscosity, compressibility, vorticity, and nonlinearity but provides a useful illustration of some basic concepts of aerodynamic shape optimization. For this example, the design variables x = (m, p, t ), with these three NACA airfoil parameters treated as continuous variables; their upper and lower bounds are taken to be xL = (0.0, 0.1, 0.0) and xU = (0.1, 0.5, 0.2), respectively. In

Encyclopedia of Aerospace Engineering, Online 2010 John Wiley & Sons, Ltd. This article is 2010 US Government in the US and 2010 John Wiley & Sons, Ltd in the rest of the world. This article was published in the Encyclopedia of Aerospace Engineering in 2010 by John Wiley & Sons, Ltd. DOI: 10.1002/9780470686652.eae500

Airfoil/Wing Optimization 3
Stiffeners and ribs

Fuel tank z x Leading edge tc mc pc (a) Chord c (b) Spars Mean camber line Chord line

Trailing edge

Figure 2. (a) National Advisory Committee for Aeronautics (NACA) four-digit airfoil parameters; (b) representative internal structures (courtesy of J. A. Samareh).

all cases, the angle of attack is chosen to be = 1 . The key performance outputs are the (section) lift, drag, and pitching moment coefcients, cl , cd and cm , respectively. (The moment reference center is the quarter-chord point.) Four representative cases are considered: (i) maximize lift (f = cl ) with no constraints, (ii) maximize lift with inequality constraints on the pitching moment (g1 = cm 0.04) and drag (g2 = cd 0.02), (iii) minimize drag (f = cd ) with inequality constraints on the pitching moment (g1 = cm 0.04) and lift (g2 = cm 0.30), and (iv) minimize lift/drag (f = cl /cd ) with inequality constraints on the pitching moment (g1 = cm 0.04) and lift (g2 = cl 0.30. (There are no equality constraints in these examples.) The gradient-based optimizations employ Sequential Quadratic Programming, using nite differences to compute the gradients and the BroydenFletcherGoldfarbShanno (BFGS) method to approximate the Hessian2 . The starting point for the optimizations is x0 = (xL + xU )/2. The optimal solutions to these four problems are given in Table 1. The pitching moment constraint is active in cases (ii) and (iv). The drag constraint is also active in case (ii). The lift constraint is active in case (iii). The corresponding airfoils are shown in Figure 3. The key performance parameters are also included in the table. The impact of constraints is especially dramatic in the contrast between the results for the rst two cases (maximization of lift). The conceptual model for this simple example has neglected several important physical effects, most noticeably those of viscosity and compressibility. The resulting optimal designs are by no means representative of airfoils that would be useful in practice.

0.2

0.1 z

0 0.1
max c l max c l [gi : cm and cd ] min cd [gi : cm and cl ]

0.2 0 0.2 0.4 x

min cl / cd [gi : cm and c l ]

0.6

0.8

Figure 3. Optimal shapes for the NACA four-digit airfoil example.

Another widely used objective function is the difference (in the least squares sense) between the surface pressure distribution and a target pressure distribution: f (x ) =
S

|| p( x; x) ptarget ( x) ||2

(8)

where the integral is taken over the airfoil or wing surface S of interest, p is the pressure in the ow produced by an airfoil described by the design variables x, and ptarget is the target pressure. The basic concept is illustrated in Figure 4. The solid line is the pressure coefcient (Cp ) for the present (baseline) airfoil. The dotted line is the desired (target) pressure distribution. Many rules have been built up over the

Table 1. Design variables and optimal solutions for the airfoil examples. Case 1 2 3 4 f cl cl cd cl /cd gi cm and cd cm and cl cm and cl m 0.10 0.029 0.016 0.028 p 0.50 0.10 0.30 0.11 t 0.20 0.15 0.10 0.10 min f 1.53 0.44 0.0010 369 cl 1.529 0.436 0.300 0.403 cd 0.0030 0.0020 0.0019 0.0011 cm 0.321 0.040 0.036 0.040

