This is a preprint of the following article, which is available at: http://mdolab.engin.umich.

edu
A. C. Gray and J. R. R. A. Martins. Geometrically nonlinear high-fidelity aerostructural optimization
for highly flexible wings. AIAA SciTech, 2021.
The original article may differ from this preprint and is available at:
https://doi.org/10.2514/6.2021-0283.

Geometrically Nonlinear High-fidelity Aerostructural

Optimization for Highly Flexible Wings
Alasdair C. Gray and Joaquim R. R. A. Martins
Department of Aerospace Engineering, University of Michigan, Ann Arbor, MI, 48109

Abstract
Over the past decade, advances in multidisciplinary design optimization (MDO) have
enabled the optimization of aircraft wings using high-fidelity simulations of their coupled
aerodynamic and structural behavior. However, as their aspect-ratios increase, these wings
increasingly exhibit geometrically nonlinear behavior that cannot be correctly modeled by
typical linear structural analysis methods. Although many low-fidelity aeroelastic analysis
and optimization tools feature geometrically nonlinear finite element models, all examples to
date of high-fidelity aerostructural optimization have been restricted to linear finite element
models due to concerns over computational cost and complexity. In this work, we develop a
tool for high-fidelity, geometrically nonlinear structural and aerostructural optimization and
demonstrate it by performing detailed structural sizing optimization of a high aspect-ratio
wingbox using high-fidelity RANS CFD and a geometrically nonlinear shell finite element
formulation. By comparing the results of aerostructural analysis and optimization using
both geometrically linear and nonlinear structural finite element formulations, we find that
geometric nonlinearity results in a 5-10% increase in the bending stress and optimized struc-
tural mass of the wingbox. Geometrically nonlinear analysis not only results in deflected
wing shapes and lift distributions which are more physically accurate, but also captures new
loading phenomena within the wingbox, both of which lead to more realistic structural sizing.
We also show that the increase in computational cost when optimizing with geometrically
nonlinear analysis can be as low as 15%, meaning these formulations should not be viewed
as prohibitively expensive.

Nomenclature

R Residual θ Under-relaxation Factor

∆X, ∆Y, ∆Z Displacement Components E Strain Energy
λ Load Factor Fex External Force
λstiff Stiffener Pitch Fin Internal Force
 
KT Tangent Stiffness Matrix Fx,cr Critical Compressive Load

1
Fxy,cr Critical Shear Load TF Force Transfer Process
hstiff Stiffener Height TX Displacement Transfer Process
tskin Skin Thickness η Normalized Semispan
tstiff Stiffener Thickness θtail Tail Rotation Angle

1 Introduction
Whether for environmental or economic reasons, aircraft manufacturers are constantly striv-
ing to design and build more efficient aircraft. One of the conceptually simplest ways to
increase the aerodynamic efficiency of aircraft is to increase the aspect-ratio of their wings.
Consequently, the past five decades have seen a steady increase in the aspect-ratio of com-
mercial aircraft wings.

11 B787 B777x
10 A330
B777-300ER A350
AR 9 A310
B767-400ER
B777
8 B767 B747-400
A300
7 B747 L1011

1970 1980 1990 2000 2010 2020

Year
Figure 1: The aspect-ratio of commercial transport aircraft wings has increased 50% in the
past 50 years [1].

These developments have required two distinct areas of innovation; first, the development
of novel airframe technologies such as advanced composite materials, active maneuver load
alleviation through control surface deflections, and passive load alleviation through aeroelas-
tic tailoring of wing structures, and second, the development of new design methodologies
to extract the full potential of these new airframe technologies. The need for new design
methodologies for high-aspect-ratio wings (HARW) stems from two separate problems. The
first is that as the aspect-ratio of wings grow, they naturally become more flexible, leading
to a stronger coupling between their structural and aerodynamic behavior, requiring coupled
analysis techniques. Secondly, many of the novel airframe technologies mentioned above,
that are key enablers of HARW, bring with them an increased number of design parameters,
further complicating the already complex task of wing design.
In response to this need, there has also been increased interest in the application of
multidisciplinary design optimisation (MDO) methods to aircraft design and, in particular,
aerostructural wing design. However, due to the replacement of engineering intuition with
simulation results, the computational methods used in MDO frameworks must be capable
of modeling the physical phenomena that limit the design space in reality. If not, there is
a common tendency for the optimizers to take advantage of gaps in the modeled physics,
producing unrealistic optimal designs. HARW exhibit geometrically nonlinear behavior, as
a result of their high flexibility, that cannot be correctly modeled using linear finite element

2
(FE) methods. Using linear models may lead to inaccurate sizing of wing structures, both
due to the direct inaccuracy in the linear stress calculation and inaccuracy in the deflected
wing shape which consequently effects the aerodynamic load distribution over the wing.
Geometrically nonlinear FE methods for structural analysis are a mature technology
available in most commercial structural analysis codes and have been incorporated into a
number of aeroelastic analysis frameworks in the past two decades [2–15]. However, for
aerostructural analysis and design, these methods use low-fidelity beam FE models, and
often low-fidelity panel-based aerodynamic models. Although these simplified models have
been proven capable of accurately predicting the global deflected shape of a wing, they are
not able to predict detailed stress distributions within the wingbox components due the
condensation inherent in their formulation. Conversely, high-fidelity approaches, that can
accurately trade-off transonic cruise drag and structural mass, have so far been limited to
linear FE methods [16–18].
This paper begins with a review of geometrically nonlinear aeroelastic phenomena and the
existing computational tools that have been used to analyze and design HARW considering
geometrically nonlinear behavior. Next, we present our implementations of geometrically
nonlinear solvers for structural and aeroelastic analysis with the MACH (MDO of Aircraft
Configurations at High-fidelity) framework, before using them to demonstrate the effect of
geometric nonlinearity on the analysis of the undeflected common research model (uCRM),
a benchmark high-aspect-ratio commercial aircraft model. Finally, we compare the results
of structural and aerostructural optimization of the uCRM with high-fidelity, linear and
geometrically nonlinear FE models.

1.1 Background
MDO is a powerful tool for aerostructural design of aircraft due to tight couplings between
aerodynamics and structures in aircraft design, this utility will only be increased as future
aircraft wings become more slender and flexible. The MACH framework, represents the
current state of the art in high-fidelity aerostructural optimization, allowing for concurrent
3


optimization of wing shape and structure using O 10 design variables using RANS com-
putational fluid dynamics (CFD) and detailed FE wingbox models. MACH has been used to
investigate the potential benefits of some of the new airframe technologies enabling HARW.
Burdette studied the effect of wing morphing in various works [19, 20] and Brooks extended
the capabilities of TACS in order to study the performance benefits of tow-steered composite
wings [21, 22].
A parallel trend has been in the development of aeroelastic optimization frameworks
which use lower fidelity panel aerodynamic models and often, though not always, beam FE
models. These frameworks generally cannot be used for wing shape or planform optimization
as they cannot accurately predict the transonic and viscous effects that are critical for drag
prediction and thus cannot accurately trade-off drag and structural mass. They are therefore
typically limited to minimizing structural mass for a fixed wing-shape. By reducing model
complexity however, frameworks such as PROTEUS [23] and UM/NAST [5] allow for gains in
analysis and optimization complexity, some examples of which include transient loadcases [23–
25], fatigue life constraints [26], simultaeneous optimization of structural sizing and control
laws [25], and, most relevant to this work, geometrically nonlinear structural models [14, 23,
24, 26–30].
Linear structural analysis methods rely on the assumption that the displacements, strains,
stresses, and forces in a structure are all linearly related to one another. For many engineering

3
applications, where structures should remain in the linear elastic region and undergo small
displacements, this is a good assumption and simplifies any analysis to the solution of a single
system of linear equations that describes the equilibrium of external and internal forces on
the structure in its undeformed configuration. However, under large deformations, and par-
ticularly large rotations, these assumptions break down and geometrically nonlinear analysis
is required, in which we balance internal and external forces in the structure’s deformed con-
figuration. Wingboxes, consisting almost entirely of thin shell structures, can exhibit strong
geometric nonlinearity through 3 primary mechanisms:

Large rotations: Structural elements that have a large discrepancy between in-plane and
out-of-plane stiffness, such as shells and beams, are sensitive to large out of plane
rotations as these significantly reorient the predominantly stiffened axes of the structure.

Stress stiffening: Like a taught cable, when highly loaded, in-plane stresses in shells are
great enough that they contribute to their out-of-plane stiffness when deformations
cause them to become oriented with out-of-plane external forces. Tensile in-plane
stresses lead to increased out-of-plane stiffness whilst compressive membrane stress
causes decreased out-of-plane stiffness which eventually causes buckling, a critical fail-
ure mode for shell structures.

Follower forces: Wings are primarily acted upon by pressure loads, which remain normal
to the wing’s surfaces as they deform and rotate.

In thin-walled beams subject to bending, stress stiffening due to the compressive and tensile
stresses in the upper and lower surfaces of the beam cause a flattening of its cross-section.
This effect was discovered by engineers building early aircraft wings from thin steel shells
and is named after Brazier [31] who first published work on phenomena and showed how it
can lead to the buckling of a beam’s cross-section. In a wingbox, Brazier effects lead to large
compressive loads in the wing’s ribs. These Brazier loads are typically the critical loads in the
structural sizing of ribs, particularly for buckling and consequently, structural optimization
of wingboxes using linear FE models results in unrealistically lightweight rib designs [32]
In order to perform accurate aerostructural analysis with large displacements, both the
structural and aerodynamic solvers must use geometrically nonlinear formulations. Howcroft
et al. [10] compared a selection of linear and nonlinear aerostructural analysis tools in order
to demonstrate a series of geometrically nonlinear aerostructural effects, concluding that:
• Correctly modeling aerodynamic forces as follower forces results in larger deformations
with a nonlinear structural model but lower deformations with a linear structural model.

• Linear structural models do not correctly capture the reduction in effective span that
occurs under large bending deformations and therefore overpredict lift.

• While in linear aeroelasticity drag forces are of little influence, under large tip defor-
mations drag forces on the outboard portion of the wing can induce significant wash-in
torsion and thus have a significant effect on the spanwise twist and load distributions
along a wing.
Both Garcia [4] and Smith et al. [33] demonstrated this drag-torsion effect at higher fidelity
by coupling nonlinear FE beams to CFD models of highly flexible wings. For an unswept
wing, Garcia found that the effect lead to a roughly 10% increase in total lift and a 25%
increase in wing root bending moment.

