31st AIAA Applied Aerodynamics Conference AIAA 2013-2649

June 24-27, 2013, San Diego, CA

An Adjoint-Based Aerodynamic Shape Optimization
Methodology for Fairing Systems

M. R. Colonno1, F. Palacios1, T. Economon2, A. K. Lonkar2, and J. J. Alonso3
Stanford University, Dept. of Aeronautics & Astronautics, Stanford, CA, 94305

The optimization of a fairing system is formulated in terms of the overall performance of
the launch vehicle-fairing system as measured by the total velocity increment within a
maximum load constraint at the fairing-vehicle interface. This objective requires a time-
accurate treatment of the vehicle’s aerodynamics and, when adjoint sensitivities are used for
gradient information, a time-accurate adjoint functional as well. The fairing’s geometry is
Downloaded by STANFORD UNIVERSITY on July 28, 2013 | http://arc.aiaa.org | DOI: 10.2514/6.2013-2649

parameterized in terms of a local spherical coordinate system which can be perturbed via
the manipulation of spline control points or Hicks-Henne bump functions. The local surface
perturbation and the continuous adjoint capabilities of the SU2 design suite allow for
accurate sensitivities of the aerodynamic performance to the geometric parameters for use
with gradient-based optimization algorithms. A sample medium-lift launch vehicle is used as
an example to demonstrate the method and results of the optimization presented in addition
to possibilities to improve the method.

A = area
A = matrix of linear constraints
A = inviscid flux Jacobian matrices
b = vector of linear constraint values
c = vector of nonlinear constraint
CA = aerodynamic axial force coefficient
CM = aerodynamic moment coefficient
CN = aerodynamic normal force coefficient
D = drag force
d = composite core thickness or beam cross-section measure
d = force projection vector
E = total energy per unit mass
Fale = convective fluxes in arbitrary Langrangian-Eulerian form
FA = axial force
FN = normal force
g0 = standard Earth gravity
H = stagnation enthalpy
Isp = specific impulse
i,j = indices
J = objective function
jS = scalar function defined on surface S
L = length
m = mass
n = unit normal vector
p = static pressure
q = dynamic pressure

Engineering Research Associate, Department of Aeronautics & Astronautics, Stanford University, AIAA Senior Member.
Ph.D. Candidate, Department of Aeronautics & Astronautics, Stanford University.
Associate Professor, Department of Aeronautics & Astronautics, Stanford University, AIAA Senior Member.

American Institute of Aeronautics and Astronautics

Copyright © 2013 by Michael Colonno. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission.

R = radius
R(U) = system of governing flow equations
S = solid wall boundary (design surface) or reference area
t0 = initial time
tf = final time
U = vector of conservative variables
uΩ = control volume boundary velocity
V∞ = free stream velocity
v = flow velocity vector in inertial frame
x = axial direction
x = vector of design variables
α = angle of attack
β = side-slip angle
ρ = density
Downloaded by STANFORD UNIVERSITY on July 28, 2013 | http://arc.aiaa.org | DOI: 10.2514/6.2013-2649

ρ∞ = free-stream density
ϕ = flow adjoint velocity vector
Γ = domain boundary
ψ = vector of flow adjoint variables
Ω = problem domain
σ = stress
γ = ratio of specific heats
∇(⋅) = gradient operator
∇⋅ (⋅) = divergence operator
∂n(⋅) = surface-normal gradient operator, nS ⋅ ∇(⋅)
⋅ = vector inner product
x = vector cross product
⊗ = vector outer product
δ(⋅) = first variation


A = axial direction
bt = boat tail section property
cyl = cylindrical section property
E = Earth
f = fairing property
fd = at fairing deployment
M = moment
N = normal direction (to axis)
P = propellant
PL = payload
S = structural
v = vehicle
∞ = ambient or free stream
0 = initial


1,2,…,n = launch vehicle stage
* = optimal

American Institute of Aeronautics and Astronautics

Copyright © 2013 by Michael Colonno. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission.