Encyclopedia of Aerospace Engineering, Online 2010 John Wiley & Sons, Ltd. This article is 2010 US Government in the US and 2010 John Wiley & Sons, Ltd in the rest of the world. This article was published in the Encyclopedia of Aerospace Engineering in 2010 by John Wiley & Sons, Ltd. DOI: 10.1002/9780470686652.eae500

4 Aerospace System Optimization

3 SHAPE DEFINITION
Aerodynamic shape optimization methods using nonlinear analysis tools are faced with signicant challenges in shape parameterization, volume mesh generation, and sensitivity analysis. Although these components are routine for many optimization problems in other disciplines, they are nontrivial for aerodynamic shape optimization methods that yield realistic shapes robustly and efciently. The surface of an aircraft, a wing, or even an airfoil is an innite-dimensional object, but it must be parameterized as a nite-dimensional object. (The parameters of the shape correspond to the design variables for aerodynamic shape optimization.) The shape parameterization should compactly cover the design space of interest while presenting no undue difculties. In particular, parameterization-induced waviness and discontinuities in (the slope and curvature of) the surface are undesirable. These features are undesirable because they present manufacturing and CFD analysis difculties (since ow elds are quite sensitive to such discontinuities). Sometimes, optimizing the shape of the entire surface is desired, but other times optimization is only applied to a portion of the surface. The alternatives summarized below apply in both cases. The usual custom for aerodynamic shape parameterization is to represent the surface as a baseline surface plus a perturbation expressed as an expansion in terms of basis functions. Consider the case of airfoils. The surface S is one dimensional, and the basis functions bk ( ) depend upon a single coordinate . The expansion is given by
d

Figure 4. Baseline and target pressure distributions (courtesy of R.L. Campbell).

years for choosing target pressure distributions to achieve desired performance measures. Methods for matching particular ow-eld characteristics rather than directly minimizing a global objective function such as drag are referred to as inverse design methods ; the more conventional alternative is called a direct optimization method . One could apply formal optimization to the inverse problem with the objective function given by equation (8). However, there is a wide variety of inverse design methods that minimize the objective in equation (8) much faster than direct optimization methods. An example of one such method is provided in Section 6 (see Labruj` ere and Slooff, 1993 for descriptions of many others). In practical applications, numerous geometric constraints are needed for obtaining an acceptable result, for example, requiring the external shape to accommodate internal structures (see Figure 2), minimal radius of curvature of the leading edge, and minimal angle of the trailing edge. Furthermore, aerodynamic shapes are required to operate over a range of conditions, especially in Mach number. This is addressed with a multi-objective optimization formulation (commonly referred to as multipoint optimization in the aerodynamic optimization community) (see Drela, 1998 for an extensive discussion of constraints and objectives for airfoil optimization). The NACA airfoil family used for the example above was convenient for this demonstration because the parameterization guaranteed a simple airfoil shape. Furthermore, the design variables corresponded directly with physical quantities for which experienced aerodynamicists have considerable intuition. However, the optimization problems of interest for many decades now require consideration of a much broader range of shapes than those covered by the NACA airfoil families. In the following section, we cover the fundamentals of parameterizations that can represent a more general range of shapes.

s( ; x) =
k =1

x k b k ( )

(9)

with the coefcients xk in the expansion serving as the design variables. The perturbation s( ; x) may be applied to whatever functions are used for the surface denition, for example, upper or lower airfoil surface, mean camber line, airfoil thickness. A variety of basis functions have been used, some global and some local. Since the leading and trailing edges of airfoils are xed in most design problems, having basis functions that vanish at both end points is desirable. The sine functions are one option for a complete set of global basis functions. Drela (1998) recommends using them in the form b k ( ) = 1 sin(k ) k (10)

They form an orthogonal set for [0, 1]. The sine basis functions are illustrated in Figure 5a. Suitable combinations of orthogonal polynomials can also be chosen to ensure that