4
 There have been few published examples of aerostructural analysis using both high-fidelity
CFD and high-fidelity geometrically nonlinear structural models. Medeiros et al. [15] devel-
oped a method for constructing geometrically nonlinear reduced-order models of high-fidelity
wingbox structures which were then coupled to a RANS CFD code for both static and dy-
namic analysis but offers little discussion of the differences between linear and nonlinear
structural models. Verri et al. [34] published details of a partitioned aerostructural analysis
tool used within Embraer which couples the commercial CFD and FE codes, CFD++ and
Nastran. To the best of our knowledge, this is the only previously published work demon-
strating the coupled analysis of a high-fidelity CFD and geometrically nonlinear full wingbox
model. The authors quote a run time of one day on 200 processors to converge a single
aerostructural analysis with a nonlinear structural model. The study compares aerostruc-
tural analysis results from a 2.5g pull-up maneuver with both linear and nonlinear structures
and the results show some of the same geometrically nonlinear phenomena discussed to this
point. The geometrically nonlinear wing shows around 20% less tip washout, resulting in a
slightly lower angle of attack and a more outwardly shifted lift distribution which results in
a more than 10% increase in bending and shear loads in the outboard sections of the wing.
Unfortunately, the work offers no analysis of stress distributions in the wingbox.
Although tools like PROTEUS have been used to optimize wing structures based on
geometrically nonlinear aeroelastic analysis, few works have explicitly investigated the differ-
ences in wing designs optimized with and without considering such effects. These implications
were studied in two recent papers by Calderon et al. [35, 36], using the low-fidelity aeroe-
lastic framework NeoCASS [6]. The results of the work found that wings optimized using
geometrically nonlinear structural analysis were lighter over the entire range of aspect ratios
studied, had roughly equal aerodynamic efficiency, and had a higher optimum aspect ratio
than those optimized based on linear structural analysis.
Based on this review of the current state of the art there appears to be a knowledge
gap in the detailed aerostructural design of HARW resulting from a capability gap between
low-fidelity aeroelastic optimization frameworks that include geometric nonlinearities and
high-fidelity aeroelastic optimization frameworks which cannot. In this work, we address
this gap by expanding the capabilities of the MACH framework to include geometrically
nonlinear finite element formulations for use in high-fidelity, gradient-based, structural and
aerostructural optimization.

2 Computational Framework
The MACH (MDO of Aircraft Configurations at High-fidelity) framework represents the
current state of the art in high-fidelity aerostructural optimization. The framework includes
high-performance CFD and FE solvers with efficient adjoint derivative implementations along
with fully differentiated modules required for coupled analysis and optimization (e.g. geom-
etry parameterization, load and displacement transfer, and mesh warping) which are used at
various points throughout this work.

2.1 Aerodynamic Analysis: ADflow

The flow solver in MACH is ADflow1 , a finite-volume CFD solver for structured multiblock
and overset meshes [37]. ADflow solves the compressible Euler, laminar Navier-Stokes, and
1
github.com/mdolab/adflow

5
Reynolds-averaged Navier-Stokes (RANS) equations with a second-order accurate spatial dis-
cretization. In this work, we solve the RANS equations with the QCR-200 Spalart–Allmaras
turbulence model, including rotation corrections. The solver employs a variety of numerical
methods to converge to a steady-state solution, including multigrid, approximate Newton-
Krylov, and Newton-Krylov algorithms [38]. The combination of these various iterative
methods makes ADflow robust and fast. ADflow also solves the discrete adjoint equations,
enabling efficient computation of derivatives independent of the number of design variables.
The solution of the discrete adjoint in ADflow relies on the ADjoint approach, which uses
algorithmic differentiation (AD) to compute partial derivatives and a Krylov method to solve
the linear system [39].

2.2 Structural Analysis: TACS

The structural solver used in MACH, and throughout this work, is TACS (toolkit for analysis
of composite structures)2 [40]. TACS was developed specifically for the kinds of poorly
conditioned FE problems typical
of thin shell structures like wingboxes, which may have
9
condition numbers of O > 10 [41]. TACS also computes gradients using the adjoint method
but, unlike ADflow, all partial derivative computations are analytic.
TACS contains two geometrically nonlinear shell element formulations. A “largerot”
formulation that is fully geometrically nonlinear and a “nonlinear” formulation that uses the
nonlinear Green-Lagrange strain relationship but retains the linear kinematic relationships.
The “nonlinear” formulation is therefore only valid under small rotations but still captures
stress stiffening effects. Since the aim of this work is to accurately model HARW under large
displacements, we use the “largerot” formulation throughout and any reference to “nonlinear”
analysis refers to the use of this formulation.

2.3 Load and Displacement Transfer: Rigid Link

To transfer loads and displacements between the non-coincident aerodynamic and structural
meshes we use the rigid link transfer (RLT) method, first proposed by Brown [42]. The
method works by first finding the nearest point on the structural mesh for each aerodynamic
surface node. The translation and rotation of the base of the link is interpolated using TACS’
own shape functions before the displacement of the aerodynamic node is then computed based
on the rigid translation and rotation of the link.
It should be noted that this method is not accurate for large rotations as the term comput-
ing the aerodynamic displacement due to structural rotations is a linearized approximation.
For large rotations this term should be replaced by a rotation matrix, as described by Kennedy
and Martins [43]. In this work, however, we retain the linearized form of the displacement
transfer scheme. This simplification is not ideal and results in a slight fictitious increase in
the chord length of wing sections under twisting, but we consider it acceptable as, for the
most part, the structural and aerodynamic meshes are coincident, rendering the rotational
term in the displacement transfer scheme negligible. Additionally, the regions of the wing
that are furthest from the wing box, and are therefore most dependent on this rotation term,
are mostly affected by the wash-in/wash-out rotation of the wingbox, which is typically below
10◦ in magnitude. The RLT method is implemented within the TACS source code and so,
like TACS, computes gradients analytically.
2
github.com/gjkennedy/tacs

6
2.4 Mesh Warping: IDWarp
During aerostructural analysis and optimization, the aircraft’s outer mould line (OML) de-
forms due to both design changes and structural deformations. As these surface deformations
occur, the CFD volume mesh must also be smoothly deformed to maintain high-quality cells.
This deformation is performed in MACH by a tool called IDWarp3 [44]. IDWarp uses an
inverse-distance weighting method as proposed by Luke et al. [45]. The advantage of this
method is that it maintains good orthogonality in near-wall cells, which is vital for RANS
meshes, at a significantly lower computational cost than elasticity based approaches. This
cost is vital for efficient aerostructural analyses as the mesh must be deformed every time
structural displacements are transferred to the CFD mesh. Derivatives of the volume mesh
coordinates with respect to surface coordinates are computed using reverse-mode AD.

2.5 Coupled Aerostructural Analysis: pyAeroStructure

PyAeroStructure interacts with the aforementioned tools in order to coordinate coupled
aerostructural analyses and gradient computations. The module can perform both the
analysis and gradient computation with either block partitioned Gauss–Seidel or monolithic
Newton-Krylov approaches [16]. In this work, we use only the partitioned solver for aerostruc-
tural analysis as this method requires no modification to work with a geometrically nonlinear
structural solver. We do however make some modifications to the solver to improve per-
formance in the geometrically nonlinear case. These modifications are described in section
3.

2.6 Geometric Parameterization: pyGeo

In this work we do not optimize wing shape or planform due to issues with the geometric
gradient computation for TACS’ nonlinear shell elements. We do however require a geometric
parameterization to control the tail rotation of our aircraft in order to satisfy trim constraints.
To achieve this, we use pyGeo4 [46] which utilizes a free-form deformation (FFD) [47] ap-
proach.

2.7 Optimization: SNOPT and pyOptSparse

To perform our optimizations, we use the sparse nonlinear optimizer, SNOPT [48], a high-
performance sequential quadratic programming (SQP) optimizer that has proven to excel at
the large-scale, sparse, constrained nonlinear optimization problems typical of high-fidelity
MDO [49]. MACH interfaces with SNOPT through pyOptSparse5 [50], an object-oriented
framework that provides a general interface to many optimizers and greatly simplifies the
task of defining large sparse Jacobians that are crucial for the performance of large-scale
optimizers like SNOPT.
3
github.com/mdolab/idwarp
4
github.com/mdolab/pygeo
5
https://github.com/mdolab/pyoptsparse

7
3 Solvers for High-Fidelity Geometrically-Nonlinear Aerostructural
Analysis
3.1 Structural Solver
To solve geometrically nonlinear structural problems in TACS we implement a Newton–
Raphson based solver with line-search stabilization, adaptive load-incrementation, and a
novel solution restarting method. TACS is written in C++ but interfaces with the MACH
framework through a Python interface known as pyTACS. Whilst TACS’ has its own python
interface, which provides more or less one to one access to its C++ code, pyTACS sits on
top of this interface and provides an additional level of abstraction, simplifying the common
tasks associated with setting up and solving a finite element problem whilst still allowing
access to low-level, BLAS style linear algebra operations on TACS’ own distributed matrix
and vector objects. We have therefore been able to implement the entirety of our solver
in user-friendly python code, while the intensive numerical work is still handled in a highly
efficient and parallelized manner in the C++ layer.
In any structural finite element
 (FE) problem, we seek to solve a system of residual
equations for the displacements, u , which drive the imbalance between internal and external
forces to zero:   
R (u) = Fin (u) + Fex (u) = 0 (1)
Using the Newton–Raphson method, we can solve these equations
  by repeatedly comput-
ing a displacement update using the tangent stiffness matrix, KT , a local linearization of
the nonlinear residual equations:

  −1 
∆ui = KT (ui ) −R (ui ) (2)
h i h i
KT (ui ) = ∂u |u=ui = ∂F∂u in (u) ∂Fex (u)
   ∂R 
u=ui
+ ∂u u=ui
(3)

