Professional Documents
Culture Documents
Keywords: optimal shape design, airfoil, Navier–Stokes, high-lift, discrete gradient, adjoint, flow
sensitivities, Newton’s method, GMRES
173.1
MARIAN NEMEC AND DAVID W. ZINGG
Y/C
12
conditioned generalized minimum residual (GM- 0 1 4 FY
1
RES) Krylov subspace method [25]. Overall, the -0.1
8 FX
gradient is obtained in just one-fifth to one-half
-0.2
0 0.25 0.5 0.75 1 1.25
of the time required for a warm-started flow solu- X/C
173.2
From Analysis to Design of High-Lift Configurations Using a Newton–Krylov Algorithm
where h (z j ) represents the minimum allowable construct a local cubic interpolant. Note that
thickness at location z j expressed as a fraction of the optimization problem is based on the discrete
the airfoil’s chord. For multi-element configura- form of the flowfield equations. Using the dis-
tions, it is also necessary to constrain the gap and crete approach, we expect the gradient to vanish
overlap distances. These constraints are required at the optimum solution. In the following sec-
in order to prevent collisions among the elements tions, we present the formulation for the penal-
and to ensure a reasonable computational grid. ized objective function, as well as the algorithms
The governing flow equations are the com- used for the flowfield evaluation (objective func-
pressible two-dimensional thin-layer Navier– tion value), the gradient evaluation, and the grid-
Stokes equations in generalized coordinates, perturbation strategy.
∂Eb(Qb; X ) ∂F
b(Qb; X ) b b
1 ∂S(Q; X ) 3.1 Objective with Constraints
+ = Re (6)
∂ξ ∂η ∂η
A penalty method is used to combine the objec-
where Q b = J 1 Q = J 1 [ρ; ρu; ρv; e℄T is the vec- tive function with the constraint equations
tor of conservative dependent state variables, ξ
and η are the streamwise and normal general- Nc
173.3
MARIAN NEMEC AND DAVID W. ZINGG
solution the design variables, and consequently and storage requirements of the incomplete fac-
the computational grid, are constants. torization. This approach is similar to the ‘di-
Eq. 9 is solved in a fully-coupled manner, agonal shift’ strategy suggested by Chow and
where convergence to steady state is achieved Saad [4]. The present preconditioning matrix is a
using the preconditioned GMRES algorithm in compromise between a preconditioner based on a
conjunction with an inexact-Newton strategy [21, first-order upwind discretization of the flowfield
16]. The main components include matrix- equations and a preconditioner based on the ac-
free GMRES(40) and a block-fill incomplete tual second-order discretization. This novel ‘in-
LU (BFILU) preconditioner. The matrix-vector termediate’ preconditioner is significantly more
products required at each GMRES iteration are effective than either of these more commonly
formed with first-order finite-differences. Right used approaches.
preconditioning is used, and the preconditioner is
based on an approximate-flow-Jacobian matrix. 3.3 Gradient Evaluation
The level of fill for most cases is 2 [BFILU(2)],
The gradient, G , of the objective function
but difficult multi-element cases may require
J [X ; Q(X )℄ is given by
BFILU(4). The approximate-factorization algo-
rithm [22, 15] is used to reduce the initial resid- dJ ∂J ∂J dQ
ual by three orders of magnitude in order to avoid G= = + (11)
dX ∂X ∂Q dX
Newton startup problems.
The approximate-flow-Jacobian matrix used where we reduce the vector of design variables,
for the preconditioner is identical to the flow- X , to a scalar in order to clearly distinguish be-
Jacobian matrix, ∂R=∂Q, except for the treat- tween partial and total derivatives. For problems
ment of the artificial-dissipation coefficients [16]. with multiple design variables, it may be helpful
Hence, the preconditioner contains the contribu- to note that G and ∂J =∂X are [1 ND ℄ row vec-
tions from all components of the residual vec- tors, ∂J =∂Q is a [1 NF ℄ row vector, and dQ=dX
tor, namely inviscid and viscous fluxes, bound- is a [NF ND ℄ matrix, where ND represents the
ary conditions, block interfaces, and the tur- number of design variables and NF represents the
bulence model. The artificial-dissipation coef- number of flowfield variables. We assume that
ficients, which include the spectral radius and the implicit function Q(X ) is sufficiently smooth
the pressure switch, are assumed to be constant even in the presence of flow discontinuities such
with respect to the flowfield variables. Fur- as shock waves. See [10, 8] for further details.