Encyclopedia of Aerospace Engineering, Online 2010 John Wiley & Sons, Ltd. This article is 2010 US Government in the US and 2010 John Wiley & Sons, Ltd in the rest of the world. This article was published in the Encyclopedia of Aerospace Engineering in 2010 by John Wiley & Sons, Ltd. DOI: 10.1002/9780470686652.eae500

Airfoil/Wing Optimization 5
1 1

0.5 bk 0 k =1 k =2 k =3 k =4 0 (a) 0.2 0.4 x 0.6 0.8 1 (b) bk

0.5

0 k =2 k =3 k =4 k =5 0.5 0 x 0.5 1

0.5

0.5

1 1

Figure 5. Examples of global basis functions: (a) sine; (b) Legendre.

the basis functions vanish at the end points. For example, one can use 1 (Lk1 ( ) Lk+1 ( )) for k 1 bk ( ) = 2 (k + 1) (11)

(Hicks and Henne, 1978) functions, which are dened by b( ) = {sin[ log(1/2)/ log(t1 ) ]}t2 (12)

for [1, 1], where Lk ( ) is the Legendre polynomial of degree k. The set of basis functions given by equation (11) is nearly orthogonal the inner product of bk and bl vanishes except for l = k 2, k, k + 2. The Legendre basis functions are illustrated in Figure 5b. These form a complete set of basis functions. They are advantageous for approximating functions with a high degree of smoothness (more than, say, four continuous derivatives), because they exhibit much faster convergence than the sine functions in such cases (see Canuto et al., 2006, Chapter 2). A noteworthy set of local basis functions specically chosen for airfoil parameterization are the HicksHenne
1

with the domain now again normalized to [0, 1]; t1 controls the location of the peak and t2 its width. Figure 6a illustrates some of these functions. Although these functions lack the completeness property possessed by the sine functions and the orthogonal polynomials, they have proven very useful in the hands of skilled designers. A more conventional choice of local basis functions are splines, which are piecewise polynomials. Figure 6b illustrates the case of cubic B-splines on uniformly distributed knots (denoted by the dots on the x -axis); the curves are labeled by the x coordinate of the center of the spline. Unlike the global basis functions shown in Figure 5, the underlying B-spline basis functions effect only local changes in the shape. On the other hand, they have only a nite number
1
c = 0.3 c = 0.5 c = 0.7

0.75 bk 0.5
t 1 = 0.25, t 2 = 0.5

0.75 bk 0.5

0.25

t 1 = 0.50, t 2 = 1.0 t 1 = 0.75, t 2 = 1.5

0.25

0 0 (a)

0 0.2 0.4 x 0.6 0.8 1 (b) 0 0.2 0.4 x 0.6 0.8 1

Figure 6. Examples of local basis functions: (a) HicksHenne; (b) B-spline.

Encyclopedia of Aerospace Engineering, Online 2010 John Wiley & Sons, Ltd. This article is 2010 US Government in the US and 2010 John Wiley & Sons, Ltd in the rest of the world. This article was published in the Encyclopedia of Aerospace Engineering in 2010 by John Wiley & Sons, Ltd. DOI: 10.1002/9780470686652.eae500

6 Aerospace System Optimization

0.2
Baseline Camber perturbation

0.2
Baseline Thickness perturbation

0.1

0.1

0.1

0.1

0.2 0 (a ) 0.2 0.4 x 0.6 0.8 1 (b)

0.2 0 0.2 0.4 x 0.6 0.8 1

Figure 7. NURBS-based perturbations for (a) airfoil camber; (b) thickness.