For many nonlinear problems, applying loads or displacements in increments can improve
convergence. Incrementation is effective because it can ensure that the current state is al-
ways close enough to the solution of the current increment to guarantee reliable Newton
convergence. Arc-length methods are an alternative to pure load or displacement imple-
mentation where both the displacement and load step are treated as unknowns that must
be solved for [51]. Doing so allows these methods to stably traverse unstable sections of a
structure’s equilibrium path and are thus necessary for simulating the kind of snap-through
and snap-back behavior seen in post-buckling analyses. In this work, we are concerned only
with modeling pre-buckling structural behavior and thus use pure load incrementation. The
conservative nature of the buckling constraints described in section 4.2 ensure that designs
produced during optimization remain stiff enough that the FE model itself does not explic-
itly buckle, which would cause convergence issues. We implement load incrementation by
redefining the residual to include a load scaling factor λ:
  
R = Fin + λ Fex (4)
To control the size of each load increment, we use the adaptive load stepping method of
Beluni and Chulya [52], where the current load step is increased or decreased based on the
actual and desired number of iterations taken to solve the previous increment.
s
Ndes
λi − λi−1 = ∆λi = ∆λi−1 (5)
Ni−1

8
 The desired number of iterations, Ndes , can then be used to control how ambitious the
solver should be with its load incrementation.
We also implement a variety of line search methods in order to stabilize our solver at larger
load increments. These include a straightforward minimum residual search, monotone and
non-monotone backtracking searches and a method proposed by Matthies and Strang [53],
where the line search seeks the point at which the residual is orthogonal to the displacement
step. In other words, finding the step size α that satisfies:
 T
E = ∆ui R (ui + α∆ui ) = 0 (6)

We refer to this method as the minimum work method as the product of a force and a
displacement represents a measure of work, even if this energy does not necessarily have an
intuitive physical interpretation. In testing the solver on representative wingbox problems, we
found the minimum work line search to be by far the best performing of these methods and,
that significant improvements in solution time could be obtained with all by foregoing the line
search on the first iteration of each load increment or when the computed Newton–Raphson
step leads to a reduction in the relevant search metric of 50% or more.
During optimization or coupled aerostructural analysis, the structural solver is called
repeatedly, with slightly different external and/or internal loads. In these cases, the previously
converged state, u∗ , no longer solves the system of equations but can be used as a good starting
point to drastically speed up the next solution. The problem in this case is to decide at what
load factor to restart the load incrementation process at. Restarting at λ = 0 would likely
waste the useful initial guess u∗ but, if the design or loading has changed significantly, it may
not be possible to converge the problem starting at the full load factor, λ = 1.
To address this problem we implement two methods for computing an “optimal” restart
load factor, λ∗ , both of which find the load factor that minimizes some measure of the solution
error. The first method finds the load factor that minimizes the resulting residual norm, a
method first proposed by Bergan [54]. The residual norm is minimized when the load factor
is chosen such that the residual is orthogonal to the external force vector. The residual
minimising load factor can then be computed as:
 T
 T  T   ∗

− Fex Fin
Fex R = Fex λ Fex + Fin (u ) = 0 ⇒ λ = 2
(7)
||Fex ||
In shell problems, terms representing the in-plane, out-of-plane and rotational residuals
can have drastically different scales, making a Euclidean norm a poor indicator of the true
magnitude of solution error. To redress this problem, many use energy-based measures of er-
ror, such as the minimum work line search criteria of Matthies and Strang [53] just described,
or the energy-based convergence criteria proposed by Bathe and Dvorkin [55]. In this spirit,
we propose a novel method that minimises the strain energy error, defined as the product of
the residual and the displacement step resulting from that residual:
T −1  T −1 
Fin + λ∗ Fex
   
∆E = R KT R = Fin + λ Fex KT (8)

Differentiating this expression with respect to the load factor and solving for the minimum
energy error gives:

9
  
T
 T
− Fex + ∆ue
∆ui Fin
λ∗ =  T
(9)
2 ∆ue Fex
  −1    −1 
∆ui = KT Fin , ∆ue = KT Fex (10)

This minimum energy restart method requires the expensive factorization of the current
tangent stiffness matrix but, as explained previously, TACS’ direct solution method means
that this factorization can be reused to compute both ∆ui and ∆ue before being reused
again for the first iteration of the restarted solution, meaning that the method does not add
a significant cost to the solution process.
If the optimal load factor is closer to the final desired load factor than the user-defined
initial increment size then it is used as the initial load factor for the solution process. If
the first load increment in the restarted solution fails to converge the solution is reset and
started from the original initial load factor. In the case the external loads have reduced or
the structure has been made stiffer since the previous solution, the optimal load factor may
be greater than 1, in which case the direction of load incrementation is reversed.

3.2 Aerostructural Solver

To perform geometrically nonlinear aerostructural analysis, we use the partitioned Gauss–
Seidel solver presented by Kenway et al. [16], where each aerostructural iteration involves
approximately solving the aerodynamic and structural problems separately and the transfer of
loads and displacements between each. The method also uses displacement under-relaxation
with an adaptive relaxation factor chosen using Aitken acceleration.
We implement our nonlinear FE solver such that it can be accessed through the same calls
to the pyTACS interface as the linear solver. As such, no modifications to the source code of
pyAeroStruct are strictly necessary in order to perform geometrically nonlinear aerostructural
analyses. We do however implement two new features which aim to reduce solution times
when using a nonlinear structural formulation.
First, because nonlinear systems do not abide by the principle of linear superposition,
the
 structural state resulting from the aerostructural solver’s displacement under-relaxation,
us , is typically not close to a solution for any scaling of the applied loading. It is therefore
typically not efficient, or even possible, to restart the subsequent structural analysis from
this damped state until late in the aerostructural solution process when displacement updates
become very small.  To remedy this, we save the converged solution of each structural analysis
before damping, ûs , and restart the next structural analysis from this undamped state.
The second feature stems from the fact that in the initial iterations of the aerostructural
solution process, the aerodynamic forces are often drastically overpredicted, potentially by
a factor of 2-3. This is due to the partial convergence of the CFD solver and the lack of
any load alleviation from the wing in its undeformed state. For a linear structural analysis,
this poses no problems as the cost of solving the pre-factorized linear system is negligible
and independent on the magnitude of the loading. However, assuming a consistent load
incrementation rate, the time to solve a nonlinear structural analysis is approximately linearly
proportional to the loading magnitude, meaning the first few structural analyses may take
2-3 times longer than usual. In the worst case, it may not be possible to solve the problem
at all if unrealistically high aerodynamic forces result in explicit buckling of the FE model.

10
 To avoid this potential issue and hopefully improve the speed of the initial structural
analyses, we use the load ramping method proposed by Sanchez et al. [56]. We do this by
ramping the maximum load scale to which the structural solver increments over the first N
Gauss–Seidel iterations following a smooth polynomial ramping function:
 n 3  n 2
λramp = −2 +3 (11)
N N
We use a fixed under-relaxation factor during the load ramp before resuming Aitken accel-
eration once the ramp is complete. To enable efficient restarting of the aerostructural solver
without restarting the load ramp from zero, we compute the optimal restart load factor using
the same minimum energy method described earlier before each call of the structural solver.
This value is then clipped to lie between the current load ramp value and 1, before being
used as the maximum load scale for the structural analysis.
We demonstrate the power of this load-ramping scheme by performing a series of geomet-
rically nonlinear aerostructural analyses on a flexible wing in Mach 0.85 flow at an angle of 8◦ .
This is a challenging case to solve due to the large displacements of the wing (22% semispan
vertical tip displacement when converged) and the presence of strong shocks and significant
flow separation in the early solution stages. Figure 2 shows the convergence histories from
a selection of the analyses, showing the norms of the aerodynamic and structural residuals
and the structural displacement vector. The load ramping results in a lower solution time
in all cases but there is no clear trend with both short (N = 2) and long (N = 8) ramps
achieving a 40% speed-up over a non-ramped solution, while an intermediate-length ramp
(N = 4) achieving only a 13% reduction. The benefit of load-ramping therefore appears to
come mostly from a reduction in the structural solution time in the first few Gauss–Seidel
iterations rather than from the stabilization of the solver.
100
f
f, 0
10 6

s 100
s, 0

10 6
300 No Load Ramp N=2 N=4 N=8
||us|| 200
100
0 365 516 593
Time (s)
Figure 2: Both short (N = 2) and long (N = 8) load ramping stages can achieve a 40%
speedup over the non-ramped solution. Most of the gains are made in the early Gauss–Seidel
iterations.

Our modified solution process is presented in algorithm 1, where the highlighted sections
indicate our modifications to the solver developed by Kenway et al. [16].

11
 Algorithm 1: MACH’s Gauss–Seidel partitioned aerostructural solver with addi-
tions for geometrically nonlinear analysis highlighted.
n o n o
Given: us
(0)
, u(0)
f
, θ(1)
while
n Notoconverged do
(n) (n−1)
Xf = TX (us ) . Transfer structure displacements and deform aero mesh
n o    
(n) (n) (n) (n−1) (n)
Find uf s.t Rf uf , Xf ≤ δf,rel Rf uf , Xf . Approximately solve

naero o
(n) (n)
Fs = TF (uf ) . Transfer aerodynamic forces to structure
if Rampingthen
n oT n o n oT n o
− ∆ui Fex + ∆ue Fin
λ∗ = n oT n o . Compute optimal restart load factor
2 ∆ue Fex
if λ∗ > 0.95 then
λ∗ = 1
Ramping = False . End ramping if λ∗ close to unity
end
n 3 n 2



λramp = −2 N +3 N . Compute ramp load factor
λ(n) = max (λramp , λ∗ ) . Choose maximum load factor
else
λ(n) = 1
end n o  
(n) (n)
Find û(n)s s.t Rs us , λ(n) Fs ≤ δs,rel λ(n) kFex k . Solve structure with scaled
load
n o n o n o
(n)
∆us = û(n)
s − u(n−1)
s
if n > 1 then
if Ramping then
θ(n) = θ(n−1)
else  T  
(n) (n−1) (n)
∆us − ∆us ∆us
θ(n) = 1 − (n) (n−1) 2 . Update under-relaxation factor using
∆us −∆us
Aitken acceleration
end
end
n o n o n o
(n)
us = u(n)
s + θ(n) ∆u(n)
s . Apply under-relaxed displacement increment
end

4 Aircraft Model
All the results presented in the subsequent sections of this paper come from analysis and opti-
mization of the undeflected common research model (uCRM), a publicly available6 benchmark
aerostructural model based on the NASA common research model (CRM) aerodynamic ge-
ometry. Two variants of the uCRM were developed by Brooks et al. [57]. The uCRM-9 is
designed to be an exact aerostructural replica of the original NASA CRM and thus shares
the same planform definition, while the uCRM-13.5 is a high aspect-ratio version of the same
6
https://mdolab.engin.umich.edu/wiki/ucrm.html

12
 Table 1: Key parameters of the uCRM planforms [57].

Parameter uCRM-9 uCRM-13.5 Unit

Aspect-ratio 9 13.5
Taper-ratio 0.275 0.25
Span 58.76 72.0 m
Root chord 11.92 11.07 m
Tip chord 2.736 2.06 m
Reference area 383.74 383.78 m2
1/4 chord sweep 35 35 ◦

Table 2: Flight conditions used in this work.

