Please cite this document as:

J. R. R. A. Martins, Gaetan K. W. Kenway, and Timothy Brooks. Multidisciplinary Design Op-

timization of Aircraft Configurations—Part 2: High-fidelity aerostructural optimization. Lecture
series, Von Karman Institute for Fluid Dynamics, Sint–Genesius–Rode, Belgium, May 2016.
This document can be found at: http://mdolab.engin.umich.edu.

Multidisciplinary Design Optimization of

Aircraft Configurations
Part 2: High-fidelity aerostructural optimization
Joaquim R. R. A. Martins, Gaetan K. W. Kenway, and Timothy Brooks
University of Michigan

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Aerodynamic shape optimization framework . . . . . . . . . . . . . . . . . . . 2
2.1 CFD solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Geometric parametrization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3 Mesh movement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Automatic differentiation adjoint . . . . . . . . . . . . . . . . . . . . . . . . 5
2.5 Optimization algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3 Aerodynamic shape optimization applications . . . . . . . . . . . . . . . . . . 8
3.1 Common Research Model wing . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Full configuration with buffet constraints . . . . . . . . . . . . . . . . . . . . 12
4 Aerostructural optimization framework . . . . . . . . . . . . . . . . . . . . . . 21
4.1 Structural solver and adjoint . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.2 Load and displacement transfer . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3 Coupled solution algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.4 Coupled adjoint derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.5 Coupled adjoint solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5 Aerostructural optimization applications . . . . . . . . . . . . . . . . . . . . . 32
5.1 Multipoint wing design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.2 Tow-steered composite optimization . . . . . . . . . . . . . . . . . . . . . . . 37
6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

1
1 Introduction
When it comes to flight, the wing is arguably the most crucial component. As the legendary
Boeing aircraft designer Jack Steiner put it, “The wing is where you’re going to fail.”. In a
book detailing the origins of the Boeing 747 Jumbo Jet, Irving (1993) writes:

“Designing the wing involved literally thousands of decisions that could add up
to an invaluable asset, a proprietary store of knowledge. A competitor could look
at the wing, measure it even, and make a good guess about its internal structure.
But a wing has as many invisible tricks built into its shape as a Savile Row suit;
you would need to tear it apart and study every strand to figure out its secrets.”

These “invisible tricks” are a reflection of the complexity involved wing modeling and design,
where flexibility couples the aerodynamic shape to the structural layout and sizing. Thus it
is of the utmost importance to consider the coupling between aerodynamic and structural
models, not only in analysis, but also in design. The “thousands of decisions” cited in the
above quote can be mapped to design variables, which involve both aerodynamic shape,
and structural design variables. The aerodynamic flow is sensitive to the slightest change in
aerodynamic shape, so it is important to parameterize the aerodynamic shape and structural
sizing with a large number variables.
When optimizing both aerodynamics and structures, we need to consider the effect of the
aerodynamic shape variables and structural sizing variables on the weight, which also affects
the fuel burn. Thus, there are complex multidisciplinary trade-offs. Numerical optimization
is a powerful tool that can perform these trade-offs automatically. Aerospace engineering
researchers recognized this as soon as multi-physics models for wings were available, estab-
lishing the field of multidisciplinary design optimization (MDO) (Haftka, 1977; Martins and
Lambe, 2013).
In Part 1 of this lecture we introduced a general framework for the solution of multidis-
ciplinary system and assembling the coupled derivative using an adjoint method (MAUD).
Here, we cover the MDO of aircraft configurations based on high-fidelity models in more
detail, with emphasis on the coupling of the aerodynamics and structures disciplines. We
show a wing design example where we tackle the compounding challenges of modeling the
wing with high fidelity, while optimizing it with respect to hundreds of design variables. We
are able to do this successfully through the use of high-performance parallel computing, fast
solvers, state-of-the-art gradient-based optimization, and an efficient approach for computing
the coupled derivatives for the aerostructural solver.
We start by reviewing our computational fluid dynamics (CFD) solver and explain how
the adjoint for this solver was developed and implemented in Sec. 2. We then review two
aerodynamic shape optimization problems whose solution was enabled by this solver in
Sec. 3. In Sec. 4 we expand the problem by coupling the aerodynamics to a structural
solver, and develop the corresponding coupled adjoint equations. Several applications of this
aerostructural analysis and design optimization framework follow in Sec. 5.

2 Aerodynamic shape optimization framework

The aerodynamic shape optimization procedure is summarized in Fig. 1, which shows each
of the modules described in this section. The optimization loop relies on a gradient-based

2
optimization algorithm to decide on the new set of design variables x at each iteration. For
each new set of design variables, we must generate a new geometry and corresponding CFD
mesh, so that the flow solver can compute the drag, lift, and moment coefficients for the
new shape. The adjoint solver then computes the derivatives required by the optimization
algorithm to perform the next design optimization iteration. This framework has been
successfully used to perform several studies (Chen et al., 2016; Kenway and Martins, 2015b;
Lyu et al., 2015; Lyu and Martins, 2014; Mader and Martins, 2013, 2014), some of which we
review in Sec. 3.

Figure 1: Aerodynamic shape optimization iteration loop.

2.1 CFD solver

The CFD solver used in the work herein is SUMad, which is a finite-volume, cell-centered
multiblock solver for the compressible Euler, laminar Navier–Stokes, and RANS equations
(steady, unsteady, and time-periodic modes) (van der Weide et al., 2006). The spatial
discretization follows the Jameson–Schmidt–Turkel scheme (Jameson et al., 1981). Central
difference augmented with artificial dissipation is used for the discretization of the inviscid
fluxes and viscous fluxes use a central discretization. SUMad provides options for a variety of
turbulence models with one, two, or four equations and options for adaptive wall functions.
The Spalart–Allmaras model (Spalart and Allmaras, 1992) is used herein unless otherwise
noted.
The solver was originally developed as a Runge–Kutta (RK) solver employing well-known
steady-state acceleration techniques including local time-stepping, implicit residual smooth-
ing, and selective evaluation of viscous fluxes on alternate stages. For RANS analysis, the
turbulent equations are solved in loosely coupled fashion using a diagonally-dominant al-
ternating direction implicit (DD-ADI) scheme with implicit boundary condition treatment.
Acceptable load balancing on large processor counts is achieved using a built-in automatic
block splitting algorithm.
Recently, two additional solution techniques were added to the solver: A diagonalized

3
ADI method for the mean-flow equations, and a fully coupled Newton–Krylov (NK) solver.
Depending on the particular problem at hand, the DADI method is 3–5 times faster than
the RK method and the NK method is 8–25 times faster.
In practice, we use the RK solver at the start of a solution and then switch to NK for the
final convergence. This provides an excellent trade-off between robustness and computational
efficiency.
The steady state solution of the RANS equations can be expressed as finding the flow
state w such that the residuals of the discretized partial differential equations for each control
volume are zero, i.e.,
R(x, w) = 0, (1)
where x is the vector of design variables, which is fixed for each flow solution. Both R and
w are vectors of dimension (5 + nturb ) × ncells , where nturb is the number of turbulence model
equations, and ncells is the number of cells in the CFD mesh.

2.2 Geometric parametrization

In this work we use a free-form deformation (FFD) volume approach (Sederberg and Parry,
1986) that we implemented (Kenway et al., 2010) and have been using extensively for aero-
dynamic (Chen et al., 2016; Kenway and Martins, 2015b; Lyu et al., 2015; Lyu and Martins,
2014; Mader and Martins, 2013, 2014) and aerostructural optimization studies (Kenway
et al., 2014a,b; Kenway and Martins, 2014; Liem et al., 2015).
The FFD approach can be visualized as embedding the spatial coordinates defining a
geometry inside a flexible volume. The parametric locations corresponding to the baseline
geometry are found using a Newton search algorithm. Once the baseline geometry has been
embedded, perturbations made to the FFD volume propagate within the embedded geometry
via the evaluation of the nodes at their parametric locations. An example of an FFD for a
wing shape optimization is shown in Fig. 2.

Figure 2: The shape design variables are FFD control point z-displacements (red spheres).

2.3 Mesh movement

The FFD approach used to parametrize the geometry applies deformations only to the surface
mesh—the part of the volume mesh that lies on the physical surface. A mesh movement
procedure is then required to propagate surface perturbations to the remainder of the CFD
volume mesh. The mesh movement algorithm used in this work is an efficient analytic inverse
distance method similar to the one described by Luke et al. (2012). Updating the mesh for
a new configuration is fast, typically requiring less than 0.1% of the CFD solution time.

4
2.4 Automatic differentiation adjoint
As previously mentioned, accurate and efficient computation of derivatives of objective and
constraint functions with respect to large numbers of design variables is crucial in the effi-
ciency of the overall optimization.
In Part 1 of this lecture, we introduced the adjoint method, which computes derivatives
of a function of interest at a cost independent of the number of design variables. The
computational cost of the adjoint method scales with the number of functions of interest,
which is usually much lower than the number of design variables for aerodynamic shape
optimization problems.
As derived in Part 1, the discrete adjoint equations can be written as,
 T  T
∂R ∂F
ψ=− , (2)
∂w ∂w
where we replaced the general state vector y by the vector of flow state variables w. The
function of interest F , in the case of aerodynamic shape optimization, is usually the drag
coefficient as the objective. Additional adjoint equations must also be solved when F is a
constraint function that depends on the flow solution—usually lift and moment coefficients.
Once the adjoint vector ψ is computed, the total derivatives can be obtained using,
df ∂F ∂R
= + ψT . (3)
dx ∂x ∂x
Again, as explained in Part 1, the partial derivatives in the equations represent an explicit
dependence that do not require convergence of the residual (the flow solution in this case).
For example, if x is a set of wing shape variables, ∂R/∂x represents the sensitivity of the
residuals at each cell to changes in the mesh due to wing shape changes, where the flow
solution w is kept fixed.
Traditionally, the differentiation for these partial derivatives is done by hand, and the
code that computes them is generated manually, which requires a long development time
and is prone to errors. In some cases, researchers have also used finite differences and the
complex-step method. Our approach is to use automatic differentiation (AD) to produce the
code that computes these partial derivatives. As explained in Part 1, AD is a tool that given
the source code of a program, automatically generates code that computes the derivatives
of specified outputs to specified inputs. In its pure form, we would use AD on the whole
CFD code including the iteration procedure. Using this brute force approach blindly incurs
a large computational cost both in terms of memory and time, although there have been
efforts towards making this feasible (Albring et al., 2016). By selectively applying AD to
compute the terms in Eqs. (2) and (3), we strike a balance between computational efficiency
and development time.
This AD adjoint (ADjoint) approach was first demonstrated by Mader et al. (2008).
Since this demonstration, we have developed different ADjoint approaches and refined the
implementation to be both more automated, and more efficient. This idea was reinforced in
the survey on CFD-based derivative computation by Peter and Dwight (2010), who wrote
“It is our belief that, at least in the medium term, industrially relevant linearized codes
will be developed by using AD to differentiate individual nonlinear routines, which are then
assembled by an expert.”

5
 In our experience, the desirable features for an ADjoint approach are:

Consistency: The code must yield derivatives consistent with the flow solution and be
verifiable. Typically we verify it against the complex-step approach and hence are
able to find small errors. Such errors cannot be found when verifying against finite-
differences, and although small, the errors could become larger in different cases.

Maintainability: For ease of maintenance, the approach should not require modification
or duplication of the original code.

Efficiency: The ADjoint code should be efficient both in terms of memory usage and com-
putational time; it should also incur no penalty to the original CFD solution code.

Ease of implementation: The implementation should be as automatic as possible, with

short development time.

In the last decade, we have experimented with different ADjoint approaches that address
these feature to various degrees.

Single cell residual: This consists in using AD on a single cell residual routine, and loop
over cells to assemble the full Jacobian (Mader et al., 2008; Martins et al., 2006).

Forward mode coloring: The complete residual routine is differentiated with AD using
coloring for efficiency and the full Jacobian is stored. (Lyu et al., 2013).