of continuous derivatives, namely, p 1 continuous derivatives for splines of order p. Hence, if p > 2, then the shape perturbation, along with its rst and second derivatives, is continuous. This includes the cubic B-splines (p = 3) shown in Figure 6b. A generalization of B-splines, called nonuniform rational B-splines (NURBS), has become a fairly widely used parameterization, particularly for complex shapes, such as full aircraft congurations. NURBS represent functions (in this case surfaces) as a rational function, with both the numerator and denominator consisting of B-spline expansions. Piegl and Tiller (1996) provide a comprehensive description of NURBS. See Samareh (2001) for a summary of their mathematical description and a short survey of various uses in aerodynamic shape optimization. An important advantage of a NURBS representation for the shape is that they are compatible with most computer-aided design (CAD)

systems. Figure 7 illustrates airfoil deformations based on NURBS expansions of the airfoil camber and thickness. Using the locations of the surface mesh points as design variables is yet another approach. The main challenge here is picking appropriate geometric constraints and/or ltering procedures to ensure a smooth surface. See Li and Krist (2005) for one among many approaches to surface smoothing. For the parameterization of wings, the same approaches apply, but, of course, there are now additional types of design variables. Some representative wing parameters with a direct aerodynamic interpretation are illustrated in Figure 8. The symbols indicate locations where the parameters are dened. The root chord, tip chord, semi-span, and leading edge sweep angle are planform variables; these are usually xed early in the design process. The airfoil section parameters
y

Leading edge sweep angle Camber and thickness Root chord Twist and shear

Tip chord

Semi-span

Figure 8. Typical design variables for wings.

Encyclopedia of Aerospace Engineering, Online 2010 John Wiley & Sons, Ltd. This article is 2010 US Government in the US and 2010 John Wiley & Sons, Ltd in the rest of the world. This article was published in the Encyclopedia of Aerospace Engineering in 2010 by John Wiley & Sons, Ltd. DOI: 10.1002/9780470686652.eae500

Airfoil/Wing Optimization 7 (camber and thickness) as well as the wing twist and shear are often the focus of wing shape design. The twist angle at a given airfoil section is the difference between the airfoil section incident angle at the root and the incident angle of that airfoil section. Similarly, the shear (dihedral) is the difference between the airfoil leading edge z coordinate for the root and the z coordinate for the particular airfoil section. CFD volume mesh for a wingfuselage conguration. The coordinates of these volume mesh-points are denoted by y in Figure 1. For aerodynamic shape optimization, changes to the surface of the airfoil resulting from design variable changes require adjustments to the volume mesh as well, and these adjustments need to occur automatically. Nowadays, computational meshes suitable for even viscous CFD analysis can be generated automatically from the surface shape denition for airfoils and wings, but this state has proven to be very elusive for complex aircraft congurations subjected to NavierStokes analyses some type of user intervention is typically needed if the mesh generation must be performed ab initio . Thompson, Soni and Weatherill (1998) provide an extensive description of methods for CFD mesh generation. Moreover, as discussed in the next section, analytically based gradients are highly desirable, and these are not generally available from ab initio mesh generation packages. Hence, many aerodynamic shape optimization processes use special processes that lend themselves to analytically based gradients for obtaining the volume mesh as a perturbation upon the mesh associated with the baseline shape. Samareh (2001) contains an overview of the various approaches to volume mesh generation.

4 MESH GENERATION
Unlike panel methods, for which a surface mesh is sufcient, CFD methods require a volume mesh for numerical solution of the state equation. For compressible ow, the state variable q is given by q = (, u, E)T (13)

where is the density, u the velocity, and the (specic) total 1 energy E = e + 2 u u , where e denotes the (specic) internal energy. For steady, inviscid ow described by the Euler equations, we have F ), x ) = r (q (x (14)

where the ux F = (u, uuT , uE + pu)T . The Navier Stokes equations have the same form as equation (14) but with additional terms in the ux function. Figure 9 illustrates the surface and symmetry plane portions of an unstructured

5 GRADIENT COMPUTATION
For gradient-based optimization methods, the derivatives of the objective and constraints with respect to the design variables (terms in parentheses in Figure 1) are needed. For nonlinear CFD methods, computation of the gradients3 using nite differences has several disadvantages: (i) computation of each gradient requires another full CFD solution since the equations are nonlinear; (ii) extensive trial and error is necessary to choose the appropriate step size for each design variable; and (iii) for some difcult CFD problems, the analysis code may be simply unable to converge to the level needed for accurate nite-difference gradients. Computation of these derivatives through quasi-analytical means is preferred. (The adjective quasi-analytical is used because the equations for the gradients are derived analytically but solved numerically.) In order to distinguish the state variable and the state equations, which are functions, from their discrete rep and resentations, which are vectors, we use the symbols q , respectively, for the latter; similarly y is the vector of the r mesh-point coordinates. For a problem with N mesh-points, and r is 5N since q has ve components at the length of q is 3N . (Boundary conditions each point, and the length of y may slightly alter these lengths.)