thermore, the preconditioning matrix is formed The difficulty in Eq. 11 is the evaluation of
with only second-difference dissipation, but the the term dQ=dX , referred to as the flow sensitivi-
second-difference coefficient is combined with ties. Evaluation of the partial derivatives, ∂J =∂X
the fourth-difference coefficient as follows, and ∂J =∂Q, is relatively straightforward and is
(2) (2) (4)
described at the end of this section. The flow sen-
dl = dr + φ dr (10) sitivities are obtained by differentiating the flow-
where the subscript r denotes the contribution field equations, Eq. 9, with respect to the design
from the right-hand side, and the subscript l de- variables, which yields the following large sys-
notes the resulting left-hand side value used in tem of linear equations
forming the preconditioner. This modification ∂R dQ ∂R
does not affect the steady-state solution. Fast = (12)
∂Q dX ∂X
convergence is obtained with the value of φ set
to 6.0, which has been determined through nu- The direct, or flow-sensitivity, method results
merical experiments. from solving Eq. 12 for the flow sensitivities
Eq. 10 improves the diagonal dominance of dQ=dX and using these values in Eq. 11 to ob-
the preconditioning matrix and reduces the work tain the gradient.
173.4
From Analysis to Design of High-Lift Configurations Using a Newton–Krylov Algorithm
173.5
MARIAN NEMEC AND DAVID W. ZINGG
4 Validation -8
-10
Before presenting design examples for high-lift
configurations, we carefully validate the perfor- -12
mance of the flow solver and the gradient com-
-14
putation algorithm. All grids for this and the fol-
0 200 400 600 800
lowing sections consist of approximately 31,000
nodes. The off-wall spacing is 2 10 6 c, the
Time (s)
distance to the outer boundary is 12c, the spac- Fig. 3 Flow-solver performance
ing at the H-topology grid stagnation points is
2 10 5 c, and the trailing-edge clustering is
2 10 3 c. The reported CPU times are obtained started flow solves typically converge in 2/3 of
on a 667 MHz Alpha 21264 processor (SPECfp the original flow solve time.
2000 rating of 562 peak).
4.2 Gradient Accuracy
4.1 Flow-Solver Performance
Finite-difference gradients provide a benchmark
A fast solution of the flowfield equations is a that is used to establish the accuracy of the gra-
critical component of an effective design algo- dient computation using the flow-sensitivity and
rithm, since an evaluation of the objective func- adjoint methods. A subsonic lift-enhancement
tion is required at each iteration of the optimizer. problem for the NLR 7301 configuration is con-
The performance of the flow solver is examined sidered. During the computation of the finite-
for the NLR 7301 configuration at M∞ = 0:25, difference gradient, the flowfield solution is con-
α = 8Æ and Re = 2:51 106 . Fig. 3 shows that verged 14 orders of magnitude. The flow-
the Newton–Krylov flow solver (denoted as NK) sensitivity and adjoint equations are converged 8
is approximately two to three times faster than the orders of magnitude.
approximate-factorization flow solver (denoted The freestream conditions are M∞ = 0:25,
as AF). For many cases, this speed-up can be α = 4Æ , and Re = 2:51 106 . We compute the
even larger. Initially, the convergence rate of gradient of the objective function, Eq. 4, with re-
both flow solvers is identical, since approximate- spect to control point 5 on the main airfoil (de-
factorization is used as a startup procedure for the noted as 5M), control point 4 on the flap (denoted
Newton–Krylov flow solver. as 4F), and the horizontal and vertical flap dis-
One of the main difficulties associated with placements (denoted as Fx and Fy , respectively),
Newton’s method is the startup procedure. The see Fig. 1. The target drag coefficient, CD , is set
Newton–Krylov flow solver is particularly well equal to the initial drag coefficient, while the tar-
suited for the design problem since once we ob- get lift coefficient, CL , is set equal to 2:2, which
tain the solution for the initial airfoil shape, we represents a 2.5% increase from the initial value.