Full reverse mode: A master routine is created such that when AD is applied to it, it
yields the desired Jacobian-vector products and derivatives.

Hybrid reverse mode: AD is applied to individual subroutines in the master routine and
the Jacobian-vector products are manually assembled using the chain rule.

The last two approaches are the most sophisticated and do the best in fulfilling the desirable
features listed above.
The single cell approach was the first demonstration of the ADjoint approach, and al-
though it was successfully used for aerodynamic shape optimization (Mader and Martins,
2013), it required code duplication, making it a poor approach from the maintainability
point of view. The forward mode coloring approach enabled us to develop a more efficient
adjoint that was much more maintainable because AD was applied to the original residual
computation routine. Using coloring techniques, we were able to reduce the computational
cost of forward mode AD (Lyu et al., 2013). However, this approach requires storing the
full Jacobian ∂R/∂w, which means that a direct solver can be used, but very large cases
might be memory limited, depending on the computer memory available. This approach was
used to solve several aerodynamic shape optimization problems (Lyu et al., 2015; Lyu and
Martins, 2014, 2015).
The last two approaches have the advantage that they do not require storing the flow
Jacobian ∂R/∂w. Instead, they yield the transpose matrix-vector product of the Jacobian
with a specified vector v, i.e., [∂R/∂w]T v. This product is then used to efficiently solve

6
the adjoint equations using an iterative solver, such as the generalized minimum residual
(GMRES) method.
The key to these two approaches is to create a routine that maps the design variables
and flow variables to the CFD residuals, as well as the functions of interest, i.e., (x, w) 7→
(R(x, w), F (x, w)). This mapping must not include iterations to solve the CFD equations,
only explicit dependencies. This master routine is not used directly, but is created for the
sole purpose of being differentiated with AD. The resulting AD code computes all the partial
derivatives needed in the adjoint equation (2) and the total derivative computation (3) using
reverse mode AD.
The adjoint linearization (2) includes both the mean flow equations as well as turbulence
model. This results in more accurate derivatives for design optimization but yields adjoint
systems that are ill-conditioned and thus more challenging to solve. We use the solvers in
PETSc (Balay et al., 1997) to solve the adjoint equations. The solution procedure is based
on a GMRES outer loop. By experimenting with the PETSc solution options, we found
that using a restricted additive Schwartz preconditioner with an overlap of 2 and an ILU
factorization with fill levels of 2 or 3 on each sub-domain results in the lowest overall CPU
time.
Strong scaling performance of both the flow solution (using the NK solver) and adjoint
solutions are shown in Fig. 2.4 for a coarse mesh (450k cells; typical for our exploratory
optimization) and a finer one (3.6M cells; typical for refining the optimized results). Both

3
10 Solution
Ideal
Adjoint

3.6M Grid
Time (s)

102

450k Grid

1
10

0 1 2 3
10 10 10 10
Processors

Figure 3: The RANS solver and adjoint solver exhibit strong parallel scaling.

meshes show good speedup for the NK flow solver through two orders of magnitude change
in processor count. The adjoint solver shows good scaling, although not quite as good as the
flow solver technique.

7
2.5 Optimization algorithm
The optimization algorithm we use for all the results presented herein is SNOPT (sparse
nonlinear optimizer) (Gill et al., 2002) through the Python interface pyOpt (Perez et al.,
2012). SNOPT is a gradient-based optimizer that implements a sequential quadratic pro-
gramming method; it is capable of solving large-scale nonlinear optimization problems with
thousands of constraints and design variables. SNOPT uses a smooth augmented Lagrangian
merit function, and the Hessian of the Lagrangian is approximated using a limited-memory
quasi-Newton method.

3 Aerodynamic shape optimization applications

We now present a few of the results that we have obtained with the aerodynamic shape
optimization framework detailed in the previous section. We start by summarizing the
solution of a benchmark RANS-based wing shape optimization problem in Sec. 3.1, and
then present results for a full aircraft configuration including trim and buffet constraints in
Sec. 3.2.

3.1 Common Research Model wing

Until recently, despite considerable research on aerodynamic shape optimization, there had
been no standard benchmark problems allowing researchers to compare results. The AIAA
Aerodynamic Design Optimization Discussion Group (ADODG) defined several benchmark
cases including a 3D RANS-based case. The baseline geometry for this case is a wing with a
blunt trailing edge extracted from the CRM wing-body geometry (Vassberg, 2011; Vassberg
et al., 2008).
The results presented herein are from Lyu et al. (2015), who have all geometries meshes
publicly available. The aerodynamic shape optimization seeks to minimize the drag coeffi-
cient by varying the shape design variables subject to a lift constraint and a pitching moment
constraint. The shape design variables are the z-coordinate movements of 720 control points
on the FFD volume (shown in Fig. 2) and the angle-of-attack. There are 750 thickness
constraints imposed in a 25 chordwise and 30 spanwise grid covering the full span and from
1% to 99% local chord. The thickness is set to be greater than 25% of the baseline thickness
at each location. Finally, the internal volume is constrained to be greater than or equal to
the baseline volume. The complete optimization problem is described in Table 3.1.
Function/variable Description Quantity
minimize CD Drag coefficient

with respect to α Angle of attack 1

z FFD control point z-coordinates 720
Total design variables 721

subject to CL = 0.5 Lift coefficient constraint 1

CMy ≥ −0.17 Moment coefficient constraint 1
t ≥ 0.25tbase Minimum thickness constraints 750
V ≥ Vbase Minimum volume constraint 1
∆zTE,upper = −∆zTE,lower Fixed trailing edge constraints 15
∆zLE,upper,root = −∆zLE,lower,root Fixed wing root incidence constraint 1
Total constraints 769

Table 1: Aerodynamic shape optimization problem

8
 Figure 4: Baseline CRM wing geometry scaled by its mean aerodynamic chord.

Figure 5 shows a detailed comparison of the baseline wing and the optimized wing. In
this figure, the baseline wing results are shown in red and the optimized wing results are
shown in blue. At the optimum, the lift coefficient target is met, and the pitching moment
is reduced to the lowest allowed value. The lift distribution of the optimized wing is much
closer to the elliptical distribution than that of the baseline, indicating an induced drag that
is close to the theoretical minimum for a planar wake. This is achieved by fine-tuning the
twist distribution and airfoil shapes.
The optimized thickness distribution is significantly different from that of the baseline,
since the thicknesses are allowed to decrease to 25% of the original thickness, and there is a
strong incentive to reduce the airfoil thicknesses in order to reduce wave drag. The volume is
constrained to be greater than or equal to the baseline volume, so the optimizer drastically
decreases the thickness of the airfoils on the outboard of the wing to the lower bounds,
where there is less volume to be gained, while increasing the thickness near the root (up to
20%), where the chords are larger and the volume-drag trade-off is more favorable. The low
outboard thickness would in practice incur a large structural weight penalty, and to trade
off the reduction in drag and increase in weight would require aerostructural optimization,
which we address in Sec. 5.
The baseline wing exhibits a front of closely spaced pressure contour lines spanning a sig-

9
 Figure 5: The optimized wing is shock-free and has 8.5% lower drag.

nificant portion of the wing, indicating a shock. The optimized wing shows parallel pressure
contour lines with uniform spacing, indicating a shock-free solution under the nominal flight
condition. This is confirmed by the shock surface plots: we can see that the baseline wing
has a shock on the upper surface, while the optimized wing does not show shocks under the
design condition. The shock elimination can also be seen on the airfoil Cp distributions. The
sharp increase in local pressure due to the shock becomes a gradual change from the leading
edge to the trailing edge.
To ensure that the result of our single-point optimization has sufficient accuracy, we
conducted a grid convergence study of the optimized design. The mesh convergence plot for
both the baseline and optimized geometry meshes is shown in Fig. 6. The zero-grid spacing
drag, which was obtained using Richardson’s extrapolation, is also plotted in the figure. We
can see that the L0 mesh has sufficient accuracy: the difference in the drag coefficient for
the L0 mesh and the value obtained for the zero-grid spacing is within one drag count. The
variation in drag coefficient between the baseline and optimized meshes is nearly constant
for each grid level, which gives us confidence that the optimization using the coarse meshes
represent the design space trends sufficiently well. Therefore, we perform the remaining
optimization studies on the coarser mesh (L2), assuming that we capture the correct design
trends.
Using this benchmark, Lyu et al. (2015) determine that there is no significant multi-
modality by showing that the optimizers always converges to the same geometry even when
starting from a wide variety of starting wing shapes. One such example is shown in Fig.7,
where the starting geometry is a random perturbation of the baseline geometry, and the final

10
 L2
0.0210

L1
baseline
L00 L0
0.0200

zero-grid spacing drag

CD
0.0190

optimized

0.0180 231M 28.8M

3.6M 451k

-0.1700
CMy

-0.1800

-0.1900
0 5E-05 0.0001 0.00015
1/GRIDSIZE2/3

Figure 6: The mesh convergence study shows that the difference between the drag value
computed with the 28.8 M grid and the zero-grid spacing drag is within one count.

result was the same as in Fig. 5.

Figure 7: The optimization manages to start from a random geometry and converge to an
optimal wing that is shock free.

11
 Lyu et al. (2015) present many other results for other problems related to this benchmark,
including multipoint optimizations. Further multipoint optimization studies are performed
by Kenway and Martins (2016b).

3.2 Full configuration with buffet constraints

For transonic wing design, aerodynamic shape optimization has demonstrated the potential
produce high performance designs, but these designs are susceptible to buffet. To address
this issue, we present a separation-based constraint formulation that can be used to constrain
buffet onset in an aerodynamic shape optimization (Kenway and Martins, 2016a). A series
of optimizations based on the AIAA ADODG wing-body-tail case are presented to show
the need for buffet-onset constraints, and demonstrate the effectiveness of the proposed
approach.
To quantify the buffet onset in aerodynamic shape optimization, we develop a new pre-
diction method based on the amount of separated flow present on the wing in a steady
RANS. We then use this method to enforce a constraint on the amount of separated flow
near the buffet onset boundary, which ensures that the optimized design has a sufficient
buffet margin, while simultaneously improving the performance at the design operating con-
ditions. The constraint ensures that the design stays within the boundary defined by the
30% margin to buffet onset airworthiness requirement. The constraint function is smooth
and its gradient with respect to the wing shape variables is computed used a discrete adjoint
implementation. We demonstrate the effectiveness of our new approach on the multipoint
drag minimization of the wing-body-tail geometry (Kenway and Martins, 2015a).
To develop a more direct way of constraining buffet onset, we analyze the physical mech-
anism that causes the loss of lift and the subsequent lowering of the lift curve slope is the
initial appearance of shock-induced flow separation. An example showing the typical pro-
gression of this type of separation with increasing angle of attack is shown in Fig. 8. The
first row in Fig. 8 shows the surface streamlines and pressure coefficient, as well as the shock
sensor in orange (Lovely and Haimes, 1999).
To determine if the flow is separated at a given location on the surface, we check if the
surface flow velocity has a component in the negative freestream direction (approximately
the negative x-axis) direction, i.e., if

V~ · V~∞
cos θ = < 0, (4)
|V~ ||V~∞ |

where θ is the angle between the local surface velocity and the freestream. Then we can
define a separation sensor as,
(
1 if cos θ ≤ 0
χ= (5)
0 if cos θ > 0

Thus χ is specific to each surface location and is a Heaviside function: It is equal to one
when the flow is separated, and equal to zero when the flow is attached. The red areas on
the surface for α = 3 ◦ and α = 3.29 ◦ in the bottom row of Fig. 8 shows the regions where
χ = 1, and they are an approximation to areas where the flow is separated.

12
Figure 8: Progression of separated flow for the CRM configuration at M = 0.85 with in-
creasing angle of attack. Top row shows the surface streamlines and pressure coefficient, as
well as reversed flow (red) and the shock (orange). Bottom row shows the separation area
integrand value from Eqn. (7).