Figure 9. Surface mesh for wing optimization in the presence of the fuselage. Reproduced with permission from Nielsen and Park (2006) c AIAA.

Encyclopedia of Aerospace Engineering, Online 2010 John Wiley & Sons, Ltd. This article is 2010 US Government in the US and 2010 John Wiley & Sons, Ltd in the rest of the world. This article was published in the Encyclopedia of Aerospace Engineering in 2010 by John Wiley & Sons, Ltd. DOI: 10.1002/9780470686652.eae500

8 Aerospace System Optimization Recall that a change in a design variable is propagated through the shape denition and mesh generation processes (Figure 1), with the computational mesh y depending upon x. In particular, the gradient of, say, the objective function with respect to the particular design variable xj is given by f df = + d xj xj
5N k =1

f q k q k xj

3N

+
l =1

f y l y l xj

(15)

Evaluation of this requires determination of the three explicit partial derivatives of f plus the q k /xj and y l /xj terms. The rst term on the right-hand side is usually straightforward to compute. (For the airfoil example in Section 2, all these derivatives vanish, as the design variables do not appear explicitly in the expressions for the lift, drag, and pitching moment.) Likewise, an analytical expression can typically be derived straightforwardly and then evaluated numerically for f/q k and f/y l . The y l /xj terms represent the inuence of the design variables upon the computational mesh. The surface denition techniques illustrated for airfoils in Figures 57, as well as some mesh perturbation techniques, lend themselves to analytical evaluation of these terms. Computation of the q k /xj terms is more involved. Consider the state equation given by equation (6). Its solution q, considered as a function of the design variables x, satises dr /dx = 0. Hence, implicit differentiation of the state equation yields the following expression:
5N k =1

equations are derived from the algebraic equations resulting from the discretization of the PDEs.) See studies by Newman et al. (1999) and Reuther et al. (1999) for extended lists of references to the early experiences of various groups with both approaches. Some of the subtle issues that must be addressed for consistent aerodynamic gradients are the treatment of boundary conditions and mesh gradients (see Giles et al., 2003; Nielsen and Park, 2006 for detailed recommendations). An approach to obtaining second-order derivatives by combining direct and adjoint techniques is described by Sherman et al. (1996). Yet another alternative to computing the gradients is the use of complex variables, which we illustrate on a simple real-valued scalar function f . In this method, the function is evaluated at the complex point x + i , where i denotes the imaginary unit, for small . A Taylor expansion yields f (x + i ) = f (x) + (i ) df + O( 2 ) dx (17)

r i q k q k xj

r i xj

3N

l =1

r i y l y l xj

(16)

Thus, for small , the real part of f (x + i ) is a good approximation to f (x), and the imaginary part divided by is a good approximation to the gradient df/dx. Note that this approach does not involve the subtraction of nearly equal quantities, as occurs for gradients computed via nite differences. Hence, round-off errors are not a concern for the complex variable approach. Provided that the source code is available, implementation of this approach can require little more than changing the type of variables from real to complex. A prototypical use of this approach for aerodynamic shape optimization is given by Nielsen and Kleb (2006).