warm-start the remaining flow solves. If the step- The values of ωL and ωD in Eq. 4 are both set
sizes during the line-search procedure are suffi- to 1.0 and there are no side constraints. Ta-
ciently small, the startup procedure using approx- ble 1 shows that there is an excellent agreement
imate factorization is not necessary. The warm- between the finite-difference, adjoint, and flow-
173.6
From Analysis to Design of High-Lift Configurations Using a Newton–Krylov Algorithm
0 S-MF D.V. 1
Table 1 Gradient accuracy S-MF D.V. 2
-1 S-MF D.V. 3
Design Finite Adjoint S-MFa S-MF D.V. 4
Adjoint
Variable Difference (% Diff.) (% Diff.)b
b -2
a matrix-free flow-sensitivity -6
b % Diff = (G GFD )=GFD 100
-7
-8
10 20 30 40 50 60 70 80
GMRES Iterations
sensitivity gradients.
Fig. 4 GMRES convergence (S-MF denotes
4.3 Gradient Computation Efficiency matrix-free flow-sensitivity, D.V. denotes design
variable)
Fig. 4 shows the convergence of the ad-
joint and flow-sensitivity equations for the lift-
enhancement problem introduced in the previous only three orders of magnitude in order to ob-
subsection. The residuals of the flow-sensitivity tain gradients of sufficient accuracy [23, 18, 12].
equation are shown for each design variable. This level of convergence is achieved in approx-
Fig. 4 highlights the influence of different right- imately 45 seconds, as shown in Fig. 5. For
hand-side vectors on the convergence of GM- the flowfield equations, we typically reduce the
RES(85). Note that the initial guess for the ad- residual by ten orders of magnitude in order to
joint and flow-sensitivity solution vectors is zero. prevent stalling of the line searches once the so-
The left-hand side of the flow-sensitivity equa- lution is close to the optimum. This conver-
tion, Eq. 12, is the flow-Jacobian matrix. This gence level is achieved in 245 seconds, and con-
matrix remains the same for each design variable. sequently, the gradient is obtained in less than
For the adjoint equation, Eq. 14, the left-hand one-fifth of the flow solve time.
side is the transpose of the flow-Jacobian ma-
trix. The transpose operator does not change the
5 Design Examples
eigenvalues of the flow Jacobian. Although the
flow-sensitivity equation converges faster than
5.1 Flap Position Optimization
the adjoint equation, the flow-sensitivity equation
must be solved for each design variable, resulting The goal of this design example is to determine
in a much longer gradient computation time. the optimal gap and overlap distances for the
Fig. 5 compares the convergence history of NLR 7301 configuration, such that the result-
the adjoint and flowfield equations with respect ing configuration achieves a higher lift coefficient
to CPU time. The time for the formation of the while maintaining the same (or lower) drag coef-
preconditioning matrices is included in Fig. 5.1 ficient. The freestream conditions are specified
It is necessary to converge the adjoint equation in Subsection 4.2. The initial values of CL and
CD are 2.145 and 0.04720, respectively. The ob-
1 In Fig. 5, the ‘flat step’ in the convergence of the jective function is given by Eq. 4, where we set
CL = 2:180 and CD equal to the initial drag coeffi-
flow solver after a three order-of-magnitude decrease in the
residual indicates the formation time of the preconditioner.