Our hypothesis is that we can correlate the value of the area where χ = 1 to buffet onset,
which is given by the integral of χ over the whole surface area of the wing. Since we need
to use this function as a constraint in an gradient-based optimization, we would like this
function to be smooth, so we use a smooth Heaviside function to blend the discontinuity as
follows,
1
χ̄ = 2k(cos θ+λ)
(6)
1+e
In the this equation, k and λ are free parameters, where k determines the sharpness of the
transition, and λ is a parameter that can be used to shift the smoothing function to the
left or right as a function of the angle. For our cell-centered solver, the values for V~ are
taken from the state variables at the cell center immediately adjacent to the wall, since the
velocities at the wall are zero when enforcing the no-slip condition.
Now, we can integrate the smooth separation sensor (6) over the surface and normalize
it by the aircraft reference area to obtain the proposed separation metric:
Z
1
Ssep = χ̄ dS. (7)
Sref S
This is equivalent to performing a weighted area integration of the sensor value shown in the
bottom row of Fig. 8. We have shown that the separation metric (7) is correlated to buffet
onset, by comparing to another method, as well as experimental data (Kenway and Martins,
2016a).
The baseline geometry defined in the ADODG Case 5 is taken directly from the 4th
Drag Prediction Workshop (Vassberg et al., 2014), and is shown in Fig. 9. A sequence
of seven design optimizations are solved to study the aerodynamic shape optimization of
the ADODG full CRM configuration, and to demonstrate the effectiveness of the proposed
approach in satisfying buffet requirements. These cases—numbered 5.1 through 5.7—are
summarized in Table 3. The objective of all optimizations (except for Case 5.7, which is

13
discussed separately) is to reduce the weighted drag coefficient at the N defined operating
conditions. The optimization problem statement can be written as:
N
X
minimize Wi CDi Quantity
i=1
with respect to Wing cross sectional shape 240
Wing twist 9
Angle of attack (αi ) N (8)
Tail rotation angle (ηi ) N
subject to CLi − CL∗ i = 0.0 N
CMyi = 0.0 N
tj ≥ tjCRM 750
Ssepi ≤ 0.04 N
Only Cases 5.4, 5.5, and 5.7 use the separation constraint, where the constraint is applied
only to the last two operating conditions. In practice, the adjoint for CD is not evaluated for
the buffet onset conditions, and conversely the separation metric adjoint is not evaluated for
the normal operating conditions. This results in a total of three adjoint solutions required for
each flight condition, which is desirable from a load balancing perspective. The constraints
that the wing thickness may not be less than the baseline configuration at any location is
particularly restrictive.
The ADODG specification for Case 5 requires the parameterization not to modify the
planform, and any shape modification must be made only in the vertical direction. Addi-
tionally, twist rotation is permitted for the wing, as well as a solid-body rotation of the
horizontal tail for trimming the aircraft. We use the FFD approach described in Sec. 2. The
FFD volume and the associated geometric design variables are shown in Fig. 9.

Figure 9: Shape and twist modifications are permitted for the wing, and the horizontal tail
can rotate.
Figures 10 and 11 show a summary of the key features of the two ADODG optimizations
(Cases 5.1 and 5.2). The baseline configuration results are shown in red, while the optimized

14
 Table 2: Operating conditions for each optimization case
Case Point Weights (Ti ) Mach CL Re M –CL plot
0.7

5.1 1 1 0.85 0.500 43.00 × 106 0.6

CL
0.5

0.4
0.8 Mach 0.85 0.9

1 6 0.7
5.2 1 0.85 0.500 43.00 × 10
3
1 0.6
43.00 × 106

CL
2 0.85 0.650
3 0.5
1
3 0.89 0.456 45.00 × 106 0.4
0.8 Mach 0.85 0.9
3

2
5.3 1 0.85 0.500 43.00 × 106 0.7

3
1 0.6
43.00 × 106

CL
2 0.85 0.650
6 0.5
1
3 0.89 0.456 45.00 × 106 0.4
0.8 Mach 0.85 0.9
6

5.4 1 1 0.85 0.500 43.00 × 106 0.7

2 0 0.85 0.650 43.00 × 106 0.6

CL
3 0 0.89 0.456 45.00 × 106 0.5

0.4
0.8 Mach 0.85 0.9

1 6
5.5 1 0.845 0.490 42.75 × 10
5
1 0.7
2 0.845 0.445 42.75 × 106
5 0.6
1

CL
3 0.845 0.408 42.75 × 106 0.5
5
1
4 0.835 0.467 42.24 × 106 0.4
0.8 Mach 0.85 0.9
5
1 6
5 0.855 0.418 43.25 × 10
5
6 0 0.85 0.650 43.00 × 106
7 0 0.89 0.456 45.00 × 106
1
5.6 1 0.845 0.490 42.75 × 106 0.7
5
1 6 0.6
2 0.845 0.445 42.75 × 10
CL

5 0.5
1
3 0.845 0.408 42.75 × 106 0.4
5 0.8 Mach 0.85 0.9
1
4 0.835 0.467 42.24 × 106
5
1
5 0.855 0.418 43.25 × 106
5
1
5.7 1 0.845 0.520 42.75 × 106
5
1 0.7
2 0.845 0.475 42.75 × 106
5 0.6
1
CL

3 0.845 0.438 42.75 × 106 0.5

5
1 6 0.4
4 0.835 0.497 42.24 × 10 0.8 Mach 0.85 0.9
5
1 6
5 0.855 0.448 43.25 × 10
5
6 0 0.85 0.690 43.00 × 106
7 0 0.89 0.469 45.00 × 106

results are shown in blue. Case 5.2 adds two additional equally-weighted operating conditions
near the buffet onset boundary. Unlike Case 5.1, for which we obtain a shock-free wing,
Case 5.2 results in double shocks at the nominal operating condition. In this case, the drag

15
at the nominal operating condition actually increased by 2.8 counts. The drag divergence
curves indicate a significant drag penalty across the lower Mach numbers but it does have a
much higher drag divergence Mach number than the baseline design.

Figure 10: High performance is achieved for the single point design at the nominal operating
condition.

While drag coefficient divergence curves yield useful insights into optimized designs, it
is particularly instructive to examine the 2D analogue in the form of M L/D contours. In
the context of transonic transport wing design, M L/D is a better measure of performance,
since it includes the benefit that flying faster has on the overall aircraft efficiency.
Contour plots for all seven optimizations including the baseline design are shown in
Fig. 12. The contours in each figure extend up to the predicted buffet onset curve shown
in red. The orange curve shows the prediction buffet onset using the ∆α = 0.1 method.
Overall, the separation metric method continues to compare well with the ∆α = 0.1 method
despite the large changes in the buffet onset boundary.
The blue curve represents the 30% margin to buffet onset boundary and is computed
directly from the red buffet onset curve. Only operating conditions below the buffet margin
curve can be considered for normal operation. The absolute maximum M L/D value for

16
Figure 11: Performance is sacrificed across a large range of Mach numbers to obtain a small
amount of improvement at the highest Mach numbers.

each configuration is shown in pink. Two specific contours for the optimization configura-
tion (one for the baseline configuration) are highlighted: the contour of 99% (M L/D)max for
the particular design is shown in blue and the contour of 99% (M L/D)max of the baseline
configuration is shown in red. The motivation for plotting these 99% contours is that airlin-
ers typically fly between the Mach number yielding maximum range (approximated by the
maximum M L/D value in the figures) and a higher Mach number that yields a 1% fuel-burn
penalty, but also decreases in the flight time. The area enclosed by both of these contours
is used to quantify the design’s robustness in these figures. The areas are scaled by a factor
of 1002 such that the area of the rectangle measuring 0.01 in M and 0.01 in CL has an area
of 1.
The design operating conditions listed in Table 2 are shown as diamonds. The operating
conditions considered for the objective function are shown in black, while the buffet onset
constraint conditions are shown in red. The first point (M = 0.85, CL = 0.65) is at the
nominal cruise Mach number and the CL value corresponding to a 1.3 g maneuver. The
second buffet point (M = 0.89, CL = 0.456) is 0.04 higher in Mach number, which is a typical
margin between a nominal cruise Mach number and the maximum Mach number (MM O )

17
condition. The lift coefficient for this condition is adjusted to give the same dimensional lift
as the nominal cruise condition.
There are two additional regions that are highlighted in black and orange. We refer to
these as integration regions. They are constructed as follows: The Mach range is from 0.83 to
0.86, which corresponds to the typical range of operating Mach numbers for an aircraft like
the CRM. The upper line corresponds to the buffet margin boundary, which is equivalent to
specifying the maximum altitude the aircraft can fly for a particular weight. The bottom line
corresponds to the reduced CL for a 4 000 ft decrease in altitude. To put it in another way, the
integration region contains all operating conditions within 4 000 ft of the buffet-constrained
ceiling and at all normal operating Mach numbers.
The contour plots give a much more complete understanding of the optimized designs.
Unsurprisingly, the single point optimization—Case 5.1—was able to produce the highest
M L/D value, which is almost exactly matched to the design operating condition. However,
without a way to constrain the buffet onset boundary, the (M L/D)max is now above the
buffet margin boundary, which means that this high performance point cannot be achieved
in practice because it falls outside the normal flight envelope. The 99% (M L/D)max contour
(blue) is small, indicating a highly localized point-design. Despite the highest M L/D value,
the average M L/D in its own integration region (orange) is 4.6% worse than the baseline
design.
For Case 5.2, the addition of operating conditions at the edge of the buffet onset envelope
has the effect of substantially improving the buffet boundary over the entire range of Mach
numbers. This case resulted in the most robust buffet onset behavior of all cases. However,
the (M L/D)max has barely improved over the baseline design (17.18 vs. 17.13). Worse, as in
Case 5.1, the high performance region lies almost entirely outside the buffet margin boundary,
rendering the high performance region unattainable. Even for this case, the M L/D average
in the integration region is slightly worse (−0.5%) than the baseline design. Note that the
increased performance afforded by the higher buffet boundary is only possible if the baseline
aircraft is buffet limited in altitude over the specific range of Mach numbers, as opposed to
thrust limited. If the aircraft were thrust limited over the integration range, the obtainable
performance would be the value obtained by integration of the black integration region.
In Case 5.3, we attempt to improve upon Case 5.2 by reducing the weighting factor for
the near-buffet conditions. For this case, the nominal operating condition has weight of 2/3,
while the remaining two points have weights of 1/6 each. The adjusted weightings yield a
much more useful design. This is the first case where a significant portion of 99% (M L/D)max
contour falls within the integration region. Additionally, the design is robust, as evidenced
by the larger area enclosed by the blue contour when compared to the baseline design. As
with the two previous cases, the increased performance is only possible if the aircraft is able
to operate at higher altitudes. The other issue with this case is that the specific weightings
were arbitrarily picked. These particular weights yielded acceptable results, but there is
no guarantee these weight values would work well for another configuration or optimization
problem.
Case 5.4 is the first optimization to use the separation sensor directly as an optimization
constraint. Case 5.4 retains the same operating conditions as Cases 5.2 and 5.3, but instead
of having the drag from the flight conditions near the buffet boundary contribute to the
drag objective function, it uses the buffet onset flight conditions to compute the separation