Equation (16) is linear in the desired q k /xj term. The size of the linear system is 5N 5N . CFD computations for wings typically use O(106 ) mesh-points, and complex congurations can easily take upwards of 108 mesh-points. Hence, the linear system is extremely large in three dimensions so large that direct solution methods are impractical. An efcient iterative solution method is surveyed in Newman et al. (1999) and described in detail in Korivi et al. (1994). For most aerodynamic shape optimization problems, the number of design variables far exceeds the number of objectives and constraints. Hence, the adjoint approach for obtaining the gradients is much more efcient than the direct approach described above. Details may be found, for example, in Newman et al. (1999) and Reuther et al. (1999). Both continuous and discrete adjoints have been used in aerodynamics. (The phrase continuous adjoint refers to one in which the continuous adjoint equations are rst derived from the governing partial differential equations (PDEs) and then discretized; a discrete adjoint is one in which the adjoint

6 OPTIMIZATION USING CFD


Over the past several decades, a broad assortment of direct optimization and inverse design methods has been applied to aerodynamic shape optimization using CFD. The principal distinctions between the approaches are (i) use of global performance measures such as lift, drag, and pitching moment in the objectives and constraints versus inverse objectives such as matching a prescribed pressure distribution, (ii) gradientbased methods versus methods using only function values, and (iii) having the optimizer invoke the aerodynamic analysis tool itself versus having it rely upon surrogate models constructed off-line. Most of the review articles cited in Section 1 concentrate on direct optimization of global performance measures using gradient-based methods that directly call CFD analysis tools. Several difculties with this approach can arise. For example, the analysis code may produce solutions that result in

Encyclopedia of Aerospace Engineering, Online 2010 John Wiley & Sons, Ltd. This article is 2010 US Government in the US and 2010 John Wiley & Sons, Ltd in the rest of the world. This article was published in the Encyclopedia of Aerospace Engineering in 2010 by John Wiley & Sons, Ltd. DOI: 10.1002/9780470686652.eae500

Airfoil/Wing Optimization 9 objective functions and constraints that are contaminated by numerical noise. Such noise gives the appearance of many local extrema and can render gradient-based methods unworkable. Another potential difculty is that there may be portions of the desired design space for which the mesh generation or the analysis process simply fails. At an even more fundamental level, the problem formulation (objective and constraints) itself can omit important considerations with the result that the optimization produces a design that is unrealistic from the perspectives of other disciplines. See the cited review articles for extensive discussion of these issues. Nevertheless, many groups have used gradient-based optimization of global performance measures without recourse to surrogate models to produce impressive CFD-based optimization results for multi-element airfoils, wings, wing fuselage congurations, and even more complex shapes. Figure 10, from Nielsen and Park (2006), shows the objective function convergence for unconstrained maximization of the lift-to-drag ratio of the vehicle illustrated in Figure 9 at transonic conditions using a NavierStokes CFD code. The optimum was achieved in a few dozen function evaluations. The surface denition for the wing utilized NURBS representations of the camber, thickness, twist, and shear (see Figure 8) perturbations from the baseline shape. (The fuselage was taken as xed.) Gradients were computed with a quasi-analytical adjoint method. One particular approach to inverse design that has seen considerable use in industry applications is based on

Figure 11. Design improvement regimes for an aircraft (courtesy of J. Hooker and A. Agelastos (Lockheed-Martin) and W. Milholen and R. Campbell (NASA)).

L / D = + 32%

10

15 20 25 Function evaluation

30

35

Figure 10. Lift/drag convergence from a wing optimization. Reproduced with permission from Nielsen and Park (2006) c AIAA.