For the adjoint equation, this time is indicated at the start cient. The weights ωL and ωD are set to 1.0. The
of the convergence history. design variables are the horizontal and vertical
173.7
MARIAN NEMEC AND DAVID W. ZINGG
Adjoint
-2 Flow Solve
Table 2 Gap-overlap optimization summary
Design CL CD Ga Ob
log( Residual )
-4
NLR 7301 2.145 0.04720 2.40 -5.31
-6 #1 2.165 0.04687 1.99 -4.28
#2 2.173 0.04677 1.95 -3.30
-8 Final 2.175 0.04675 2.02 -2.68
Target 2.180 0.0472
-10 a Gap (%c)
b Overlap (%c)
-12
0 50 100 150 200 250 300
CPU Time (s)
0
on the initial configuration. The weight, ωT , as-
sociated with the gap and overlap constraints is -0.025
set to 0:05. The gradient is computed using the
-0.05
adjoint method.
0.85 0.9 0.95 1
Table 2 and Fig. 6 summarize the results. X/C
Within a few flowfield and gradient evaluations,
the flap reaches the maximum allowable over- Fig. 6 Flap position summary
lap distance of approximately 4:3%c, at which
point the overlap penalty function becomes ac-
tive. The optimization converges to the design
#1 configuration, shown in Fig. 6. A new grid
is generated for this configuration and the corre- -5 -2.5
Cp
-2
design is obtained (see Fig. 6), where the gap and -0.5
overlap constraints are no longer active. Note that -1
0
the drag objective is satisfied for all the designs.
Consequently, the optimization is purely attempt- 0 0.5
173.8
From Analysis to Design of High-Lift Configurations Using a Newton–Krylov Algorithm
0 3
Design #2
Final Design 2.9
-1
2.8
-2
2.7
G24-O53
log( Gradient )
-3 2.6 G29-O05
Gap (%c)
2.5
-4
2.4
-5 2.3
2.2
-6
2.1
-7 1
2 2
-8 1.9
5 10 15 20 -5 -4 -3 -2 -1 0
Flow Solves and Gradient Evaluations Overlap (%c)
Fig. 8 Convergence histories for gap-overlap op- Fig. 9 Convergence to optimal gap-overlap dis-
timization tances from two distinct initial conditions
173.9
MARIAN NEMEC AND DAVID W. ZINGG
-2.5
Table 3 Flap optimization summary -5
-2
Design CL CD Ga Ob -4 NLR 7301
Final Design
NLR 7301 2.145 0.04720 2.40 -5.31 -3
-1.5
Cp
Cp
Final 2.2 0.04723 2.06 -2.40 -2
0
a Gap (%c)
b 0
Overlap (%c) 0.5
1 1
0 0.2 0.4 0.6 0.8 0 0.1 0.2 0.3
X/C X/C
0.05
NLR 7301
Initial Fig. 12 Cp distribution for main element and flap
0.025 6 D. V.
Y/C
0
mately one-fifth of the flow solve time when us-
-0.025 ing the adjoint method. Overall, the results indi-
-0.05
cate that the new algorithm provides an efficient
0.9 0.95 1 1.05 and robust tool for complex aerodynamic design
X/C
problems.
Fig. 11 Flap shape and position summary (D.V.
denotes design variable) Acknowledgments
173.10
From Analysis to Design of High-Lift Configurations Using a Newton–Krylov Algorithm
multi-element airfoils. AIAA Paper 93–0969, gorithm for complex aerodynamic design. Proc
1993. 10th Annual Conference of the CFD Society of
[6] Elliott JK and Peraire J. Aerodynamic opti- Canada, Windsor, ON, June 2002.
mization on unstructured meshes with viscous [18] Nielsen EJ. Aerodynamic Design Sensitivi-
effects. AIAA Paper 97–1849, June 1997. ties on an Unstructured Mesh Using the Navier
[7] Eyi S, Lee KD, Rogers SE, and Kwak D. High- Stokes Equations and a Discrete Adjoint Formu-
lift design optimization using Navier–Stokes lation. PhD thesis, Virginia Polytechnic Insti-
equations. Journal of Aircraft, Vol. 33, No 3, tute and State University, 1998.
pp 499–504, 1996. [19] Nielsen EJ and Anderson WK. Recent improve-
[8] Giles MB and Pierce NA. An introduction to the ments in aerodynamic design optimization on
adjoint approach to design. Flow, Turbulence unstructured meshes. AIAA J., Vol. 40, No 6,
and Combustion, Vol. 65, No 3/4, pp 393–415, pp 1155–1163, 2002.