18
Figure 12: Contours of M L/D for the baseline and each optimized configuration.
19
sensor and constrain its value. Note that there is a slight discrepancy between the operating
conditions (red diamonds) and the buffet onset boundary itself. The reason is that the buffet
onset conditions are analyzed using the JST scalar dissipation scheme, which results in a so-
lution with more dissipation. The scalar scheme provides the increased robustness necessary
for optimization, which is unnecessary for the contour plot evaluations. The more dissipa-
tive scalar scheme slightly under-predicts the area of separated flow, and thus the buffet
boundary is lower when analyzed with the matrix scheme for the contour plot. Overall, the
performance of this design is similar to that of the single-point optimization (Case 5.1). Most
of the high performance region lies outside the integration region. However, the performance
reduction is not as pronounced as with Case 5.1, with a 1.6% performance reduction in the
original integration region and almost zero for the on-design integration region. However,
there is a small improvement in the buffet onset boundary.
After analyzing the results from Cases 5.1–5.4, we noticed that the nominal design point
was always located towards the upper side or even completely outside the integration re-
gion. It is perhaps no surprise that all optimization discussed thus far failed to improve
the performance in the baseline integration region. To address this issue, we formulated a
multipoint optimization, Case 5.5. Previous optimizations performed by the authors on the
CRM wing along configuration have shown that five operating conditions arranged as a cross
in M –CL space results is highly robust designs (Kenway and Martins, 2016b). Since our goal
is to improve the performance in the original integration region, we distribute five operating
conditions as follows. The nominal Mach number is reduced to 0.845 for the first three op-
erating conditions. The first point lies on the 1.3 g buffet onset boundary, while the second
two points are at CL values corresponding to 2 000 and 4 000 ft below. The two remaining
points are 2 000 ft lower than the buffet onset boundary at ±0.01 Mach number. The buffet
onset conditions are taken at M = 0.85, CL = 0.65 and M = 0.89, CL = 0.41. The later
point is taken from the baseline design buffet onset boundary. The overall performance of
this case is superior to all previously discussed cases. The performance in the baseline inte-
gration region has now increased by 1.2%, the performance of the updated integration region
increase by 2.0%. The design is very robust, as shown by area inside the 99% (M L/D)max
contour. Further, the maximum performance point is located inside the operating envelope.
Next we developed Case 5.6 to investigate the effect of removing the buffet onset condi-
tions present in Case 5.5. We wish to answer the question: Is a multipoint optimization near
the design operating condition sufficient to ensure a robust buffet onset envelope? Unsur-
prisingly, without the buffet onset constraints, the buffet margin boundary dropped slightly
over the integration envelope, pushing the integration region into a lower performance region.
The average M L/D for the integration region is 16.88, only 0.6% higher than the baseline
configuration. This is much smaller than the 2.0% improvement obtained in Case 5.5.
For the last case, Case 5.7, we formulate a different design optimization problem. We wish
to remove the requirement of specifying fixed design lift coefficients and let the optimization
itself determine the ideal on-design condition. All cases presented thus far are lift-constrained
drag minimizations with fixed operating conditions. The fixed operating conditions also
include fixing the CL for the buffet onset locations. In the formulation of Case 5.7, we want
the optimization to adjust the single nominal operating condition directly. The remaining
operating conditions are then explicitly linked to this design CL . More specifically, the high
CL buffet onset conditions must have 1.35 times the lift of the nominal cruise Mach. The

20
high Mach buffet case must have the same physical lift as the nominal operating condition
at M = 0.89. Finally, the remaining operating conditions move vertically in sync with the
changing design CL . The modified optimization formulation is as follows:
N
X
maximize Wi Mi Li /Di Quantity
i=1
with respect to Wing cross sectional shape 240
Wing twist 9
Angle of attack (αi ) N
Tail rotation angle (ηi ) N (9)
Design CL 1
subject to CLi − CL∗ i = 0.0 N
CMyi = 0.0 N
tj ≥ tjCRM 750
Sepi ≤ 0.04 N

Note that operating conditions (diamonds) shown in Fig. 12 are the optimized values.
The optimization increased the nominal design CL from the initial value of 0.490 (the same
value as used in Cases 5.5 and 5.6) and increased it to 0.520, an increase of 0.03 in CL . This
increase is only possible due to a corresponding increase in the buffet onset boundary. The
previous optimizations, especially Case 5.2, showed that there can be a significant penalty
in cruise drag for a higher buffet boundary. For Case 5.7, we have given the optimizer
sufficient information to make this trade-off optimal. This results in a slightly higher average
performance than Case 5.5 (17.13 vs. 17.11), as well as a higher buffet onset boundary. The
design is also highly robust, exhibiting the largest 99% (M L/D)max contour of all cases.

4 Aerostructural optimization framework

Coupling aerodynamic and structural numerical models to compute the static aeroelastic
shape of lifting surfaces is essential when designing lifting surfaces that are flexible. Even
small changes in shape can have a large effect on the aerodynamic performance, and multiple
flow conditions result in multiple shapes. This is particularly important for swept wings,
where bend-twist coupling can result in large changes in the twist distribution. In this section
we extend the aerodynamic shape optimization framework by coupling it to a structural
solver, and demonstrate the developed approach in two wing design problems.

4.1 Structural solver and adjoint

The structural solver used in this work is the Toolkit for the Analysis of Composite Struc-
tures (TACS) (Kennedy and Martins, 2014). TACS includes an adjoint solver that is able
to handle the structural design variables, which in our case are the thicknesses of the struc-
tural members. Parallelism is achieved within TACS by using an element-based partitioning
of the finite-element mesh. This partitioning is used to parallelize the factorization of the
stiffness matrix, the computation and assembly of the stiffness matrix and structural resid-
uals, and the computation of the functions of interest and their derivatives. Typically, the

21
factorization of the stiffness matrix is the most costly operation. To parallelize the ma-
trix factorization and back solutions, TACS uses a Schur-complement-based parallel direct
solver. In this technique, each processor independently computes the local contribution to a
reduced linear system that is formed from the all unknowns on the domain interface. This
reduced problem is the global Schur complement. In TACS, the local contributions to the
global Schur complement are computed using a block-based parallel factorization. After the
global Schur complement is computed, it is factored in parallel using a sparse block-cyclic
algorithm that achieves excellent parallel performance. This direct method enables us to
solve poorly conditioned thin-shell structural problems with condition numbers O(109 ) in an
efficient manner. For general nonlinear analysis the structural discipline residuals are
S (u) = 0 (10)
where u is the vector of structural displacements. For linear analysis, this equation can be
written as S (u) = Ku − F , where K is the linear stiffness matrix and F is the load vector.

4.2 Load and displacement transfer

Our load and displacement transfer scheme follows the work of Brown (1997). In this ap-
proach, rigid links are used to extrapolate the displacements from the structural surface to
the CFD surface, as shown in Fig. 13. These rigid links are constructed between the aerody-
namic surface mesh points and the points on the structural model lying closest to this set of
points. The consistent force vector is determined by employing the method of virtual work,
ensuring that the force transfer is conservative. The integration of the forces is performed
on the aerodynamic mesh and is transmitted back through the rigid links to the structure.
ur × r Displaced aerodynamic surface, XS
Displaced
Structural element
r

ur
uA

Structural element ut Jig aerodynamic surface, XJ

r
Rigid link

Figure 13: Load-displacement transfer operation

The design jig shape, XJ , is uniquely determined by the optimization design variables x.
A single coordinate on the perturbed surface, XS , is given by
XS = XJ + uA = XJ + ut + ur × r (11)

22
where r is the vector that connects a point on the aerodynamic surface to the closest location
on the structural mesh, ut is the translational component of the displacement, and ur is the
small-angle approximation of the rotations in the global reference frame, as shown in Fig. 13.
We can compactly represent the displacement transfer from u to uA as
 
∂uA
uA = T u = u (12)
∂u
where T is the generalized transfer matrix. To transfer loads from the aerodynamic discipline
to the structural discipline, we employ the transpose operation,
 
T ∂F
F = T FA = FA (13)
∂FA
where FA are the forces on the CFD nodes. More details on these transfers can be found in
Brown (1997) and Martins et al. (2005).

4.3 Coupled solution algorithms

We formulate the aerostructural analysis problem using a two-field formulation. The aero-
dynamic analysis is the first field and the structural analysis is the second. Using this
formulation, we can write the governing equations of both disciplines as a function of the
fluid states, w, structural states, u, and design variables x. The latter is a vector consisting
of global variables and local variables. Global variables affect the two disciplines directly,
while local variables affect only a single discipline in a direct manner. Geometric variables
that change the aircraft wetted surface—or outer mold line (OML)—are global variables. For
example, an airfoil thickness variable will not only change the aerodynamic shape but also
the height of the ribs and spars in the internal structure. Local variables include the angle
of attack, which directly affects only the aerodynamics, and the thicknesses of the structural
members, which affect only the structures. Combining the residual equations from the aero-
dynamic and structural disciplines, we write the combined residual of the multidisciplinary
system as " #
A (w, u; x)
R= = 0. (14)
S (u, w; x)
The aerostructural analysis consists in finding a solution, (w, u), that satisfies these coupled
residual equations.

4.3.1 Nonlinear block Gauss–Seidel method (NLBGS)

The traditional process for solving the coupled aerostructural equations (14) is to use a
nonlinear block Gauss–Seidel (NLBGS) method (Martins et al., 2005; Maute et al., 2001).
In this approach, the aerodynamic analysis is first partially converged and the aerodynamic
forces are evaluated. The forces are then transferred to the structural analysis and the
corresponding displacements are computed. Finally, the displacements are transferred back
to the aerodynamic analysis, the mesh is deformed, a new CFD solution is found, and
this iterative loop continues until the coupled convergence criterion is met. The NLBGS
procedure is listed in Algorithm 1.

23
Algorithm 1 Nonlinear block Gauss–Seidel method
1: Given: w(0) , u(0) , XJ , kmax
2: Initialize: θ ∈ (0, 1]
3: for k ← 1, kmax do
(k) (k−1)
4: XS ← T u +  XJ . Transfer displacements from CSM to CFD
(k) (k)
5: XV ← W XS . Deform volume mesh to match surface
 
(k)
6: A(k) ← A w(k−1) , XV . Evaluate initial CFD residual
7: while kA (wi ; x)k > A A(k) do . CFD partial convergence criterion
(k−1) (k−1)
8: wi+1 ← wi + ∆wi . Iterate CFD
9: i←i+1
10: end while
(k−1)
11: w(k) ← wi   . New CFD solution
(k) (k)
12: FA ← FA w(k) , XS . Evaluate aerodynamics forces
(k)
13: F (k) ← T TFA  . Transfer forces from CFD to CSM
14: S (k) ← S u(k−1) , F (k) . Evaluate initial CSM residual
 
15: while S u(k) ; x > S S (k) do . CSM partial convergence criterion
(k−1) (k−1)
16: ui+1 ui← + ∆ui . Iterate CSM
17: i←i+1
18: end while
(k−1)
19: u(k) ← uf . Temporary new CSM solution
(k) (k) (k−1)
20: ∆u ← u − u . Compute displacement increment
21: if k > 1 then !
∆u(k) − ∆u(k−1) · ∆u(k)


22: θ ←θ 1−
. Adapt under-relaxation with Aitken
k ∆u(k) − ∆u(k−1) k2
acceleration
23: end if
24: u(k+1) ← u(k) + θ∆u(k) . Compute new structural states
(k) (1) (k) (1)
25: if A < AS A and kS k < AS kS k then . Aerostructural convergence
criterion
26: break
27: end if
28: end for

Three tolerances are defined for the coupled analysis. The relative tolerances, A and
S , are the tolerances required by the aerodynamic and structural disciplines, respectively,
The aerodynamic solver tolerance is typically O 10−1 , while


for each NLBGS iteration.
S is typically O 10−3 or smaller. The third tolerance, AS , is the aerostructural solution

tolerance and represents the feasibility of the interdisciplinary coupling. Typical values for
AS range from 10−3 for an engineering solution accurate to three decimal places to 10−6 ,

24
which is the typical value used when performing design optimization. Aitken acceleration
(Irons and Tuck, 1969) (Line 22 of Algorithm 1) is employed to dynamically choose the
under-relaxation factor to accelerate convergence. One advantage of this method is that
each disciplinary solver can be used without modification. For tightly coupled aerostructural
problems with large displacements, however, this method may converge slowly or not at all.

4.3.2 Coupled Newton–Krylov method (CNK)

The second approach that we use for the aerostructural solution is a fully coupled Newton–
Krylov (CNK) method. Algorithm 2 lists the pseudocode for evaluating the coupled nonlinear
residual, R.