Condition Mach Lift Altitude (ft)

2.5 g maneuver 0.64 2.5×MTOW 0
-1 g maneuver 0.64 −MTOW 0

aircraft. The uCRM-13.5 wing retains the same reference area, sweep and 1/4 chord mean
aerodynamic chord (MAC) position in order to remain feasible with the same fuselage, en-
gine and tail configuration. While the uCRM-9 is typical of the aspect-ratios seen in current
transport aircraft, the uCRM-13.5 represents the kind of wing aspect-ratios that may be
present on the next generation of transport aircraft entering service in the next two decades.
For this reason, we focus exclusively on the uCRM-13.5 in this work in order to study the
importance of considering geometric nonlinearity in the design of next-generation transport
aircraft.
As is explained in a later section, we are unable to optimize the wing’s geometry when
using nonlinear element formulations in this work and so focus solely on optimization of the
structural design of the uCRM wing, albeit considering coupled aeroelastic behavior. For this
reason, we consider only two flight conditions; a 2.5 g and a -1g maneuver condition, both of
which are summarized in Table 2.

4.1 Aerodynamic Model

Since this work is limited to symmetric flight conditions we use half-meshes of the uCRM
for CFD analyses. The aircraft OML used for these analyses is a wing-body-tail geometry,
with no rudder or engine nacelle. The publicly available model contains CFD meshes in
both structured multiblock and overset form. The coarse, medium and fine versions of these
meshes contain approximately 1, 3 and 9 million cells respectively. Unless otherwise specified,
we use the coarse multiblock mesh throughout this work.
When performing aeroelastic analysis and optimization, we trim the uCRM by varying
its angle of attack and horizontal tail rotation. This twisting of the horizontal tail is achieved
using the FFD approach described in section 2.6. As shown in figure 3, this involves applying
a rigid rotation to the FFD control points enclosing the tail.

4.2 Structural Model

The uCRM-13.5 wingbox contains upper and lower skins, trailing and leading-edge spars, 58
ribs oriented perpendicular to the leading edge spar and an additional engine mount panel

13
Figure 3: The FFD volumes used to parameterise the uCRM-13.5, with the volume controlling
tail rotation highlighted.

between the two ribs at the Yehudi break. Each rib and each section of the contiguous
skins and spars between the ribs is considered a separate panel whose sizing variables can
take on their own values. The wingbox is rigidly clamped at the symmetry plane of the
aircraft and the outer edge of the rib at the wing-fuselage junction is fixed in the vertical and
chordwise directions. The FE meshes of the wingbox consist of 4-node, second-order, MITC
shell elements, with the coarse, medium and fine meshes having approximately 1.5, 3.5 and
5.5 × 105 degrees of freedom. As with the CFD meshes, we use the coarse mesh throughout as
Brooks et al. [57] showed that stresses were more or less constant across the three structural
mesh levels.
In the skin panels, blade stiffeners are aligned with the leading edge spar, thus the stiff-
eners in the skin sections inside the fuselage are not swept whilst those outside the fuselage
are. The stiffeners on the rib and spars are oriented vertically. We do not model these
stiffeners directly but instead include their effect by adding equivalent stiffness terms to the
skin shell elements as described by Brooks et al. [21], essentially “smearing” the stiffness of
stiffeners over the entire panel area. As shown in figure 4b, this smeared stiffener model is
parameterized by 3 additional design variables per panel for the thickness, height, and pitch
of the stiffeners in addition to the skin thickness variable. This parameterization is one of
the primary advantages of the smeared stiffener approach over explicitly modeled stiffeners,
which can only easily be given thickness design variables. Material failure is predicted using
the Von-Mises criterion at 3 locations through the thickness of the virtual panel; the upper
and lower surface of the skin, which are in a 2D plane stress state, and at the extrema of the
stiffener blade, which is assumed to be under purely axial stress.
Another advantage of the smeared stiffener model over explicit stiffeners is the ability to
predict buckling failure without the need for any additional eigenvalue based analyses. To
do this, TACS defines a force based failure envelope:
2
Fxy
Fx
+ 2 ≤1 (12)
Fx,cr Fxy,cr
Based on critical compressive, Fx,cr , and shear, Fxy,cr , loads in the local axis of the panel
stiffeners. At each element these critical loads are calculated using relations presented by
Stroud and Agranoff [58] for global panel buckling, inter-stringer skin buckling and stringer
buckling of an infinitely wide flat panel of the same length as the wingbox panel the element
is a part of [21].
We use aluminum as the material of the wingbox, with an elastic modulus of 70 GPa, Pois-
son ratio of 0.3, density of 2780 kg m−3 and yield strength of 420 MPa. Although the latest

14
 tskin

tstiff hstiff

λstiff
(b) The parameterization of a smeared stiffener panel in TACS.
(a) The boundary conditions ap-
plied to the uCRM wingboxes [57].

Figure 4: Structural modelling of the uCRM-13.5

generation of transport aircraft sport almost entirely composite wing structures, compos-
ites significantly increase the complexity of the structural design parameterization and the
resulting optimization problems. Studying only aluminum structures therefore allows this
work to focus solely on differences in analysis and optimization results caused by geometric
nonlinearity rather than the minutiae of composite design optimization.

5 Analysis Studies
Before moving to comparing the results of high-fidelity design optimizations with linear and
nonlinear structural formulations, we compare structural and aerostructural analyses under
identical conditions with the two formulations. Doing so will allow us to more easily trace
differences in optimized designs back to modeling effects.

5.1 Structural Analysis

We first perform pure structural analysis of the uCRM-13.5 with linear and nonlinear element
formulations under identical loading. We apply aerodynamic loads generated from linear
aerostructural analyses at the two maneuver conditions mentioned previously along with fuel
loads representing a full fuel load and inertial loads from the wingbox itself as well as the
engine and leading-edge/trailing-edge devices.
Figure 5 shows the distributions of the Von Mises and buckling failure criteria in the
uCRM-13.5 wingbox in the 2.5 g maneuver condition. The wingbox on the right of each
figure shows the linear analysis values subtracted from the nonlinear analysis values, leaving
red regions indicating areas with higher stresses in the nonlinear case while blue areas show
areas where stresses are lower in the nonlinear case. Also shown is the relative difference in
these criteria along a slice taken through the upper skin of the wingbox.
The presence of almost no blue in the rightmost wingbox in each figure indicates that
the stresses in the nonlinear analysis are equal to or higher than those in the linear analysis
throughout the entire wingbox. In particular, over the majority of the upper and lower
wing skins, the Von-Mises stress is 10-20% greater in the nonlinear case. This increase in
bending stress is due to the nonlinear analysis correctly maintaining the bending moment
produced by the lift forces on the outboard portions of the wing as they rotate inboard with
the wing. Since the linear analysis only considers the undeformed configuration of the wing,
only vertical force components induce a bending moment and the inboard component of the

15
rotated lift force is felt simply as an axial compression of the wingbox. The increase in the
buckling criterion in the upper skin is roughly half the magnitude seen in the Von-Mises
stress, peaking at 10%. A much greater increase in the buckling criterion is seen in the ribs,
strongly suggesting the presence of Brazier loads.
To confirm this suggestion, figure 6 compares the axial stress in the stiffener-wise (verti-
cal) direction of the ribs. The stress shown is computed at the mid-section of the shell, thus
negating any axial stresses due to bending. The results show a drastic increase in the com-
pressive stress in the ribs, particularly in the region just outboard of the engine mount where
the buckling criteria also increased greatly. The compressive stresses in the linear analysis,
if present at all, are smaller than in the nonlinear case by at least an order of magnitude.

(a) Comparison of Von-Mises criterion distribu-

(b) Comparison of buckling criterion distribution.
tion.

Von Mises 0.20

fNL fL 0.15
fL
0.10

Buckling 0.05

0.00 0 10 20
Root Fuselage Engine Fuel Tank Tip
Junction Mount End

(c) Relative differences in the two failure criteria along the upper wing skin.

Figure 5: Nonlinear analysis under identical loads leads to a 10-20% increase in bending
stress in the uCRM-13.5 wingbox.

Figure 7 shows the chordwise, spanwise and vertical components of the displacement and
wing twist, all extracted from points along a line on the upper skin of the wingbox. The
twist distributions in the linear and nonlinear cases are almost identical, with the nonlinear
case showing fractionally more washout. This is caused by higher bending curvature in the
nonlinear case, which results in more washout due to the wing’s geometric bend-twist coupling
and is more than enough to cancel out any decrease in washout due to drag-torsion effects
which appear to be negligible. This is a somewhat predictable outcome given that, in a well-
designed wing, there should be very little drag, or even a small amount of thrust produced
by the wing-tip.
Also shown in figure 7 are the chordwise (∆X), spanwise (∆Y ) and vertical (∆Z) com-

16
Figure 6: The presence of Brazier loads in the nonlinear results is confirmed by examining
the axial stress in the ribs.

ponents of the displacement. From these curves it is possible to quantify the effective tip-
shortening in the nonlinear case, around 3% of the wing’s semispan. Also of note is that the
chordwise deformation is reversed between the two cases with the wing tip moving forwards
in the nonlinear case. This is due to the same bending kinematics that cause the tip shorten-
ing effect of the wing, which result in a forward displacement of the wing-tip due the wing’s
backward sweep.

0.2721 0.000
X (m) 0.209
Linear 0.0000

Nonlinear Y (m)
0.2719 1.075
7.008 0.0
Z (m)

y (deg)
0.000 11.6
0.0 0.5 1.0 0.0 0.5 1.0

Figure 7: Geometrically exact bending kinematics correctly capture the inward and forward
deformation of the wingtip under bending but there appears to be no strong drag-torsion
effects.