2000. [20] Nocedal J and Wright SJ. Numerical Optimiza-
[9] Godin P, Zingg DW, and Nelson TE. High-lift tion. Springer–Verlag Inc., New York, 1999.
aerodynamic computations with one- and two- [21] Pueyo A and Zingg DW. Efficient Newton–
equation turbulence models. AIAA J., Vol. 35, Krylov solver for aerodynamic computations.
No 2, pp 237–243, 1997. AIAA J., Vol. 36, No 11, pp 1991–1997, 1998.
[10] Gunzburger MD. Introduction into mathemati- [22] Pulliam TH. Efficient solution methods for the
cal aspects of flow control and optimization. In- Navier-Stokes equations. Technical report, Lec-
verse Design and Optimization Methods, Brus- ture Notes for the von Kármán Inst. for Fluid
sels, Belgium, April 1997. von Kármán Institute Dynamics Lecture Series, Brussels, Belgium,
For Fluid Dynamics. January 1986.
[11] Kim CS, Kim C, and Rho OH. Sensitivity anal- [23] Reuther JJ, Jameson A, Alonso JJ, Rimlinger
ysis for the Navier–Stokes equations with two- MJ, and Saunders D. Constrained multipoint
equation turbulence models. AIAA J., Vol. 39, aerodynamic shape optimization using an ad-
No 5, pp 838–845, 2001. joint formulation and parallel computers, part 2.
[12] Kim H-J, Sasaki D, Obayashi S, and Naka- Journal of Aircraft, Vol. 36, No 1, pp 61–74,
hashi K. Aerodynamic optimization of super- 1999.
sonic transport wing using unstructured adjoint [24] Rumsey CL and Ying SX. Prediction of high
method. AIAA Journal, Vol. 39, No 6, pp 1011– lift: Review of present CFD capability. Progress
1020, 2001. in Aerospace Sciences, Vol. 38, No 2, pp 145–
[13] Kim S, Alonso JJ, and Jameson A. Design opti- 180, 2002.
mization of high–lift configurations using a vis- [25] Saad Y and Schultz MH. GMRES: A gener-
cous continuous adjoint method. AIAA Paper alized minimal residual algorithm for solving
2002–0844, January 2002. nonsymmetric linear systems. SIAM J. Sci. &
[14] Nelson TE, Godin P, De Rango S, and Zingg Stat. Comput., Vol. 7, No 3, pp 856–869, 1986.
DW. Flow computations for a three-element air- [26] Spalart PR and Allmaras SR. A one-equation
foil system. Canadian Aeronautics and Space turbulence model for aerodynamic flows. AIAA
Journal, Vol. 45, No 2, pp 132–139, 1999. Paper 92–0439, January 1992.
[15] Nelson TE, Zingg DW, and Johnston GW. Com- [27] Van Dam CP. The aerodynamic design of
pressible Navier–Stokes computations of mul- multi-element high-lift systems for transport
tielement airfoil flows using multiblock grids. airplanes. Progress in Aerospace Sciences,
AIAA J., Vol. 32, No 3, pp 506–511, 1994. Vol. 38, No 2, pp 101–144, 2002.
[16] Nemec M and Zingg DW. Newton–Krylov algo- [28] Van Den Berg B. Boundary layer measurements
rithm for aerodynamic design using the Navier– on a two-dimensional wing with flap. NLR TR
Stokes equations. AIAA J., Vol. 40, No 6, 79009 U, January 1979.
pp 1146–1154, 2002.
[17] Nemec M and Zingg DW. A Newton–Krylov al-
173.11