modifying the surface curvature in proportion to the desired change in surface pressure. The surface geometry is modied at the same time that the ow solver is converging; hence, this method costs little more than a single CFD analysis. Details can be found in the study by Campbell (1992). Figure 11, taken from joint Lockheed-Air Force-NASA work, illustrates the various regions of the aircraft to which this inverse design method was applied. (PAI refers to propulsion-airframe integration.) Collectively, these improvements in the local aircraft shape enabled the design to meet its performance goals. The most common non-gradient-based optimization methods that have been applied to aerodynamic shape optimization are genetic algorithms (see Holst, 2005 for an example). These have exhibited the customary advantage of greater likelihood of nding the global optimum, as well as the accompanying disadvantage of taking signicantly more computational time for convergence than gradient-based methods. In aerodynamics, as in other disciplines, surrogate models are extensively used in optimization. In some cases, these are used to deal with noisy analysis results. Linear aerodynamics methods are particularly prone to producing noisy results. (Some surrogate models such as quadratic or cubic response surfaces smooth the noise, but some other surrogate models do not.) In other cases, surrogates are used to deal with portions of the design space that cause difculties for the analysis code, since the surrogate model enables the optimization to proceed even in the face of occasional analysis failures. Yet another use is to permit efcient searches for global optima using, say, pattern-search or genetic algorithm techniques. For the relatively expensive Euler and NavierStokes models, use of surrogates entails the usual compromise between accuracy (relative to the CFD model) and speed. All types of surrogates design of experiments, kriging, neural networks, radial basis functions have been fruitfully employed on this

Encyclopedia of Aerospace Engineering, Online 2010 John Wiley & Sons, Ltd. This article is 2010 US Government in the US and 2010 John Wiley & Sons, Ltd in the rest of the world. This article was published in the Encyclopedia of Aerospace Engineering in 2010 by John Wiley & Sons, Ltd. DOI: 10.1002/9780470686652.eae500

L / D ratio

10 Aerospace System Optimization application. These types of surrogates all suffer from the curse of dimensionality, limiting the number of design variables in practice to O(10). Of course, one would like optimization using a surrogate to converge to the optimum corresponding to that for the underlying analysis code. Several optimization methods that judiciously mix calls to the surrogate with calls to the actual code have been proven mathematically to converge to the true optimum. See Booker et al. (1999) for discussion of one method in a general setting, and see Alexandrov et al. (2001) for another method with an aerodynamic shape optimization application; in practice, the latter seems to reduce the overall cost by a factor of 35 compared with always invoking the CFD code. When the surrogate is just a lower delity model, for example, using an Euler code instead of a NavierStokes code or a coarse-grid solution in place of a ne-grid solution, then the number of design variables is not limited as it is for generic surrogates. In addition to the considerations mentioned above, one needs to weigh the computational costs of the various methods, at least for CFD. (Linear aerodynamics methods are so fast on desktop computers that CPU time is rarely an issue, regardless of the choice of optimization method.) Roughly speaking, in terms of the cost of a single CFD analysis, an inverse design method takes O(1) analysis time, gradient-based optimization using the adjoint formulation takes O(10) analysis time, and non-gradient-based methods (including surrogate ones) take O(100 10 000) analysis time. All these methods have their niches. Non-gradientbased methods and surrogate models enable exploration of large regions of the design space. Gradient-based methods facilitate renement of a local optimum, and inverse design methods help ne-tune performance in local regions of the surface.

NOTES
1. The underlying computations are performed with a TM code of L.N. Sankar, found on-line at Matlab http://www.ae.gatech.edu/people/lsankar/AE3903/. 2. Performed with the MatlabTM routine fmincon. 3. These gradients are often called sensitivity derivatives .