Algorithm 2 Coupled nonlinear residual computation

1: function R(w,u)
2: XS ← T u + XJ . Transfer displacements
3: XV ← W (XS ) . Deform volume mesh to match surface
4: A ← A (w, XV ) . Evaluate CFD residuals
5: FA ← FA (w, XS ) . Evaluate aerodynamics forces
6: F ← T T FA . Transfer forces
7: S ← S(u, F ) . Evaluate CSM residuals
8: R ← (A, S) . Combine residuals
9: return R
10: end function

The procedure is similar to one iteration of the NLBGS method, except that instead of
computing an approximate solution update to the state variables, we evaluate only resid-
uals. We use an inexact Newton–Krylov approach to solve the coupled equations (14) by
computing the approximate Newton update,
∂A ∂A "
 
# " #
 ∂w ∂u  ∆w A (w)
 ∂S ∂S  =− . (15)
∆u S (u)
∂w ∂u
After this update is computed, the state variables w and u are updated with ∆w and ∆u,
respectively.
To reduce the memory requirement, we use a matrix-free method. The matrix-vector
products required by the Krylov method are approximated with finite differences. In theory,
with this approach, we need only residual evaluations to solve the coupled problem. However,
Krylov methods require effective preconditioning for acceptable performance, especially for
very large systems with millions of degrees of freedom. We have implemented a block-Jacobi
preconditioner that reuses the preconditioner of the original discipline solvers. This has two
advantages: 1) the preconditioners for each discipline can be applied in parallel, because this
approach eliminates contributions from the off-diagonal terms, which are never explicitly
stored, and 2) since both of our discipline solvers use implicit methods, this preconditioner
simply reuses the same linear solution method from the original solvers. Since the coupled

25
preconditioning method varies from iteration to iteration, we must use a flexible variant of a
Krylov method to solve the coupled Newton update (15). We choose to use FGMRES (Saad
and Schultz, 1986), which has been shown to work well for a wide range of large asymmetric
systems.

4.4 Coupled adjoint derivatives

The critical component of the aerostructural framework is an efficient method for the com-
puting coupled derivatives. Using the approach outlined by Martins et al. (2005), we write
the adjoint equations for the coupled aerostructural system. This is a specific case of the
unifying chain rule presented in Part 1 of this lecture. The residuals, state variables, and
adjoint variables for the aerodynamic discipline are A, w, and ψ, respectively, and the cor-
responding variables for the structural discipline are S, u, and φ. The total derivative of the
function of interest, f , is
 " #
∂f ∂f ∂f dwx
df x = + . (16)
∂x ∂w ∂u dux

We write the total derivative of the residuals as

∂A ∂A ∂A "
   
" # #
dAx  ∂x   ∂w ∂u dwx
=  ∂S  +  ∂S ∂S  dux = 0. (17)

dSx
∂x ∂w ∂u
Substituting the solution of Eq. (17) into Eq. (16) to eliminate the total derivatives we obtain

 ∂A ∂A −1 ∂A
   

∂f ∂f ∂f  ∂w ∂u   ∂x 
df x = −  ∂S ∂S   ∂S  . (18)
∂x ∂w ∂u
| {z∂w ∂u } ∂x
ΨT

This yields the coupled adjoint equations,

∂A ∂A T " # 
 
T
 ∂w ∂u  ψ ∂f ∂f
 ∂S ∂S  = . (19)
φ ∂w ∂u
∂w ∂u
After the solution for the coupled adjoint equations (19) is obtained, the following equation
can be used to compute the total derivative:
   
∂f T ∂A T ∂S
df x = −ψ −φ . (20)
∂x ∂x ∂x

Implementing the coupled adjoint, including all the required partial-derivative terms, is
a challenging endeavor. Furthermore, ensuring that the partial-derivative computations and
solution methods are efficient and exhibit good parallel scalability is even more difficult.

26
 The aerodynamic diagonal block, ∂A/∂w, in the coupled adjoint equations (19), repre-
sents the Jacobian that we already explained in Sec. 2.
The first off-diagonal block, ∂A/∂u, contains the derivatives of the aerodynamic residual
with respect to the structural displacements. This is a challenging matrix to compute in
a two-state aerostructural formulation. What makes this term particularly challenging is
that a single structural degree of freedom (DOF) can affect many aerodynamic cells. Using
the mesh deformation described in Sec. 2.3, any structural DOF that perturbs a CFD node
results in a dense column due to the solution of the linear system of equations. With a
large number of structural surface DOFs and a large number of CFD nodes, storing this
matrix would require an excessive amount of memory. Instead, we compute this term in a
matrix-free fashion using the chain rule:
 T  T  T  T
∂A ∂XS ∂XV ∂A
ψ= ψ (21)
∂u ∂u ∂XS ∂XV

where ∂A/∂XV contains the derivatives of the aerodynamic residual with respect to vari-
ations of all the volume mesh coordinates. The sparse structure of this matrix is due to
the stencil used for the finite-volume computation. We compute this matrix using reverse-
mode AD, in a similar fashion to ∂A/∂w. This matrix is computed in parallel and stored in
distributed memory, which allows us to compute the transpose matrix-vector products in a
relatively inexpensive manner.
The matrix of derivatives of the mesh deformation, ∂XV /∂XS , requires a careful im-
plementation to be computationally efficient. A summary of the full mesh-deformation
procedure is shown in Fig. 14.

Bézier
YS = RXS Kuu YV − Kus YS = 0 interpolation
XS YS YV XSW XV

Transfinite interpolation

Figure 14: Solid mesh warping procedure

For the transpose matrix-vector products required for the adjoint computation, the nor-
mal deformation procedure of Fig. 14 is done in reverse. Reverse-mode AD is used to compute
the variation of all the surface nodes (XS ) and the variation of the volume supernodes (YV ).
The variation of the volume supernodes is related to the surface supernodes through the
solution of the mesh adjoint equation given by
T
Kuu ψM = −YV , (22)

where ψM is the mesh adjoint. Finally, the variation YS is propagated to the full mesh surface
using the transpose of the restriction operator. This procedure is illustrated in Fig. 15.
This computation is performed just once for each transpose matrix-vector product. To
ensure overall scalability of the coupled adjoint solution, all of the computations are executed
in parallel with acceptable efficiency. In this case, since the linear system of equations for
the mesh deformation is symmetric and fixed, a parallel LU decomposition is performed

27
 T
Kuu ψM = −YV Reverse mode
T T
Bézier
XS = XS A + R YS YS = Kus ψM interpolation
XS YS YV XSW XV

Reverse mode transfinite interpolation, XS A

Figure 15: Adjoint procedure for solid mesh warping

once using the distributed version of SuperLU. Subsequent mesh perturbations and adjoint
solutions can be computed quickly with a back-substitution operation. Finally, since XS is
given by XS = XJ + T u, the partial-derivative matrix (∂XS /∂u)T is simply T T , which is the
load transfer operation.
The final aerodynamic residual partial derivative, ∂A/∂x, is computed using another
chain rule:      
T ∂A T ∂A ∂XV ∂XS
ψ =ψ . (23)
∂x ∂XV ∂XS ∂x
The two right-most terms are identical to the terms in the computation of ∂A/∂u (21),
The other off-diagonal block in the coupled adjoint equations (19), ∂S/∂w, represents
the derivatives of the structural residuals with respect to the aerodynamic states. The only
contribution to this matrix is due to the aerodynamic forces, FA , and is given by,
 T  T
∂S ∂F
φ=− φ (24)
∂w ∂w
 T  T
∂FA ∂F
=− φ
∂w ∂FA
 T
∂FA
=− T φ.
∂w

We compute and store ∂FA /∂w and ∂FA /∂XS , which we refer to as the coupling matrices.
These terms are computed using reverse-mode AD applied to the CFD force evaluation rou-
tine. Unlike ∂A/∂w or ∂A/∂XV , only cells on the wetted surface have nonzero components.
Thus, the cost of storing these terms is very low.
The remaining diagonal block, ∂S/∂u, represents the derivative of the structural residuals
with respect to the structural states. For linear aerostructural analysis, this term is not
simply the linear stiffness matrix, K. Given the structural residuals (10), we find that the
applied forces, F , are actually explicit functions of u. As the flexible structure deforms, the
surface normals on the CFD mesh change orientation, giving rise to a nonlinear following
force. To obtain derivatives, this effect must be included. To obtain the matrix-vector
products required by the Krylov-type solution strategy, we compute this term in the following

28
manner:
 T  T !
∂S ∂F
φ= K− φ (25)
∂u ∂u
  T  !
∂F ∂FA ∂XS
= K− φ
∂FA ∂XS ∂u
 T !
T ∂FA
= K −T T φ.
∂XS

This product is formed by transferring the φ vector to the aerodynamics using a displace-
ment transfer, finding the ∂FA /∂XS transpose matrix-vector product, and then transferring
the result back to the structural solver using a load transfer. Since this operation requires
two synchronous data transfers, these products are evaluated only as required, and the exact
∂S/∂u is never formed.
The partial derivatives of the structural residuals with respect to the design variables,
∂S/∂x, required in total-derivative equation (18), are evaluated analytically.
The product of the partial derivatives and the adjoint vector ψ T ∂S/∂x is evaluated
analytically on an element-by-element basis, where geometric and material variables are
handled using different approaches. For the geometric variables, the computation proceeds in
two stages. First, the product ψ T ∂S/∂XJ is computed, where XJ are all the node locations.
Next, the product of ψ T ∂S/∂XJ and the derivative of the node locations with respect to the
design variables, ∂XJ /∂x, is computed. This computation is arranged so that it is efficient
when the number of geometric design variables is large. For the derivative of the material
design variables, the sparse dependency of the design variables is exploited, and only elements
with a nonzero contribution to ψ T ∂S/∂x are computed.
The right-hand side of the coupled adjoint system (19) is the derivative of the function
of interest with respect to the system states. We consider only aerodynamic or structural
functions, that is, functions that could be used in a single-discipline analysis. The derivatives
of typical aerodynamic functions such as CL , CD , and CM are computed with reverse-mode
AD, in a similar fashion to the coupling matrices. We analytically evaluate the derivatives
of typical structural functions, such as individual element stresses, σi , or the Kreisselmeier–
Steinhauser (KS) aggregation functions (Poon and Martins, 2007; Wrenn, 1989). For the
structural functions, ∂I/∂w is zero. However, for aerodynamic functions involving surface
pressure or traction integration, ∂I/∂u is nonzero and is given by
  
∂f ∂f ∂XS
= (26)
∂u ∂XS ∂u
 
∂f
= T.
∂XS

The resulting procedure is similar to that used in the computation of ∂S/∂u, expressed in
Eq. (25).
The evaluation of the partial-derivative terms in the total-derivative equation (20) is
challenging because the load and displacement transfer operations have a dependency on the

29
geometric design variables. Figure 13 shows how the displaced surface coordinates, XS , are
extrapolated from the structural domain. The complication arises from the fact that the
length of the extrapolation vector, r, changes with the design variables.
For aerodynamic functions, we compute ∂I/∂XS using reverse-mode AD, similarly to
what was done for the coupling matrices in Eq. (24). Then, for each design variable, we
evaluate the derivative of the perturbed shape, XS , with respect to the design variables, x,
by computing
 
∂f ∂f ∂XS
= (27)
∂x ∂XS ∂x
∂ (ur × r)
 
∂f ∂XJ
= + .
∂XS ∂x ∂x
Since the transfer-operation matrix T is never explicitly formed, we evaluate a single entry of
the resulting vector sequentially in a matrix-free fashion. A similar correction must be made
for the derivative of the structural states with respect to the geometric design variables, since
the load transfer depends on the design variables.
Finally, the evaluation of the ∂XS /∂x term is used to complete the computation of
∂A/∂x. These operations are relatively costly, since in our implementation they require a
synchronous transfer of data from both disciplines for each design variable.

4.5 Coupled adjoint solver

As with the aerostructural system of equations, we consider two approaches to solve the cou-
pled adjoint system (19): a segregated approach and a monolithic approach. The segregated
approach is the lagged coupled adjoint approach (Martins et al., 2004), which corresponds
to a linear block Gauss–Seidel algorithm. We also investigate a monolithic approach, which
uses a Krylov method applied directly to the coupled linear system.