5.2 Aerostructural Analysis

We now compare geometrically linear and nonlinear aerostructural analyses, both trimmed
to the same 2.5g lift level, to demonstrate how the structural nonlinearities just discussed

17
Table 3: Aerodynamic functions of interest from the trimmed aerostructural 2.5 g maneuver
analyses

Value Units Linear Nonlinear ∆ (%)

α ◦ 5.082 5.097 0.3
θtail ◦ 0.373 0.624 67.5
CL 0.586 0.586 0.0
CL,wing 0.495 0.491 -0.9
CL,tail 0.0115 0.0148 28.9
CL,fuse 0.0798 0.0808 1.2
CD cts 264.3 269.6 2.0
CD,wing cts 113.5 115.4 1.7
CD,tail cts 19.57 22.17 13.3
CD,fuse cts 131.2 132 0.6
CM y 0 0 0
CMy ,wing -0.0586 -0.0411 -29.9
CMy ,tail -0.0584 -0.0745 27.4
CMy ,fuse 0.117 0.115 -1.3
COPwing,x m 0.684 0.483 -29.3
COPwing,y m 14.02 13.73 -2.1

affect the aeroelastic performance of the aircraft. We use the fine CFD mesh to maximize the
accuracy with which the aerodynamic forces are computed but retain the coarse FE mesh.
Table 3 presents some aerodynamic quantities of interest from the two analyses, including
a component-wise breakdown of the lift, drag and pitching moment coefficients, and the
coordinates of the centre of pressure (COP) of the wing relative to the uCRM’s centre of
gravity (COG)7 .
The majority of the differences in these results can be attributed to the geometrically
nonlinear bending kinematics discussed in the previous section, which changes the intricate
balance of lift and pitching moment between the wing, tail and fuselage. In the nonlinear
case, the inward deflection of the wing caused by geometrically nonlinear bending kinematics
reduces the effective span of the wing by 3.7%, it is therefore initially surprising that the non-
linear case requires only a 0.3% increase in angle of attack to achieve the same lift coefficient
as the linear case. However, the same bending kinematics also shift the wing’s COP forward,
reducing the pitching moment arm of the wing’s lift around the COG by almost 30%. The
horizontal tail is therefore required to produce almost 30% more lift in the nonlinear analysis.
This increase in tail lift, and a small increase in fuselage lift due to the increased angle of
attack, counteract the loss of wing lift due to span shortening and thus reduce the increase
in the angle of attack required to achieve the same lift. The nonlinear case shows a 2%
increase in drag which comes mostly from the wing and tail. The increase in tail drag can
be attributed to its increased level of lift while the source of the wing drag increase is less
clear. We propose it may be caused by an increase in induced drag due to the wing’s lower
effective span.
Figure 8 compares the wingbox stress distributions from the two aerostructural analyses.
7
The COP is not a unique point and can be defined anywhere along the line of the resultant aerodynamic
force. We compute the location of the COP in the same horizontal plane as the COG

18
Contrary to the results of our structural analysis comparison, in these cases, the geometrically
nonlinear analysis leads to a reduction in the bending stress in the wing, with the Von-Mises
and buckling failure criteria being reduced by up to 15% in the upper and lower wing skins.
Some of this decrease can be attributed to the inboard shift of the wing’s COP, but we consider
it unlikely that this entirely explains these results as it represents only a 2% reduction in the
moment arm of the wing’s lift force. Also of note, is that, although the stresses over the
majority of the wing skins is lower, the peak stresses are higher, by around 10%. These
peak stresses occur around the engine mount and we primarily attribute their increase in the
nonlinear to be the result of the simplistic manner in which the inertial load of the engine is
introduced to the structure, which is not valid under finite rotations.

(a) Comparison of Von-Mises criterion distribu-

(b) Comparison of buckling criterion distribution.
tion.
Buckling 0.00

0.05
fNL fL
fL 0.10

Von Mises 0.15

0.20
0 10 20 30
Root Fuselage Engine Fuel Tank Tip
Junction Mount End

(c) Relative differences in the two failure criteria along the upper wing skin.

Figure 8: In contrast to structural analysis, geometric nonlinearity in aerostructural analysis

leads to an up to 15% decrease in bending stress in the uCRM-13.5 wingbox.

6 Optimization Studies
We now present the results of a series of structural and aerostructural optimizations of the
uCRM-13.5 In doing so we aim to assess how geometrically nonlinear phenomena affect the
optimal structural and aeroelastic design of high-aspect-ratio wing (HARW).
We first present structural optimizations performed under the same fixed aerodynamic
loads used in section 5.1. Next we perform aerostructural optimization of the uCRMs, using
an identical problem formulation but utilizing coupled aerostructural analysis. Our initial
intent would have been to simultaneously optimize the wing shape, planform, and struc-

19
tural sizing, as has been done in many previous works using MACH. However, issues with
derivative computations with respect to geometric design variables with TACS’ nonlinear
element formulations restrict us to considering only structural sizing in these studies. These
optimizations still provide useful insight as, unlike in a structural optimization under fixed
loads, when using coupled aerostructural analysis and adjoint derivatives, the optimizer is
able to aeroelastically tailor the wingbox in order to positively influence the aerodynamic
loads it experiences. In all the optimizations we present herein, the objective function to be
minimized is the mass of the wingbox.

6.1 Structural Optimization

6.1.1 Design Variables
Figure 9 and Table 4 summarise the parameterization of the uCRM-13.5 structural design.
Each of the 283 panels in the wingbox is parameterized by four design variables controlling the
thickness of the panel skin and the thickness, height and pitch of the smeared panel stiffeners.
In this parameterization, all panels have design variables for skin thickness, stiffener thickness
and stiffener height, we also give the rib and spar panels stiffener pitch variables, while the
stiffener pitch across all upper and lower skin panels are controlled by two global design
variables. This leads to a total of 1029 structural sizing variables. This is a departure from
the parameterization originally used in the conception of the uCRM-13.5, where all ribs
shared the same stiffener design variables [57]. We however found, in early optimizations,
that this approach lead to the majority of ribs being greatly over-sized, rendering the effect
of Brazier loads negligible.

Figure 9: Structural parameterization of the uCRM-13.5 wingbox.

6.1.2 Constraints
We apply constraints on the Von Mises buckling failure criteria in the 2.5 g and -1 g maneuver
conditions. We split the wingbox into four separate regions, upper skin, lower skin, ribs and
engine mount, and spars, within which, the element Von Mises and buckling values are
combined into a single constraint value using Kreiselmeier-Steinhauser (KS) aggregation.
In the 2.5 and 1 g load cases, we apply Von Mises constraints to all regions and buckling
constraints on all regions except the lower skin. In the -1 g case we constrain only the lower

20
 Table 4: Summary of the design variables used for the uCRM structural optimisation.
Components Quantity Bounds
Variable Description Lower Skin Upper Skin Ribs Spars Lower Upper
tskin Local Skin Thicknesses 3 3 3 3 286 3 mm 250 mm
tstiff Local Stiffener Thicknesses 3 3 3 3 286 2.5 mm 250 mm
hstiff Local Stiffener Heights 3 3 3 3 286 25 mm 200 mm
λstiff local Stiffener Pitch 3 3 169 150 mm 300 mm
λstiff Global Stiffener Pitch 3 3 2 150 mm 300 mm
Total 1029

skin buckling value. This gives a total of 8 failure and buckling constraints, to which we
apply a safety factor of 1.5.
To enforce a realistic structural design, we apply a series of linear constraints to the
panel design variables. On all panels, we constrain the stiffener and skin thicknesses to be
within 2.5 mm of each other. Additionally, to avoid abrupt changes in panel sizing we apply
adjacency constraints to limit the change in sizing variables between adjacent panels and
consecutive ribs. The change in skin and stiffener thicknesses is limited to 2.5 mm and the
change in stiffener height and pitch to 10 mm. These constraints, although numerous, are
linear and are therefore handled efficiently by SNOPT, without the need for repeated gradient
calculations. In total, this gives 1284 constraints, which are summarized in Table 5.

Table 5: Constraints applied to the baseline structural optimisation problem.

Components Quantity
Constraint Description Lower Skin Upper Skin Ribs Spars
1.5KSvm, 2.5g 2.5 g Von Mises failure 3 3 3 3 4
1.5KSbuckling, 2.5g 2.5 g buckling failure 3 3 3 3
1.5KSbuckling, -1g -1 g buckling failure 3 1
tskin,i − tstiff,i Stiffener-skin thickness difference 3 3 3 3 286
tskin,i − tskin,i+1 Skin thickness adjacency 3 3 3 3 275
tstiff,i − tstiff,i+1 Stiffener thickness adjacency 3 3 3 3 275
hstiff,i − hstiff,i+1 Stiffener height adjacency 3 3 3 3 275
λstiff,i − λstiff,i+1 Stiffener pitch adjacency 3 3 165
Total 1284

6.1.3 Results
Figure 10 shows the history of the mass of each wingbox component group during the opti-
mization and the percentage difference between the optimized masses using linear and non-
linear FE formulations. Both the linear and nonlinear optimizations achieve a significant
decrease in mass over the baseline uCRM wingbox design and satisfy both the optimality
and feasibility convergence criteria. The wingbox optimized with nonlinear analysis is 7%
heavier than the linear optimized design, an increase that is consistent with the increased
bending stresses observed in the structural analysis studies. This mass increase from linear
to nonlinear optimized wingbox is greatest in the two skins and the ribs.
In figures 11 and 12 we plot the spanwise distribution of the structural sizing of wing
skins and ribs. Rather than plot the distribution of each panel sizing variable separately, we
plot what we refer to as aggregated sizing values, which combine the stiffness contribution of
both the panel skins and stiffeners into a single value. The “axial effective thickness” is the

21
 6000
2.5g Upper
Skin Mass
Linear Nonlinear
5000 6.8%
6000
2.5g Lower
Skin Mass 5000
7.7%

2.5g Rib+mount 1600

Mass
7.2%
1400

2.5g Spars 700

Mass
600
3.8%
0 800
Major Iteration

Figure 10: Structural optimization of the uCRM-13.5 with nonlinear structural analysis
results in a 7% greater mass than with linear analysis. The greatest differences are in the
skins and ribs.