REFERENCES
Abbott, I.A. and von Doenhoff A.E. (1959) Theory of Wing Sections, Dover Publications, New York. Alexandrov, N.M., Lewis, R.M., Gumbert, C.R., Green, L.L. and Newman, P.A. (2001) Approximation and model management in aerodynamic optimization with variable-delity models. J. Aircraft, 38(6), 10931101. Booker, A.J., Dennism, J.E., Jr., Frank, P.D., Serani, D.B., Torczon, V. and Trosset, M.W. (1999) A rigorous framework for optimization of expensive functions by surrogates. Struct. Optim., 17(1), 113. Campbell, R.L. (1992) An approach to constrained aerodynamic design with application to airfoils. NASA-TP-3260. Canuto, C., Hussaini, M.Y., Quarteroni, A., and Zang, T.A. (2006) Spectral Methods, Fundamentals in Single Domains, Springer, New York. Drela, M. (1998) Pros and cons of airfoil optimization, in Frontiers of Computational Fluid Dynamics (eds D.A. Caughey, and M.M. Hafez), World Scientic, Hackensack, pp. 363381. Giles, M.B., Duta, M.C., M uller, J.-D. and Pierce, N.A. (2003) Algorithm developments for discrete adjoint methods. AIAA J., 41(2), 198205. Hicks, R.M. and Henne, P.A. (1978) Wing design by numerical optimization. J. Aircraft, 15(7), 407412. Holst, T. (2005) Genetic algorithms applied to multi-objective aerospace shape optimization. J. Aero. Comput. Info. Comm., 2, 217235. Korivi, V.M., Taylor, A.C. Ill, Newman, P.A., Hou, G.J.-W. and Jones, H.E. (1994) An approximate factored incremental strategy for calculating consistent discrete CFD sensitivity derivatives. J. Comput. Phys., 113(2), 336346. Labruj` ere, Th.E. and Slooff, J.W. (1993) Computational methods for the aerodynamic design of aircraft components. Ann. Rev. Fluid Mech., 25, 183214. Li, W. and Krist, S. (2005) Spline-based airfoil curvature smoothing and its applications. J. Aircraft, 42(4), 10651074. Mohammadi, B. and Pironneau, O. (2004) Shape optimization in uid mechanics. Ann. Rev. Fluid Mech., 36, 255279. Newman, J.C. III, Taylor, A.C. III, Barnwell, R.W., Newman, P.A. and Hou, G.J.-W. (1999) Overview of sensitivity analysis and shape optimization for complex aerodynamic congurations. AIAA J., 36(1), 8796

7 DISCLAIMER
This chapter is declared a work of the U.S. Government and is not subject to copyright protection in the United States.

ACKNOWLEDGMENTS
The author gratefully acknowledges numerous discussions with his colleagues at the NASA Langley Research Center for all their contributions to his understanding of this subject. The numerical examples in Section 3 utilized publicly available codes of L.N. Sankar. J.A. Samareh supplied the code used for Figure 7.

Encyclopedia of Aerospace Engineering, Online 2010 John Wiley & Sons, Ltd. This article is 2010 US Government in the US and 2010 John Wiley & Sons, Ltd in the rest of the world. This article was published in the Encyclopedia of Aerospace Engineering in 2010 by John Wiley & Sons, Ltd. DOI: 10.1002/9780470686652.eae500

Airfoil/Wing Optimization 11
Nielsen, E.J. and Kleb, W.L. (2006) Efcient construction of discrete adjoint operators on unstructured grids using complex variables. AIAA J., 44(4), 827836. Nielsen, E.J. and Park, M.A. (2006) Using an adjoint approach to eliminate mesh sensitivities in computational design. AIAA J., 44(5), 948953. Piegl, L. and Tiller, W. (1996) The NURBS Book, Springer, New York. Reuther, J.J., Jameson, A., Alonso, J.J., Rimlinger, M.J. and Saunders, D. (1999) Constrained multipoint aerodynamic shape optimization using an adjoint formulation and parallel computers. Parts 1 and 2. AIAA J., 36(1), 5174. Samareh, J.A. (2001) Survey of shape parameterization techniques for high-delity multidisciplinary shape optimization. AIAA J., 39(5), 877884. Sherman, L.L., Taylor, A.C. III, Green, L.L., Newman, P.A., Hou, G.J.-W. and Korivi, V.M. (1996) First- and second-order aerodynamic sensitivity derivatives via automatic differentiation with incremental iterative methods. J. Comput. Phys., 129(2), 307 331. Thompson, J.R., Soni, B.K. and Weatherill, N.P. (eds) (1998) Handbook of Grid Generation, CRC Press, Boca Raton.

Encyclopedia of Aerospace Engineering, Online 2010 John Wiley & Sons, Ltd. This article is 2010 US Government in the US and 2010 John Wiley & Sons, Ltd in the rest of the world. This article was published in the Encyclopedia of Aerospace Engineering in 2010 by John Wiley & Sons, Ltd. DOI: 10.1002/9780470686652.eae500

You might also like