4.5.1 Linear block Gauss–Seidel method (LBGS)

The analog of the NLBGS method for coupled adjoint equations is the linear block Gauss–
Seidel (LGBS) method. This method expresses the interdisciplinary coupling as additional
forcing terms to the right-hand side of each set of disciplinary adjoint equations. If we
separate the single matrix equation into the two blocks we obtain:
 T  T  T
∂A (k) ∂f ∂S
ψ = − φ(k−1) (28)
∂w ∂w ∂w
   T  T  T
∂S (k) ∂f ∂A (k) ∂S
φ = − ψ − φ(k−1) .
∂u K ∂u ∂u ∂u F

The subscripts K and F in the term ∂S/∂u represent the contributions from the stiffness
matrix and external forces, respectively. Since the external-force component is costly to
compute and involves synchronous communication between disciplines, it is lagged. The same
applies to the contribution from the off-diagonal term. We can iterate between these two
equations until we reach the desired convergence level. The main advantage of this approach
is that the discipline adjoint solvers can be re-utilized by simply adding the appropriate

30
right-hand side forcing terms. In practice, we only partially converge each discipline before
a data exchange is made. Additionally, depending on the problem, it may be necessary
to apply an under-relaxation factor to the structural update for enhanced stability. The
pseudocode for this approach is listed in Algorithm 3.

Algorithm 3 Block Gauss–Seidel coupled adjoint solution method

1: Given: ψ (0) , φ(0) , kmax , θ ∈ (0, 1] . ψ (0) , φ(0) can be from a previous solution
2: if Aerodynamic  Function
 then
∂f T ∂f
3: ←T . Cross-discipline function partial derivative
∂u ∂XS
4: end if
5: for k ← 1, kmax do
6: φA ← T φ(k−1) . Transfer structural adjoint
 T  T  T
(k) ∂f ∂A ∂FA
7: RA ← − ψ (k−1) − φA . Aerodynamic adjoint RHS
∂w ∂w ∂w
∂A T
 
(k)
8: ∆ψ = RA . Partially solve CFD adjoint update
∂w
9: ψ (k) ← ψ (k−1) + ∆ψ . Update aerodynamic adjoint
 T  T
∂XV ∂A
10: P ← TT ψ (k) . Transfer aerodynamic contribution to structural RHS
∂XS ∂XV
∂FA T
 
T ∂S
11: Q←T T φ(k−1) . Force contribution
∂X ∂u
 ST
(k) ∂f
12: RS ← − Kφ(k−1) − P − Q . Compute structural RHS
∂u
(k)
13: K∆φ = RS . Partially solve CSM adjoint update
(k) (k−1)
14: φ ←φ + θ∆φ . Under-relaxed structural adjoint update
(k) (1) (k) (1)
15: if RA < SA RA and kRS k < SA kRS k then . Coupled adjoint convergence
check
16: break
17: end if
18: end for

4.5.2 Coupled Krylov method (CK)

The second approach that we use to solve the coupled adjoint equation (19) is a fully coupled
monolithic method. This is the first use of a monolithic solution method for the coupled
adjoint equations on a large-scale aerostructural problem.
Krylov subspace methods are particularly attractive for this type of problem, since they
require only matrix-vector products, allowing the use of our pre-existing matrix-free adjoint
operators. In our case, the diagonal blocks of the Jacobian in Eq. (19) are stored, but
the off-diagonal terms are not explicitly stored. Effective preconditioning is critical to the
performance of Krylov methods, especially on large systems of equations such as the ones
we are solving.

31
 We choose to use a block-Jacobi preconditioner for the coupled system, since this ap-
proach ignores the off-diagonal terms and allows the aerodynamic and structural precondi-
tioning to be carried out in parallel. For the aerodynamic block preconditioner we reuse
the preconditioned Krylov subspace method used for the aerodynamic adjoint. In this case
however, we run a fixed number of GMRES iterations, typically between 10 and 20.
For the structural block of the preconditioner, we use the matrix factorization of K. The
pseudocode for the linear adjoint operator for the coupled system is listed in Algorithm 4.
The matrix-vector products are computed in a matrix-free fashion.

Algorithm 4 Coupled Krylov method linear operator

1: function Mult(X) . Compute Jacobian-vector product with X
2: (XA , XS ) ← X . Extract aerodynamic and structural components
 T
∂A
YA ← XA
3: ∂w . Evaluate diagonal contributions in parallel
YS ← KXS
∂FA T
 
4: YA ← YA + T XS . Add aerodynamic off-diagonal term
∂w
 T  T
∂XV ∂A
5: YS ← YS + T T XA . Add structural off-diagonal term
∂XS ∂XV
∂FA T
 
T ∂S
6: YS ← YS + T T XS . Add force contribution
∂XS ∂u
7: Y ← (YA , YS ) . Combine aerodynamic and structural components
8: return Y
9: end function

5 Aerostructural optimization applications

5.1 Multipoint wing design
Similarly to the cases presented in the previous section, we use the CRM as our baseline
aircraft configuration. Since CRM geometry only provides the 1 g wing shape, we had to
develop a structural wing box model and a jig wing shape, which we call the undeformed
CRM (uCRM) (Kenway et al., 2014a), shown in Figure 16.
The objective for each optimization is to minimize the following function, which involves
both fuel burn and takeoff gross weight (TOGW):
N
X
Obj = Wi (β × Fuel Burn + (1.0 − β) TOGW) . (29)
i

Extreme β values of 0 and 1 yield the TOGW and fuel burn objectives, respectively. The
real objective function for a transport aircraft lies somewhere between these two extremes
and depends on how the airline uses the aircraft, fuel price, and other factors.
For each flight condition considered in our multipoint optimization, we compute the fuel
burn and TOGW as if the entire mission was completed in a cruise-climb fashion at the L/D

32
Figure 16: uCRM aerostructural model. The wingbox structural model is shown on the left
and the aerodynamic model on the right. The black outline denotes the jig shape. The CFD
surface mesh resolution is also visible.

ratio evaluated at that analysis point. For simplicity, we ignore the fuel burn associated with
the taxi, take-off climb, and descent phases. Using the Breguet range equation applied to
the full design range we compute the TOGW and fuel burn.
The flight conditions considered in our optimization are shown in Fig. 17. The basic
operating condition is M = 0.85, CL = 0.5, at 34 000 ft. The lift at this operating condition
corresponds to a point in the flight where 1/2 of the total mission fuel has been burned.
The next four operating conditions, Points 2 through 5, are evaluated at the same physical
lift (fuel fraction of 0.5). We vary the Mach number by ±0.01, and the CL is modified
accordingly. Points 4 and 5 vary the altitude by ±2000 ft, which yields larger variations in
the lift coefficient. The final two normal operating conditions—Points 6 and 7—vary the fuel
fraction from near empty (10%) to nearly full (90%). The main reason for including these
points, is to capture the effects due to higher or lower wing bending and the accompanying
twist variations. For these cases the altitudes have been adjusted so that the CL is close to
the nominal value.
The final two operating conditions are for enforcing the buffet onset constraints, whose
formulation we introduced in Sec. 3.2. For this optimization study, we have chosen to include
these two conditions directly in the objective function. The first buffet condition is chosen
to be 1.3 g higher than the nominal design point (Point 1). The second buffet point, is at
the maximum operating Mach (MMO ) condition, which we have assumed to be M = 0.89 for
the CRM configuration. The lift at this condition is the same as the nominal flight condition
and thus the increased Mach number results in a lower coefficient of lift.
In addition to the cruise and buffet operating conditions, we must also define several
maneuver loading conditions that produce representative loads to size the wing structure.
The three maneuver load conditions are as follows:

33
 340

1.3g
290
+40 t, -3000 ft

Weight (t)
0.04M

+/- 2000 ft
+/- 0.01M
240

-40 t, +4000 ft

30000
32000
34000 190
Alti 0.9
tude 36000
(ft) 0.88
38000 0.86 h
40000
0.84 Ma c
0.82

Figure 17: Flight operating conditions considered in the optimizations.

2.5 g symmetric pull up : For transport aircraft this is one of the most critical sizing
conditions. This analysis is performed at TOGW, M = 0.6415, sea-level altitude with
full fuel. This Mach number is chosen to correspond to an assumed VD speed of 425
knots.

−1.0 g symmetric push over : The primary effect of this maneuver condition is the siz-
ing of the lower wing surface dictated by local buckling constraints. The analysis is
performed at the same conditions as the 2.5 g condition.

Quasi-steady gust load : The purpose of this condition is to ensure the operating stresses
at normal operating conditions are not too high. It can also be considered as a surrogate
for a gust load that may be expected to load the wing sufficiently quickly that the
passive aeroelastic tailoring cannot redistribute the load inboard fast enough. This
maneuver condition is analyzed near the Mach cross-over altitude of 28 000 ft, M =
0.85, TOGW and full fuel load.

The wingbox structure is assumed to be constructed of a 7000 series aluminum alloy with
a limiting stress of 420 Mpa. The 2.5 g and −1.0 g conditions incorporate an additional 1.5
safety factor as required by regulations. The third flight condition utilizes a 2.67 safety factor
to ensure a sufficient structural margin for a gust encountered near the cruising altitude.
The full optimization problem (objective, design variables, and constraints) is summa-
rized in Table 5.1. When it comes to design variables, there are several groups: geometric
design variables, aerodynamic variables, engine model variables, and structural variables.
The geometric and structural design variables are shown in Fig. 18.
We have selected a compact set of variables that are able to efficiently manipulate the
wing planform: span, sweep, and chord. When combined, the span and chord variables

34
 Figure 18: Geometric and structural optimization variables.

control the wing area and aspect ratio. Since the chord variable is applied evenly to the
entire span, the taper ratio of the initial configuration remains constant. The planform
variables yield tremendous design freedom typically not found in high-fidelity aerodynamic
only optimization. In particular, the optimization process is free to explore the trade-off
between increasing the wingspan to reduce the span-loading and thus induced drag at the
expense of a potentially heavier structure. Likewise, the sweep angle may be modified in
response to changes in the airfoil thickness distributions as well as accounting for the change
in structural mass. The remaining geometric variables are 7 twist variables, a tail rotation
angle for each flight condition, and 192 shape variables controlling the detailed cross sections.
The aerodynamic variables include the angle of attack at each of the cruise and maneuver
conditions and the nominal cruise altitude. This single altitude variable is used to drive the
operating altitude for all flight conditions. Each of the normal operating conditions have a
throttle variable. This variable determines the thrust requested from the engine, which is
necessary to meet the D = T constraint.
The remainder of the design variables are used for the parametrization of the structural
wingbox. The first four variables are the stiffener pitch of the upper skin, lower skin, leading
edge spar and trailing edge spar. We assume that the stiffener pitch is thus constant across
each component. The panel-based smeared stiffness approach results in three design variables
for the skin or spar in each rib bay: panel thickness, stiffener thickness, and stiffener height.
An additional variable, the panel length, is used to simplify panel buckling computations
and is constrained to match the physical panel length (which changes as a function of the
geometric design variables) through an equal number of nonlinear constraints. Altogether,
there are 972 design optimization variables.
Even with high fidelity analyses, the optimization problem requires many constraints.
Each of the cruise and maneuver operating conditions require that lift equals weight and