thickness of an equivalent unstiffened panel with the same axial stiffness (in the stiffener-wise
direction) as the stiffened panel, this is a proxy for both the axial strength and mass of each
panel. The “bending effective stiffness” follows the same rationale but for bending stiffness
and is thus a proxy for the buckling resistance of each panel.
The sizing distributions of the upper and lower wing skins in both optimized designs are
qualitatively similar save for a few small discrepancies. The increase in the axial effective
thickness of the upper and lower skins from linear to nonlinear optimized designs is relatively
uniform along the entire wingspan which is again consistent with the increase in bending
stresses observed in the structural analysis studies. The exception to this are the peaks in
thickness that occur regularly every 4 panels along the upper skin in the nonlinear optimized
design, particularly in the region just outboard of the engine mount. These spikes are caused
by distorted elements in the FE which lead to unrealistic stress spikes and thus require an
overly reinforced panel to satisfy the maximum failure constraints. The other interesting
difference in the two designs appears near the root of the lower skin. The bending stiffness
of the linear optimized lower skin is significantly higher over the portion of the wing between
the fuselage junction and engine mounting location, perhaps indicating that the compressive
stresses seen in the -1 g maneuver condition are greater in the linear case. In contrast, the
bending stiffness of the lower skin of the center wingbox (inside the fuselage) is more than
8×8 greater in the nonlinear optimized design. The rationale for this increase is unclear,
as the lower skin of the central wingbox section is not buckling critical in either design, we
believe that the optimizer may have found that the greatly increased bending stiffness in this
region allows it to save material in surrounding regions of the wingbox.
The impact of Brazier loads on the rib sizing is clear, with the optimizer having given the
ribs between the fuselage junction and engine mount noticeably greater bending stiffness in
the nonlinear optimized design. It is worth noting, however, that the axial effective thickness
of these ribs does not increase significantly, meaning that the optimizer was able to shift
8
The bending effective thickness is only greater by a factor of 2 but bending stiffness is proportional to the
cube of thickness.

22
 25 Linear Nonlinear
20
Axial
Effective 15
Thickness
(mm)
10
5

50 U_SKIN
L_SKIN
Bending 40
Effective
Thickness 30
(mm)
20
10
Root Fuselage Engine Fuel Tank Tip
Junction Mount End

Figure 11: The increase in skin stiffness in nonlinear optimized design is consistent with the
increase in stress seen in the structural analysis studies.

material in the rib panels from the skin to the stiffeners, without increasing their overall
mass. Likely, the majority of the difference in rib mass between the linear and nonlinear
optimized designs occurs in the fuselage junction rib which sees significantly greater loading
than the rest of the ribs.
20
Linear Nonlinear
15
Axial
Effective
Thickness
(mm) 10

40 RIBS
Bending 30
Effective
Thickness
(mm) 20

10
Root Fuselage Engine Fuel Tank Tip
Junction Mount End

Figure 12: Brazier loads necessitate greater bending stiffness in the ribs near the engine
mount and fuselage junction.

6.2 Aerostructural Optimization

Finally, in this section, we perform aerostructural optimization of the uCRM wingboxes.
The formulation of the aeroelastic optimization problem is near-identical to the structural
optimization formulation with the following changes, aimed at reducing the computational
cost and complexity of the problem. First, we revert to a parameterization more similar to

23
that used in the original design of the uCRM-13.5, where stiffener heights and pitches on all
ribs are and all spars are identical, reducing the number of design variables to 694 and the
number of linear sizing constraints to 835. Additionally, we reduce the allowable change in
sizing variables between adjacent panels to 0.5 mm.
To ensure the aircraft is correctly trimmed in each flight condition, we place constraints
on the total pitching moment of the aircraft and on the total lift, which must equal the
uCRM-13.5’s maximum take-off weight (MTOW). To satisfy these constraints, the angle of
attack and tail-rotation at each flight condition are added as design variables. We perform
a pre-optimization trimmed aerostructural analysis to obtain the values of these variables
that ensure our initial design satisfies these trim constraints. To increase the aerostructural
coupling of the optimization problem, with the hope of accentuating any differences in linear
and nonlinear optimized designs, we assume that the mass of the wingbox affects the aircraft’s
MTOW and consequently the lift required in each flight condition. As was done by Brooks
et al. [57], we multiply the wingbox mass by 1.25 to produce an estimate of the total wing
mass, we then add this mass to a reference mass which is calibrated such that the baseline
uCRM wingbox design results in the original uCRM-13.5 MTOW of 268 × 103 kg.
We run each flight point in parallel on 144 Intel Xeon Platinum 8160 Skylake processor
cores and run the optimization for 2 days. In this time, the optimizations achieve only
intermediate convergence, producing designs which are significantly lighter than the initial
point, and are feasible to within 2% of the failure criteria, but do not appear to have quite
converged to an optimum. We however believe that the results still offer some useful, if not
fully conclusive insights into the effect of geometric nonlinearity on optimal HARW design.
Figure 13 contains two separate comparisons; on the left comparing between the linear and
nonlinear aerostructural optimizations, and on the right comparing between the nonlinear
aerostructural optimization and a nonlinear structural optimization with the same problem
formulation but using fixed loads, as was done in section 6.1.
The difference in mass between the linear and nonlinear aerostructurally optimized designs
is 4.5%, slightly less than the 7% difference seen between the structurally optimized designs
in section 6.1. This is somewhat surprising given that in section 5.2, geometrically nonlinear
aerostructural analysis resulted in lower bending stresses in the majority of the wingbox.
We believe that the nonlinear optimized design may end up heavier despite this due to the
increased peak stress around the engine mount which, due to the tight adjacency constraints
used in this problem, ultimately dictates the sizing of a significant portion of the wing.
Another cause of this mass increase appears to be that, the linear optimized design is able to
achieve slightly more passive load alleviation, judging by the lift and twist distributions in
figure 13e which show the linear design exhibits more outwash over the midspan of the wing
and thus has a slightly more inboard lift distribution than in the nonlinear case. Comparing
the structural sizing of the two designs in figure 13c, this appears to be because the optimizer
has reinforced the leading-edge spar significantly more around the midspan in the linear case.
We cannot however, be sure that the optimizer would not have matched this reinforcement
in the nonlinear case, given more time.
The difference between the aerostructurally and structurally optimized designs is also
primarily in the prioritization of reinforcement towards the leading-edge spar. This is because,
during the structural optimization, where the optimizer has no influence over the aerodynamic
loading of the wing, the most effective design strategy is to concentrate reinforcement as far
from the neutral axis of the wingbox as possible, in the upper and lower skins. However,
when we optimize with coupled aerostructural analysis, the optimizer is able to induce more
passive load-alleviation in the wing by shifting reinforcement to the leading-edge spar. This

24
reduces the bending moment induced in the wingbox and allows for a lighter structure, which
in turn reduces the amount of lift that must be produced by the wing and leads to further
weight savings. All in all, this leads to the aerostructurally optimized wing being 25.7%
lighter than the structurally optimized wing.
10 1
10 2

10 2
Linear Nonlinear
Optimality 10
4
Optimality
10 3 10 6 Struct Opt Aerostruct Opt
10 2
10 4

Feasibility 10
3
Feasibility
10 8
10 4

15000 15000

2.5g Mass 12500 2.5g Mass 12500

-25.7%
10000 10000
4.5%
0 137 169 0 137 321
Major Iteration Major Iteration

(a) Optimization history. (b) Optimization history.

(c) Sizing distributions. (d) Sizing distributions

2 2
1 Linear 1 Struct Opt
Normalised 0 Nonlinear Normalised 0 Aerostruct Opt
Lift Lift
1 Elliptical 1 Elliptical
2 2
0 1 0 1

0 0
Twist ( ) Twist ( )
5 5
2.5g 2.5g
-1g -1g
10 10
0 1 0 1

(e) Normalized lift and twist distributions. (f) Normalized lift and twist distributions.

Figure 13: A comparison between linear and nonlinear aerostructurally optimized wings
(left), and nonlinear structurally and aerostructurally optimized wings (right).

25
7 Computational Cost
One of the reasons we believe there have been few, if any, examples of high-fidelity, geomet-
rically nonlinear aerostructural analysis and optimization published is the greatly increased
cost of solving nonlinear FE problems as compared to linear ones. In this work, we therefore
aimed not only to demonstrate the capability of performing such analyses and optimiza-
tions, but also to quantify the increase in cost associated with the switch to a geometrically
nonlinear formulation. For standalone structural analyses with our solver, the addition of
geometric nonlinearity increases the computational cost by a factor of around 20 over linear
analyses. This increase is directly proportional to the number of Newton–Raphson iterations
taken to solve the nonlinear problem and we therefore expect it to vary depending on the
stability and load factor of the problem at hand. When performing structural optimization,
the nonlinear structural solver can be restarted from the last converged solution each time it
is called which reduces this slowdown factor to around 10. When performing aerostructural
analysis, we typically observed a much smaller cost increase of around 40% as the majority
of the solution time is spent in the CFD solver. This of course greatly depends on the rela-
tive sizes of the fluid and structural meshes. In the aerostructural optimizations performed
in this work, the computation time is dominated by the coupled adjoint solver which is no
more expensive when using a nonlinear structural formulation and the penalty for adding
nonlinearity therefore reduces to around 15%.