35
 Function/variable Description Quantity
XN
minimize Wi (β × Fuel Burn + (1.0 − β) TOGW)
i
with respect to xspan Wing span 1
xsweep Wing sweep 1
xchord Wing chord 1
xtwist Wing twist 7
xairfoil FFD control points 192
xalphai Angle of attack at each flight condition 12
xηi Tail rotation angle at each flight condition 12
xthrottlei Throttle setting for each cruise flight con- 7
dition
xaltitude Cruise altitude 1
XCG CG position 1
xskin pitch Upper/lower stiffener pitch 2
xspar pitch LE/TE Spar stiffener pitch 2
xribs Rib thickness 45
xpanel thick Panel thickness skins/spars 172
xstiff thick Panel stiffener thickness skins/spars 172
xstiff height Panel stiffener height skins/spars 172
xpanel length Panel length skin/spars 172
Total design variables 972

subject to L = ni W Lift constraint 12

CMyi = 0.0 Trim constraint 12
T =D Thrust constraint 7
1.08D − Tmax < 0 Excess thrust constraint 7
tLE /tLEInit ≥ 1.0 Leading edge radius 20
tTE /tTEInit ≥ 1.0 Trailing edge thickness 20
Vwing > Vfuel Minimum fuel volume 1
xCG − 1/4M AC = 0 CG location at 1/4 chord MAC 1
Lpanel − xpanel length = 0 Target panel length 172
KSstress ≤ 1 2.5 g Yield stress 4
KSbuckling ≤ 1 2.5 g Buckling 3
KSbuckling ≤ 1 -1.0 g Buckling 3
KSbuckling ≤ 1 1.78 g Yield stress 3
KSbuckling ≤ 1 1.78 g Buckling 4
xpanel thicki − xpanel thicki+1 ≤ 0.005 Skin thickness adjacency 168
xstiff thicki − xstiff thicki+1 ≤ 0.005 Stiffener thickness adjacency 168
xstiff heighti − xstiff heighti+1 ≤ 0.005 Stiffener height adjacency 168
xstiff thick − xpanel thick < 0.005 Maximum stiffener-skin difference 172
∆zTE,upper = −∆zTE,lower Fixed trailing edge 8
∆zLE,upper = −∆zLE,lower Fixed leading edge 8
Total constraints 961

Table 3: Summary of aerostructural design optimization problem.

that the configuration is trimmed (CMy = 0.0). Each of the normal operating conditions
have two additional constraints related to the engine model. The first is that the thrust must
equal the drag, which is met by the engine throttle setting variable. This is important, as
the TSFC is a function of throttle setting. The second is an excess thrust constraint, which
ensures sufficient climb margin (or equivalently thrust to drag ratio) at the normal operating
conditions.
The 2.5 g maneuver condition uses four KS stress constraint aggregation functions: one
for the upper wing surface, a second for the lower wing surface, one for the spars, and a

36
final one for the ribs. Three buckling constraints are used for the upper skin, spars, and
ribs respectively. For the −1.0 g condition, three buckling constraints are used: one for the
lower skin, one for the spars and a final for the ribs. Finally, for the quasi-steady gust
load condition, there are three yield stress constraints and four buckling constraints. We
use several hundred adjacency constraints that ensure there are no large changes in the
properties of the skin or stiffener between adjacent panels. Finally, there are 16 constraints
for the FFD airfoil shape variables that keep the leading and trailing edges from moving in
the z direction. This is required to eliminate a “shearing twist” design mode, and allows
the pure rotation (nonlinear) twist to modify the overall wing twist. In total, there are 961
optimization constraints, for the case 4 optimization.
We present two optimizations: one that minimizes the fuel burn (β = 1), and another
that minimizes TOGW (β = 0). This enables us to see the effect of the objective choice on
the optimized designs. The resulting optimizations are shown in Figs. 19 and 20.
The fuel burn minimization shown in Fig. 19 dramatically increased the wing span, at
the cost of a heavier wing structure. The wingbox thickness distribution in the upper right
of Fig 19 confirms that the wing skins are much thicker. There is also an increase in area,
and lower thickness to chord ratios.
The TOGW minimization (Fig. 20) shows very different trends. The wingspan is short-
ened slightly compared to the original configuration and the wingbox is thickened over the
mid-span section. Outboard however, the wing is thinner than the baseline.

5.2 Tow-steered composite optimization

In the last 30 years since their introduction into aerospace applications, composites have
become increasingly used, making up as much as 50% of modern aircraft by weight. Consid-
ering this fact, it is surprising that most aircraft today are only scratching the surface of the
true potential of composite technology with traditional uniaxial fibers. With the introduc-
tion of automated fiber placement machines, the tow direction in laminae is now allowed to
be steered spatially throughout each layer. This process is known as composite tow steering
and has been shown to have improved performance over its uniaxial fiber counterpart with
no additional weight penalty. With modern aircraft wings moving toward higher aspect ra-
tios, which inevitably leads to larger deflections, it is reasonable to assume that a tow-steered
composite structure can be tailored to outperform its unsteered counterpart. However, given
the highly coupled nature of the aerodynamics and structural response of the problem it is
not obvious nor intuitive to find the composite fiber pattern that would yield an optimum
result. To address this issue, we perform the simultaneous design of aerodynamic shape,
structural sizing, and tow steering angles, while considering the wing flexibility.
Two aerostructural optimizations were performed: one with the full tow steering parametriza-
tion , which we will call the steered case, and one where the tow orientation was held fixed,
the unsteered case, but the laminate thickness and other variables were still allowed to vary.
The unsteered case is done in order to isolate the benefits due to the tow steering process.
Each optimization requires four aerostructural analysis: a cruise condition for evaluating
the performance, and three conditions for which the structural constraints are enforced: a
−1.0 g dive, 2.5 g pull-up, and 1.0 g gust maneuver condition. The parameters for these four
conditions are listed in Table 4, where TOGW is the Takeoff Gross Weight of the aircraft.

37
 Figure 19: Fuel burn minimization result.

A summary of the tow-steered optimization problem is shown in Table 5.

Table 4: Optimization flight condition parameters

Parameter Cruise 2.5g -1.0g 1.0g gust
Mach number 0.85 0.64 0.64 0.86
Altitude (ft.) 37 000 0 0 27 300
Weight LGW + 0.2 × FB TOGW TOGW TOGW

As a baseline for our optimizations, we use the undeflected Common Research Model
13.5 (uCRM 13.5) is a 13.5 aspect ratio variant of the CRM. The purpose of the uCRM-
13.5, similarly to that of the uCRM-9 developed by Kenway et al. (2014a), is to define a jig
wing geometry and its corresponding structural layout to be used as a baseline for future
aerostructural studies. The high aspect ratio wing planform was defined by taking the
planform of the uCRM 9 and increasing the span while keeping the reference area constant

38
 Figure 20: TOGW minimization result.

until an aspect ratio of 13.5 was achieved. The outer mold line (OML) was then defined
through a single point aerostructural shape optimization of a conventional aluminum wing
with the objective of minimizing fuel burn. A three-view of the uCRM-13.5 is shown in
Fig. 21. The wing planform and wing box structure used for the optimization are shown in
Fig. 22. In addition to aerodynamic loads the structural model also considers the inertial
effects of the non-structural masses. This includes a 7,500 kg engine along with a number of
leading edge and trailing edge masses concentrated at discrete locations along the span of
the wing, as shown in Fig. 22. Finally, the analysis also includes the distributed weight of
the fuel in the wing during flight. This inertial load is applied as a distributed traction over
the lower surface of the wing skin.
The objective of the optimization problem is to minimize the fuel burn of the aircraft,
calculated using the Breguet range equation.
The design variables used in the optimization can be broken down into three categories:
aerodynamic, geometric, and structural. The aerodynamic design variables consists of 4
angles of attack, one for each flight condition. The geometric design variables consist of
216 wing FFD variables which control the wing cross-sectional shape, 8 twist variables,

39
 Variable/function Description Quantity
Minimize FB Fuel burn

with respect to αi Angle of attack for each flight case 4

γ Wing twist 10
xshape Wing FFD variables 216
φi Tail trim angle for each case 4
cp
θwing,U Wing upper skin tow variables 120
cp
twing,U Wing upper skin thickness variables 120
cp
θwing,L Wing lower skin tow variables 120
cp
twing,L Wing lower skin thickness variables 120
trib Rib thickness variables 58
tcp
LS Leading edge spar thickness variables 20
tcp
TS Trailing edge spar thickness variables 20
Total design variables 812

subject to L = ni W Load factor 4

cimy = 0 Trimmed flight 4
tLE /tLEinit ≥ 1.0 Leading edge radius 20
tT E /tT Einit ≥ 1.0 Trailing edge thickness 20
KSstress < 1.0 Yield stress 9
KSbuckling < 1.0 Buckling 9
t > tgage Gage thickness 258
1 1
− < KSκ < Fiber path turning radius 112
rmin rmin
−ψ0 < KSψ < ψ0 Fiber path divergence 112
KS||∇t|| < ||∇t0 || Thickness variation 221
Total constraints 769

Table 5: Tow-steered wing box minimization problem

and 4 horizontal tail incidence angles used to trim each flight condition. The structural
design variables consist of 58 rib thickness variables (one for each rib) and 40 spar thickness
variables (20 for each spar). The remaining structural variables for the steered case are the
tow-steering variables for the upper and lower skins of the wing. These are the thickness and
main tow-steering orientation values set at the control points on the upper and lower wing
box skins. There are 120 control points for each surface, each of which corresponding to a
thickness and tow orientation variable for a total of 480 design variables. For the unsteered
optimization, the steer orientation variables are fixed while the thickness variables are free to
vary. The total number of design variables for the tow-steered optimization is 812 variables.
To produce meaningful results, a number of constraints had to be applied to the opti-
mization, as summarized in Table 5. The structural sizing of the wing box is dictated by
three maneuver conditions. The first set of constraints ensure that the lift produced by the
aircraft in each condition of flight had to match the weight of the aircraft in that condition
multiplied by the load factor of that condition. In this case, there four lift constraints, one
for each condition. In addition to this constraint, the pitching moment for each flight condi-
tion is constrained to be zero such that each flight conditioned is trimmed, for another four
constraints. The leading edge radius was constrained from decreasing in order to maintain
high lift performance of the wing, and the trailing edge thickness was constrained to ensure
manufacturability, adding another 40 constraints. The failure and buckling constraints were
aggregated into three groups: ribs and spars, upper skin, and lower skin. These constraints

40
 Figure 21: Three-view of the uCRM-13.5 model

Figure 22: uCRM-13.5 planform and wing box with non-structural masses

were enforced for the 3 maneuver flight conditions, resulting in a total of 18 constraints: 9
for failure and 9 for buckling.
The next set of constraints are related to the divergence and curvature of the tow-steered
skins. The fiber curvature was constrained to have a minimum turning radius of 70 in, which
corresponds to a magnitude of curvature of 0.0143 in−1 . Both constraints were aggregated
over each panel of the wing box for a total of 224 constraints (112 on each side). The total

41
number of constraints for the optimization problem adds up to 769.
The results of the tow-steered and unsteered optimization are shown in Fig. 23. We
can see that there is a 1% improvement, in the fuel burn of the tow-steered relative to
the unsteered wing. From the spanwise lift distribution of the two wings for the cruise
and 2.5 g maneuver, we can see that the difference between the cruise and 2.5 g maneuver
conditions, which corresponds to the passive load alleviation that we seek. Although there
is some amount of passive load alleviation, the steered case does not show significantly more
load alleviation when compared to the unsteered case. This may be due to the fact that
without stiffeners defined in the model, the buckling constraint dominates the optimization.
Further evidence for this can be seen in the fact that the thickness distribution changes
most noticeably on the lower skin of the wing, where the buckling would be less critical for
the design. In addition, a greater difference between optimal cruise and maneuver loads is
achieved for other objectives that emphasize structural weight, such as takeoff gross weight
(Kenway et al., 2014a). Despite the small difference in load alleviation the optimizer is able
to decrease the wing weight of the steered wing by 13% relative to the unsteered case. This is
most probably because the optimizer has the freedom to place the composite tow orientation
closer to the orientation of the principal stress in high stressed regions.
Figure 24 shows the a comparison between the tow-steered and unsteered patterns for
each ply. Due to the buckling constraints and the absence of stringers the wing box panels
are much thicker than in a realistic wing box. For both wings, the optimizer adds additional
laminate thickness at the engine mounting location to deal with the local stress concentration.
The main tow pattern for the upper and lower skin of the wing box both steer back towards
the trailing edge at the root and the gradually sweep forward towards the tip.