8 Conclusions
As new airframe technologies and improved computational design tools allow for the aspect-
ratio of aircraft wings to grow, the ability to model geometrically nonlinear aeroelastic be-
havior is increasingly critical. However, to date, this has only been possible with low fidelity
tools. In this work we presented examples of structural and aerostructural optimization using
high-fidelity RANS CFD and geometrically nonlinear shell FE models.
We implemented a Newton–Raphson based solver with multiple line search methods and
an adaptive load incrementation strategy within the MACH framework to allow for the
solution of geometrically nonlinear structural analyses using MACH’s high-performance finite
element library, TACS. For coupled aerostructural analysis, we found that the load scale
ramping strategy proposed by Sanchez et al. [56] can reduce time to reach a coupled solution
by up to 40%.
Under identical loading, geometrically nonlinear structural analysis results in greater
bending stresses throughout the wing, particularly in the upper and lower skins where in-
creases of 10-15% were observed in the Von Mises and buckling failure criteria. We proposed
that this increase is due to the nonlinear analysis correctly maintaining the bending moment
caused by lift forces on the outboard portions of the wing which are rotated inboard, a conclu-
sion consistent with previous works demonstrating that linear analyses underpredicts bending
deformation when subject to follower forces. The presence of Brazier loads in the nonlinear
analyses causes a substantial increase in the compressive axial stress and buckling failure cri-
teria in the wingbox ribs. Correct modeling of finite-rotation kinematics in the geometrically
nonlinear analysis results in a significant inward and forward deflection of the wingtip, on
the order of 5% of the wing semispan. Any geometrically nonlinear drag-torsion effects, if
present, have little effect on the twisting deformation of the wing, which is dominated by the
geometric bend-twist coupling caused by wing sweep.
The overall lift and drag of an aircraft are insensitive to the levels of geometric nonlinearity

26
seen in the uCRM-13.5, varying by only 1-2% in a 2.5 g maneuver condition. The strongest
effect of the nonlinearity on aerodynamic behavior was seen in the pitching moment, due to
the span shortening and subsequent shift in the wing’s center of pressure under large bending
deformations. This resulted in large relative changes in the tail rotation required to trim the
aircraft but the absolute values of the changes were on the order of fractions of a degree.
The nonlinear analysis resulted in 5-10% lower bending stresses over the majority of the wing
skins, contrary to the results seen when comparing structural analyses under fixed loading.
Geometric nonlinearity did not appear to have a significant effect on the level of passive load
alleviation achieved by the wing.
In our structural optimizations, the increase in bending stresses seen in the structural
analysis studies lead to a 7% increase in the mass of the wingbox optimized using nonlin-
ear analysis. Brazier loads in the nonlinear analysis lead the optimizer to design ribs with
noticeably higher bending stiffness, but that the mass penalty for this increase was insignif-
icant. When optimizing using coupled aerostructural analysis, the optimizer was able to
aeroelastically tailor the wing and produced a fundamentally different design, with the lead-
ing edge spar being significantly reinforced to increase passive load alleviation. The mass
increase from linear to nonlinear aeroelastically optimized wingboxes was smaller than seen
in the structural optimizations but remains inconsistent with the overall decrease in bend-
ing stresses demonstrated when comparing geometrically linear and nonlinear aerostructural
analysis comparisons. We posited that this unexpected mass increase is caused by unrealis-
tic stress peaks around areas where inertial loads are applied from the engine, which has a
cascading effect on the wingbox sizing due to a restrictive set of adjacency constraints.
We also reported that introducing geometric nonlinearity results in an order of magnitude
cost increase for structural analysis and optimization. However, when incorporated in a high-
fidelity aerostructural analysis or optimization, the majority of the computational cost still
lies in the CFD portion of the process and the cost penalty for switching to a geometrically
nonlinear structural formulation can be as low as 15%.

Acknowledgments
A. C. Gray was supported in this work by a Justus & Louise van Effen Research Grant from
TU Delft and wishes to thank Roeland De Breuker for his advice and guidance during the
project. Computational resources for this work were provided by the Extreme Science and
Engineering Discovery Environment (XSEDE), which is supported by the National Science
Foundation grant number ACI-1548562, through allocation TG-DDM140001.

References
[1] Castellani, M., Cooper, J. E., and Lemmens, Y., “Nonlinear static aeroelasticity of high-
aspect-ratio-wing aircraft by finite element and multibody methods,” Journal of Aircraft,
Vol. 54, No. 2, 2017, pp. 548–560. doi:10.2514/1.C033825.

[2] Dowell, E., Edwards, J., and Strganac, T. W., “Nonlinear aeroelasticity,” Collection
of Technical Papers - AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics
and Materials Conference, Vol. 5, No. 5, 2003, pp. 3824–3847. doi:10.2514/2.6876.

[3] Patil, M. J., and Hodges, D. H., “On the importance of aerodynamic and
structural geometrical nonlinearities in aeroelastic behavior of high-aspect-ratio

27
 wings,” Journal of Fluids and Structures, Vol. 19, No. 7, 2004, pp. 905–915.
doi:10.1016/j.jfluidstructs.2004.04.012.

[4] Garcia, J. A., “Numerical investigation of nonlinear aeroelastic effects on flexible

high-aspect-ratio wings,” Journal of Aircraft, Vol. 42, No. 4, 2005, pp. 1025–1036.
doi:10.2514/1.6544.

[5] Su, W., “Coupled Nonlinear Aeroelasticity and Flight Dynamics of Fully Flexible Air-
craft,” Phd, University of Michigan, 2008. doi:10.1007/s13398-014-0173-7.2.

[6] Cavagna, L., Ricci, S., and Riccobene, L., “Structural sizing, aeroelastic analysis, and
optimization in aircraft conceptual design,” Journal of Aircraft, Vol. 48, No. 6, 2011, pp.
1840–1855. doi:10.2514/1.C031072.

[7] Harmin, M. Y., and Cooper, J. E., “Aeroelastic behaviour of a wing including geo-
metric nonlinearities,” Aeronautical Journal, Vol. 115, No. 1174, 2011, pp. 767–777.
doi:10.1017/S0001924000006515.

[8] Palacios, R., Wang, Y., Wynn, A., and Karpel, M., “Condensation of large finite-element
models for wing load analysis with geometrically-nonlinear effects,” IFASD 2013 - In-
ternational Forum on Aeroelasticity and Structural Dynamics, 2013, pp. 1–25.

[9] Bartels, R. E., Scott, R. C., Allen, T. J., and Sexton, B. W., “Aeroelastic analysis of
SUGAR truss-braced wing wind-tunnel model using FUN3D and a nonlinear structural
model,” 56th AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials
Conference, 2015. doi:10.2514/6.2015-1174.

[10] Howcroft, C., Cook, R., Calderon, D., Lambert, L., Castellani, M., Cooper, J. E., Lowen-
berg, M. H., and Neild, S., “Aeroelastic modelling of highly flexible wings,” 15th Dy-
namics Specialists Conference, AIAA, San Diego, 2016. doi:10.2514/6.2016-1798.

[11] Castellani, M., Cooper, J. E., and Lemmens, Y., “Nonlinear static aeroelasticity of high-
aspect-ratio-wing aircraft by finite element and multibody methods,” Journal of Aircraft,
Vol. 54, No. 2, 2017, pp. 548–560. doi:10.2514/1.C033825.

[12] Ritter, M., Teixeira, P. C., and Cesnik, C. E., “Comparison of nonlinear aeroelastic meth-
ods for maneuver simulation of very flexible aircraft,” AIAA/ASCE/AHS/ASC Struc-
tures, Structural Dynamics, and Materials Conference, 2018, 2018. doi:10.2514/6.2018-
1953.

[13] Riso, C., Di Vincenzo, F. G., Ritter, M., Cesnik, C. E., and Mastroddi, F., “Nonlin-
ear aeroelastic trim of very flexible aircraft described by detailed models,” Journal of
Aircraft, Vol. 55, No. 6, 2018, pp. 2338–2346. doi:10.2514/1.C034787.

[14] Lupp, C. A., and Cesnik, C. E. S., “A Gradient-Based Flutter Constraint Including Ge-
ometrically Nonlinear Deformations,” 2019 AIAA/ASCE/AHS/ASC Structures, Struc-
tural Dynamics, and Materials Conference, AIAA, San Diego, California, 2019.

[15] Medeiros, R. R., Cesnik, C. E. S., and Coetzee, E. B., “Computational Aeroelastic-
ity Using Modal-Based Structural Nonlinear Analysis,” AIAA Journal, 2019, pp. 1–10.
doi:10.2514/1.j058593.

28
[16] Kenway, G. K. W., Kennedy, G. J., and Martins, J. R. R. A., “Scalable Parallel Ap-
proach for High-Fidelity Steady-State Aeroelastic Analysis and Derivative Computa-
tions,” AIAA Journal, Vol. 52, No. 5, 2014, pp. 935–951. doi:10.2514/1.J052255.

[17] Kennedy, G. J., Kenway, G. K. W., and Martins, J. R. R. A., “High Aspect Ratio Wing
Design: Optimal Aerostructural Tradeoffs for the Next Generation of Materials,” Pro-
ceedings of the AIAA Science and Technology Forum and Exposition (SciTech), National
Harbor, MD, 2014. doi:10.2514/6.2014-0596.

[18] Kenway, G. W. K., and Martins, J. R. R. A., “High-fidelity aerostructural optimiza-

tion considering buffet onset,” Proceedings of the 16th AIAA/ISSMO Multidisciplinary
Analysis and Optimization Conference, Dallas, TX, 2015. AIAA 2015-2790.

[19] Burdette, D. A., and Martins, J. R. R. A., “Design of a Transonic Wing with an Adap-
tive Morphing Trailing Edge via Aerostructural Optimization,” Aerospace Science and
Technology, Vol. 81, 2018, pp. 192–203. doi:10.1016/j.ast.2018.08.004.

[20] Burdette, D. A., and Martins, J. R. R. A., “Impact of Morphing Trailing Edge on Mission
Performance for the Common Research Model,” Journal of Aircraft, Vol. 56, No. 1, 2019,
pp. 369–384. doi:10.2514/1.C034967.

[21] Brooks, T. R., Martins, J. R. R. A., and Kennedy, G. J., “High-fidelity Aerostruc-
tural Optimization of Tow-steered Composite Wings,” Journal of Fluids and Structures,
Vol. 88, 2019, pp. 122–147. doi:10.1016/j.jfluidstructs.2019.04.005.

[22] Brooks, T. R., Martins, J. R. R. A., and Kennedy, G. J., “Aerostructural Trade-offs for
Tow-steered Composite Wings,” Journal of Aircraft, 2020. doi:10.2514/1.C035699.

[23] Werter, N. P., and De Breuker, R., “A novel dynamic aeroelastic framework for aeroe-
lastic tailoring and structural optimisation,” Composite Structures, Vol. 158, 2016, pp.
369–386. doi:10.1016/j.compstruct.2016.09.044, URL http://dx.doi.org/10.1016/j.
compstruct.2016.09.044.

[24] Rajpal, D., Gillebaart, E., and De Breuker, R., “Preliminary aeroelastic design of com-
posite wings subjected to critical gust loads,” Aerospace Science and Technology, Vol. 85,
2019, pp. 96–112. doi:10.1016/j.ast.2018.11.051, URL https://doi.org/10.1016/j.
ast.2018.11.051.