6 Summary
In Part 2 of this lecture, we presented a framework for aerodynamic and aerostructural
design optimization of aircraft wings. The methods used in this framework successfully
tackle the compounding challenges of modeling the wing with high fidelity, while optimizing
it with respect to hundreds of design variables. The effectiveness of this framework hinges
the use of high-performance parallel computing, fast solvers, state-of-the-art gradient-based
optimization, and an efficient and accurate approach for computing the derivatives for the
aerostructural solver via the coupled adjoint method. We demonstrated the effectiveness of
the methods by presenting the results of four design optimization cases. Other cases have
been solved, and the interested reader can refer to other published work (Chen et al., 2016;
Garg et al., 2015; James et al., 2014; Kenway and Martins, 2016b; Liem et al., 2015; Lyu
and Martins, 2014, 2015; Mader and Martins, 2013).

References
Albring, T., Sagebaum, M., and Gauger, N. R. (2016). New Results in Numerical and
Experimental Fluid Mechanics X: Contributions to the 19th STAB/DGLR Symposium
Munich, Germany, 2014, chapter A Consistent and Robust Discrete Adjoint Solver for
the SU2 Framework—Validation and Application, pages 77–86. Springer International
Publishing, Cham.

42
Figure 23: Comparison tow-steered (right) vs unsteered (left) aerostructural optimization.

Balay, S., Gropp, W. D., McInnes, L. C., and Smith, B. F. (1997). Efficient Management of
Parallelism in Object Oriented Numerical Software Libraries, pages 163–202. Birkhäuser
Press.

Brown, S. A. (1997). Displacement extrapolation for CFD+CSM aeroelastic analysis. In

Proceedings of the 35th AIAA Aerospace Sciences Meeting, Reno, NV. AIAA 1997-1090.

Chen, S., Lyu, Z., Kenway, G. K. W., and Martins, J. R. R. A. (2016). Aerodynamic shape
optimization of the Common Research Model wing-body-tail configuration. Journal of
Aircraft, 53(1):276–293. doi:10.2514/1.C033328.

Garg, N., Kenway, G. K. W., Lyu, Z., Martins, J. R. R. A., and Young, Y. L. (2015). High-
fidelity hydrodynamic shape optimization of a 3-D hydrofoil. Journal of Ship Research,
59(4):209–226. doi:10.5957/JOSR.59.4.150046.

Gill, P. E., Murray, W., and Saunders, M. A. (2002). SNOPT: An SQP algorithm for
large-scale constrained optimization. SIAM Journal of Optimization, 12(4):979–1006.
doi:10.1137/S1052623499350013.

43
 Figure 24: Comparison of optimal tow-steered vs unsteered aerostructural designs.

Haftka, R. T. (1977). Optimization of flexible wing structures subject to strength and

induced drag constraints. AIAA Journal, 14(8):1106–1977. doi:10.2514/3.7400.

Irons, B. M. and Tuck, R. C. (1969). A version of the aitken accelerator for computer
iteration. International Journal for Numerical Methods in Engineering, 1(3):275–277.

Irving, C. (1993). Wide-Body—The Triumph of the 747. William Morrow and Company,
New York.

James, K. A., Kennedy, G. J., and Martins, J. R. R. A. (2014). Concurrent aerostruc-

tural topology optimization of a wing box. Computers and Structures, 134:1–17.
doi:10.1016/j.compstruc.2013.12.007.

Jameson, A., Schmidt, W., and Turkel, E. (1981). Numerical solution of the Euler equations
by finite volume methods using Runge-Kutta time stepping schemes. AIAA Paper 81-1259.

Kennedy, G. J. and Martins, J. R. R. A. (2014). A parallel finite-element framework for large-

scale gradient-based design optimization of high-performance structures. Finite Elements
in Analysis and Design, 87:56–73. doi:10.1016/j.finel.2014.04.011.

44
Kenway, G. K., Kennedy, G. J., and Martins, J. R. R. A. (2010). A CAD-free approach
to high-fidelity aerostructural optimization. In Proceedings of the 13th AIAA/ISSMO
Multidisciplinary Analysis Optimization Conference, Fort Worth, TX. AIAA 2010-9231.

Kenway, G. K. and Martins, J. R. R. A. (2016a). Aerodynamic shape optimization of the

CRM configuration including buffet-onset conditions. In 54th AIAA Aerospace Sciences
Meeting, San Diego, CA. American Institute of Aeronautics and Astronautics.

Kenway, G. K. W., Kennedy, G. J., and Martins, J. R. R. A. (2014a). Aerostructural

optimization of the Common Research Model configuration. In 15th AIAA/ISSMO Mul-
tidisciplinary Analysis and Optimization Conference, Atlanta, GA. AIAA 2014-3274.

Kenway, G. K. W., Kennedy, G. J., and Martins, J. R. R. A. (2014b). Scalable parallel

approach for high-fidelity steady-state aeroelastic analysis and derivative computations.
AIAA Journal, 52(5):935–951. doi:10.2514/1.J052255.

Kenway, G. K. W. and Martins, J. R. R. A. (2014). Multipoint high-fidelity aerostructural

optimization of a transport aircraft configuration. Journal of Aircraft, 51(1):144–160.
doi:10.2514/1.C032150.

Kenway, G. K. W. and Martins, J. R. R. A. (2015a). AIAA ADODG Case 5: CRM wing-

body-tail optimization at flight reynolds number. Technical report, AIAA.

Kenway, G. K. W. and Martins, J. R. R. A. (2015b). Multipoint aerodynamic shape opti-

mization investigations of the Common Research Model wing. In Proceedings of the AIAA
Science and Technology Forum and Exposition (SciTech), Kissimmee, FL.

Kenway, G. K. W. and Martins, J. R. R. A. (2016b). Multipoint aerodynamic shape optimiza-

tion investigations of the Common Research Model wing. AIAA Journal, 54(1):113–128.
doi:10.2514/1.J054154.

Liem, R., Kenway, G. K. W., and Martins, J. R. R. A. (2015). Multimission aircraft fuel burn
minimization via multipoint aerostructural optimization. AIAA Journal, 53(1):104–122.
doi:10.2514/1.J052940.

Lovely, D. and Haimes, R. (1999). Shock detection from computational fluid dynamics
results. In Proceedings of the 14th Computational Fluid Dynamics Conference, Norfolk,
VA.

Luke, E., Collins, E., and Blades, E. (2012). A fast mesh deformation method
using explicit interpolation. Journal of Computational Physics, 231(2):586–601.
doi:10.1016/j.jcp.2011.09.021.

Lyu, Z., Kenway, G. K., and Martins, J. R. R. A. (2015). Aerodynamic shape optimization
investigations of the Common Research Model wing benchmark. AIAA Journal, 53(4):968–
985. doi:10.2514/1.J053318.

45
Lyu, Z., Kenway, G. K., Paige, C., and Martins, J. R. R. A. (2013). Automatic differentiation
adjoint of the Reynolds-averaged Navier–Stokes equations with a turbulence model. In
21st AIAA Computational Fluid Dynamics Conference, San Diego, CA.

Lyu, Z. and Martins, J. R. R. A. (2014). Aerodynamic design optimization studies of a

blended-wing-body aircraft. Journal of Aircraft, 51(5):1604–1617. doi:10.2514/1.C032491.

Lyu, Z. and Martins, J. R. R. A. (2015). Aerodynamic shape optimization of an adaptive

morphing trailing edge wing. Journal of Aircraft, 52(6):1951–1970. doi:10.2514/1.C033116.

Mader, C. A., Kenway, G., and Martins, J. R. R. A. (2008). Towards high-fidelity

aerostructural optimization using a coupled ADjoint approach. In Proceedings of the
12th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, Victoria,
BC. AIAA 2008-5968.

Mader, C. A. and Martins, J. R. R. A. (2013). Stability-constrained aerodynamic shape

optimization of flying wings. Journal of Aircraft, 50(5):1431–1449. doi:10.2514/1.C031956.

Mader, C. A. and Martins, J. R. R. A. (2014). Computing stability derivatives and

their gradients for aerodynamic shape optimization. AIAA Journal, 52(11):2533–2546.
doi:10.2514/1.J052922.

Martins, J. R. R. A., Alonso, J. J., and Reuther, J. J. (2004). High-fidelity aerostruc-

tural design optimization of a supersonic business jet. Journal of Aircraft, 41(3):523–530.
doi:10.2514/1.11478.

Martins, J. R. R. A., Alonso, J. J., and Reuther, J. J. (2005). A coupled-adjoint sensitivity

analysis method for high-fidelity aero-structural design. Optimization and Engineering,
6(1):33–62. doi:10.1023/B:OPTE.0000048536.47956.62.

Martins, J. R. R. A. and Lambe, A. B. (2013). Multidisciplinary design optimization: A

survey of architectures. AIAA Journal, 51(9):2049–2075. doi:10.2514/1.J051895.

Martins, J. R. R. A., Mader, C. A., and Alonso, J. J. (2006). ADjoint: An approach for
rapid development of discrete adjoint solvers. In Proceedings of the 11th AIAA/ISSMO
Multidisciplinary Analysis and Optimization Conference, Portsmouth, VA. AIAA 2006-
7121.

Maute, K., Nikbay, M., and Farhat, C. (2001). Coupled analytical sensitivity analysis and op-
timization of three-dimensional nonlinear aeroelastic systems. AIAA Journal, 39(11):2051–
2061.

Perez, R. E., Jansen, P. W., and Martins, J. R. R. A. (2012). pyOpt: A Python-based

object-oriented framework for nonlinear constrained optimization. Structural and Multi-
disciplinary Optimization, 45(1):101–118. doi:10.1007/s00158-011-0666-3.

Peter, J. E. V. and Dwight, R. P. (2010). Numerical sensitivity analysis for aerody-

namic optimization: A survey of approaches. Computers and Fluids, 39(3):373–391.
doi:10.1016/j.compfluid.2009.09.013.

46
Poon, N. M. K. and Martins, J. R. R. A. (2007). An adaptive approach to constraint aggre-
gation using adjoint sensitivity analysis. Structural and Multidisciplinary Optimization,
34(1):61–73. doi:10.1007/s00158-006-0061-7.

Saad, Y. and Schultz, M. H. (1986). GMRES: A generalized minimal residual algorithm for
solving nonsymmetric linear systems. SIAM Journal on Scientific and Statistical Comput-
ing, 7(3):856–869. doi:10.1137/0907058.

Sederberg, T. W. and Parry, S. R. (1986). Free-form deformation of solid geometric models.

SIGGRAPH Comput. Graph., 20(4):151–160. doi:10.1145/15886.15903.

Spalart, P. and Allmaras, S. (1992). A one-equation turbulence model for aerodynamic flows.
In 30th Aerospace Sciences Meeting and Exhibit.

van der Weide, E., Kalitzin, G., Schluter, J., and Alonso, J. J. (2006). Unsteady turboma-
chinery computations using massively parallel platforms. In Proceedings of the 44th AIAA
Aerospace Sciences Meeting and Exhibit, Reno, NV. AIAA 2006-0421.

Vassberg, J. (2011). A unified baseline grid about the common research model wing/body for
the fifth aiaa cfd drag prediction workshop (invited). In 29th AIAA Applied Aerodynamics
Conference.

Vassberg, J. C., DeHaan, M. A., Rivers, S. M., and Wahls, R. A. (2008). Development of a
common research model for applied CFD validation studies. AIAA 2008-6919.

Vassberg, J. C., Tinoco, E. N., Mani, M., Rider, B., Zickuhr, T., Levy, D. W., Broder-
sen, O. P., Eisfeld, B., Crippa, S., Wahls, R. A., Morrison, J. H., Mavriplis, D. J., and
Murayama, M. (2014). Summary of the fourth AIAA computational fluid dynamics drag
prediction workshop. Journal of Aircraft, 51(4):1070–1089. doi:10.2514/1.c032418.

Wrenn, G. A. (1989). An Indirect Method for Numerical Optimization Using the

Kreisselmeier-Steinhauser Function. Technical report, NASA Langley Research Center,
Hampton, VA.

47