[25] Stanford, B., “Aeroservoelastic Optimization of a Transport Aircraft Wingbox under

Stochastic Gust Constraints,” Journal of Aeroelasticity and Structural Dynamics, Vol. 6,
No. 1, 2018, pp. 21–41. doi:10.3293/asdj.2018.48.

[26] Rajpal, D., Kassapoglou, C., and De Breuker, R., “Aeroelastic optimization of com-
posite wings including fatigue loading requirements,” Composite Structures, Vol. 227,
No. August, 2019, p. 111248. doi:10.1016/j.compstruct.2019.111248, URL https:
//doi.org/10.1016/j.compstruct.2019.111248.

[27] Lancelot, P., and De Breuker, R., “Transonic Flight and Movable Load Modelling for
Wing-Box Preliminary Sizing,” 18th International Forum on Aeroelasticity and Struc-
tural Dynamics, , No. June, 2019, pp. 28–30.

29
[28] Krüger, W. R., Dillinger, J., De Breuker, R., and Haydn, K., “Investigations of pas-
sive wing technologies for load reduction,” CEAS Aeronautical Journal, Vol. 10, No. 4,
2019, pp. 977–993. doi:10.1007/s13272-019-00393-2, URL https://doi.org/10.1007/
s13272-019-00393-2.

[29] Natella, M., and De Breuker, R., “The effects of a full-aircraft aerodynamic model on
the design of a tailored composite wing,” CEAS Aeronautical Journal, Vol. 10, No. 4,
2019, pp. 995–1014. doi:10.1007/s13272-019-00366-5, URL http://dx.doi.org/10.
1007/s13272-019-00366-5.

[30] Bordogna, M. T., Lancelot, P., Bettebghor, D., and De Breuker, R., “Aeroelastic tailor-
ing for static and dynamic loads with blending constraints,” 17th International Forum on
Aeroelasticity and Structural Dynamics, IFASD 2017, Vol. 2017-June, No. June, 2017.

[31] Brazier, L. G., “On the Flexure of Thin Cylindrical Shells and Other “Thin” Sections,”
Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences,
Vol. 116, No. 773, 1927, pp. 104–114. doi:10.1098/rspa.1927.0125.

[32] Stanford, B. K., and Dunning, P. D., “Optimal Topology of Aircraft Rib and Spar
Structures Under Aeroelastic Loads,” Journal of Aircraft, Vol. 52, No. 4, 2015, pp. 1298–
1311. doi:10.2514/1.C032913, URL http://arc.aiaa.org/doi/10.2514/1.C032913.

[33] Smith, M. J., Patil, M. J., and Hodges, D. H., “CFD-based analysis of nonlinear aeroelas-
tic behavior of high-aspect ratio wings,” 19th AIAA Applied Aerodynamics Conference,
, No. c, 2001. doi:10.2514/6.2001-1582.

[34] Verri, A. A., Bussamra, F. L., de Morais, K. C., Becker, G. G., Cesnik, C. E., Luque
Filho, G. B., and de Oliveira, L. C., “Static loads evaluation in a flexible aircraft us-
ing high-fidelity fluid–structure iteration tool (E2-FSI): extended version,” Journal of
the Brazilian Society of Mechanical Sciences and Engineering, Vol. 42, No. 1, 2020.
doi:10.1007/s40430-019-2154-4.

[35] Calderon, D. E., Cooper, J. E., Lowenberg, M. H., and Neild, S. A., “On the Effect
of Including Geometric Nonlinearity in the Sizing of a Wing,” AIAA/ASCE/AHS/ASC
Structures, Structural Dynamics, and Materials Conference, 2018, Kissimmee, 2018.
doi:10.2514/6.2018-1680.

[36] Calderon, D. E., Cooper, J. E., Lowenberg, M., Neild, S. A., and Coetzee, E. B., “Sizing
High-Aspect-Ratio Wings with a Geometrically Nonlinear Beam Model,” Journal of
Aircraft, Vol. 56, No. 4, 2019, pp. 1455–1470. doi:10.2514/1.c035296.

[37] Mader, C. A., Kenway, G. K. W., Yildirim, A., and Martins, J. R. R. A., “ADflow—An
open-source computational fluid dynamics solver for aerodynamic and multidisciplinary
optimization,” Journal of Aerospace Information Systems, 2020. doi:10.2514/1.I010796.

[38] Yildirim, A., Kenway, G. K. W., Mader, C. A., and Martins, J. R. R. A., “A Jacobian-
free approximate Newton–Krylov startup strategy for RANS simulations,” Journal of
Computational Physics, Vol. 397, 2019, p. 108741. doi:10.1016/j.jcp.2019.06.018.

[39] Kenway, G. K. W., Mader, C. A., He, P., and Martins, J. R. R. A., “Effective Adjoint
Approaches for Computational Fluid Dynamics,” Progress in Aerospace Sciences, Vol.
110, 2019, p. 100542. doi:10.1016/j.paerosci.2019.05.002.

30
[40] Kennedy, G. J., and Martins, J. R. R. A., “A Parallel Finite-Element Frame-
work for Large-Scale Gradient-Based Design Optimization of High-Performance
Structures,” Finite Elements in Analysis and Design, Vol. 87, 2014, pp. 56–73.
doi:10.1016/j.finel.2014.04.011.

[41] Kennedy, G. J., “Aerostructural analysis and design optimization of composite aircraft,”
Ph.D. thesis, University of Toronto, Toronto, ON, Canada, December 2012.

[42] Brown, S., “Displacement extrapolations for CFD+CSM aeroelastic analysis,” 38th
Structures, Structural Dynamics, and Materials Conference, American Institute of Aero-
nautics and Astronautics, 1997. doi:10.2514/6.1997-1090.

[43] Kennedy, G. J., and Martins, J. R. R. A., “A parallel aerostructural optimization frame-
work for aircraft design studies,” Structural and Multidisciplinary Optimization, Vol. 50,
No. 6, 2014, pp. 1079–1101. doi:10.1007/s00158-014-1108-9.

[44] Secco, N., He, G. K. W. K. P., Mader, C. A., and Martins, J. R. R. A., “Efficient Mesh
Generation and Deformation for Aerodynamic Shape Optimization,” AIAA Journal,
2020. (In press).

[45] Luke, E., Collins, E., and Blades, E., “A Fast Mesh Deformation Method Using Explicit
Interpolation,” Journal of Computational Physics, Vol. 231, No. 2, 2012, pp. 586–601.
doi:10.1016/j.jcp.2011.09.021.

[46] Kenway, G. K., Kennedy, G. J., and Martins, J. R. R. A., “A CAD-Free

Approach to High-Fidelity Aerostructural Optimization,” Proceedings of the 13th
AIAA/ISSMO Multidisciplinary Analysis Optimization Conference, Fort Worth, TX,
2010. doi:10.2514/6.2010-9231.

[47] Sederberg, T. W., and Parry, S. R., “Free-form Deformation of Solid Geomet-
ric Models,” SIGGRAPH Comput. Graph., Vol. 20, No. 4, 1986, pp. 151–160.
doi:10.1145/15886.15903.

[48] Gill, P. E., Murray, W., and Saunders, M. A., “SNOPT: An SQP Algorithm for Large-
Scale Constrained Optimization,” SIAM Review, Vol. 47, No. 1, 2005, pp. 99–131.
doi:10.1137/S0036144504446096.

[49] Lyu, Z., Xu, Z., and Martins, J. R. R. A., “Benchmarking Optimization Algorithms for
Wing Aerodynamic Design Optimization,” Proceedings of the 8th International Confer-
ence on Computational Fluid Dynamics, Chengdu, Sichuan, China, 2014. ICCFD8-2014-
0203.

[50] Wu, N., Kenway, G., Mader, C. A., Jasa, J., and Martins, J. R. R. A., “pyOptSparse: A
Python framework for large-scale constrained nonlinear optimization of sparse systems,”
Journal of Open Source Software, Vol. 5, No. 54, 2020, p. 2564. doi:10.21105/joss.02564.

[51] Leon, S. E., Paulino, G. H., Pereira, A., Menezes, I. F., and Lages, E. N., “A unified
library of nonlinear solution schemes,” Applied Mechanics Reviews, Vol. 64, No. 4, 2011.
doi:10.1115/1.4006992.

31
[52] Beluni, P. X., and Chulya, A., “An improved automatic incremental algorithm for
the efficient solution of nonlinear finite element equations,” Computers and Struc-
tures, Vol. 26, No. 1-2, 1987, pp. 99–110. doi:10.1016/0045-7949(87)90240-9, URL
https://www.sciencedirect.com/science/article/pii/0045794987902409.

[53] Matthies, H., and Strang, G., “The solution of nonlinear finite element equations,”
International Journal for Numerical Methods in Engineering, Vol. 14, No. 11, 1979, pp.
1613–1626. doi:10.1002/nme.1620141104, URL http://doi.wiley.com/10.1002/nme.
1620141104.

[54] Bergan, P. G., “Solution algorithms for nonlinear structural problems,” Computers and
Structures, Vol. 12, No. 4, 1980, pp. 497–509. doi:10.1016/0045-7949(80)90125-X.

[55] Bathe, K. J., and Dvorkin, E. N., “On the Automatic Solution of Nonlinear Finite
Element Equations,” Computers and Structures, Vol. 17, No. 3, 1983, pp. 871–879.
doi:10.1016/0045-7949(83)90101-3.

[56] Sanchez, R., Albring, T., Palacios, R., Gauger, N. R., Economon, T. D., and Alonso,
J. J., “Coupled adjoint-based sensitivities in large-displacement fluid-structure interac-
tion using algorithmic differentiation,” International Journal for Numerical Methods in
Engineering, Vol. 113, No. 7, 2018, pp. 1081–1107. doi:10.1002/nme.5700.

[57] Brooks, T. R., Kenway, G. K. W., and Martins, J. R. R. A., “Benchmark Aerostructural
Models for the Study of Transonic Aircraft Wings,” AIAA Journal, Vol. 56, No. 7, 2018,
pp. 2840–2855. doi:10.2514/1.J056603.

[58] Stroud, W. J., and Agranoff, N., “Minimum-mass design of filamentary composite panels
under combined loads: Design procedure based on simplified buckling equations,” Tech.
Rep. TN-D-8257, NASA Langley Research Center, Hampton, VA, 1976.

32