11. and structural loading. MDO efforts of rocket systems to date have focused primarily on system-level parameters3. The loss terms include the velocity increment lost to aerodynamic drag. Eq. etc. Some previous work has been done in the area of high-fidelity rocket and fairing optimization7 in addition to some recent work surrounding the aerodynamic shape optimization of asymmetric fairings for large and unconventional payload geometries1. in the form shown below. mass properties. Ultimately. and 3) the very broad range of flow regimes encountered. focusing primarily on the minimization of aerodynamic drag and.21 have focused on high-fidelity implementation. 3) the number of stages (n). This essentially decoupled the launch vehicle and fairing design by reducing the size of the design space Downloaded by STANFORD UNIVERSITY on July 28. Evaluating this expression with values of practical systems shows that even for mf → 0 the total performance gain expressed in terms of ΔVtot is small. In order to fully capture the performance impact of the fairing design.4. aerodynamic heating. In order to address these challenges in a consistent way for shape optimization. This famous equation elegantly reveals the system-level parameters key to the performance of any rocket vehicle: 1) propulsion efficiency (Isp. Eq. a multidisciplinary optimization (MDO) is required which captures the interactions of structural loading. is a small fraction of the final stage’s initial mass and a very small fraction of the vehicle’s initial mass. in some cases. 2013 | http://arc.org | DOI: 10. measured in seconds in the formulation below). II.2514/6. The subsequent section discusses the parameterization of the system’s geometry and the associated linear constraints. The cumulative aerodynamic performance of fairing system over the relevant portion of a trajectory to atmospheric flight loads was determined. Introduction O PTIMUM profiles for launch vehicle and missile fairings have been sought since the early days of rocket flight. and 4) the trajectory’s losses (ΔVlosses) or the deviation from the ideal performance. c( x ) ≤ 0 (1) Subject to: . Ax ≤ b where c(x) denotes any nonlinear constraints and A denotes a matrix of linear constraint coefficients with corresponding values b. staging ratio) rather than directly optimizing the physical variables of the vehicle (e. Equation 2..g. with the final stage performance separated from the summation. I. skin thickness. (3). in essentially all practical systems. In order to isolate the aerodynamic shape optimization problem from the full MDO one without omitting important performance trades. constrained by one-dimensional-equivalent structural limits of the vehicle. separating it into two sub-stages at a specific propellant mass (subscript fd). implicitly assumes that the fairing is deployed sometime during the final stage flight. aerodynamic stability. . Inc. In the sections below two different objective functions are developed and the nonlinear constraint discussed. 2) the fairing’s aerodynamic properties have a considerable impact on the loading sustained by the rest of the vehicle which is not directly modeled. continuous surface adjoint formulation was used via the SU2 suite of aerodynamic design tools20. The general optimization problem statement is Minimize: J ( x) . three key challenges exist that are not typically encountered in most aerodynamic shape optimizations: 1) the ambient conditions are changing rapidly during either ascent or reentry requiring a time-accurate treatment of performance and sensitivities.5 (e. with permission. the performance measure of interest is the flight performance of the vehicle as a whole which includes consideration of disciplines beyond aerodynamics. Later sections discuss the sensitivities of J and c based on a continuous adjoint formulation and detail the optimization procedure itself.12. Published by the American Institute of Aeronautics and Astronautics.aiaa. however. and the velocity increment lost to finite-time flight in a gravitational field (ΔVg). Optimization Problem Formulation In this section the overall performance of the launch vehicle and structural constraints are developed.19. radius. For any chemically-powered launch vehicle system. A.) while MDO efforts on other aerospace applications6. and trajectory performance must be included in an integrated fashion. must be considered in the context of the launch vehicle performance as a whole. constraints were used to bound the aerodynamic loading applied to the rest of the vehicle. a time-accurate.2013-2649 available to a full-vehicle MDO. length. This. the payload mass fraction (fraction of the liftoff mass that corresponds 3 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno. The fairing mass. Performance Consider the general multi-stage rocket equation. Launch vehicle optimization is inherently multidisciplinary in nature and some simplifications are made to expose the most important effects of the fairing’s aerodynamic performance. (2).g.22. Within the aerodynamic shape optimization sub-problem of a fairing. 2) the mass ratios (full mass to empty) of the stages.

equivalently. the changing ambient conditions. the vehicle’s structural dynamics. and its control system – all of which are beyond the scope of this work. the aerodynamics of a launch vehicle at a given instant in an ascent trajectory are a function of the current altitude (atmospheric conditions) and velocity in addition to the vehicle’s orientation. The latter requires a six-degree-of freedom dynamic analysis involving the instantaneous wind conditions. Inc. (4) GmE / RE The negative sign corresponds to minimization consistent with the canonical optimization problem formulation. Finally. Hence. The constant terms merely shift the objective value and do not play a role in optimization. to the payload) is very small. Here we assume that the contribution to the drag of the rest of the vehicle. maximum q∞(t)α(t) is fixed if a value of α at maximum q∞ is assumed. This assumption represents a significant simplification but removed the need to recompute the trajectory for each design – a major corresponding reduction in computational cost.aiaa. CA and CM are evaluated at α = αmax only and the nominal (α = 0) value of CA for drag loss is given by 4 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno. neglects winds and dynamics or. This requires a method by which sensitivities can be obtained to functionals varying in time and the continuous adjoint approach used presently is summarized in a subsequent section. the lateral and axial force coefficients are assumed to be proportional to sin(α) and cos(α). is a constant which is independent of the fairing’s geometry. small savings in either mf or ΔVlosses from an optimized fairing are critical due to the high sensitivity of the relevant performance metric.org | DOI: 10. ΔVD can be expressed as the sum of the fairing’s drag and that of the rest of the vehicle. not to mention the cost of access to space. or product of dynamic pressure and angle of attack. integrated over the trajectory. (3) is evaluated directly by SU2. In order to isolate the fairing’s effects and limit the scope of the problem the vehicle’s trajectory was assumed to fixed. from a performance standpoint. This adds an additional scalar input to the optimization which defines the maximum aerodynamic loading but does not change the optimization procedure itself. The sensitivity of the ΔVD functional of Eq. The nominal trajectory.. Hence. The total velocity increment for a typical low-Earth-orbit mission is ~ 9 km/s. assumes α = 0 throughout. respectively. (2) rely upon the mass of the fairing and can hence be expressed as the sensitivity of mf to the design variables. ΔVtot J ≡− . B. and the changing vehicle mass. It must also be noted that the vehicle’s drag is only partially accounted for by the fairing. discussed in detail in a subsequent section. . Consistent with the assumption that q∞(t) is fixed. sensitivities of the objective function are required. The remaining terms of Eq. Aerodynamics Most generally. Hence. with permission. In order to save the computational expense of evaluating the aerodynamics at both α = 0 (for nominal performance) and α = αmax (for loads). Published by the American Institute of Aeronautics and Astronautics. This is discussed in the mass properties section below. the objective used here divides the total velocity increment by an orbital velocity scale for Earth. its overall impact is minimal and was neglected. While aerodynamic loading can have an impact on the trajectory close to the maximum q∞ point. (2): an integration of the drag force. essentially linearizing the aerodynamics and structural dynamics of the vehicle for small α.2514/6. Hence. When establishing an objective function for numerical evaluation it is preferable to scale both the objective and variables to ~O(1). 2013 | http://arc. A simplification frequently employed to decouple this tightly-interdisciplinary problem (and provide weather limits for launch) is to rate the vehicle to a maximum q∞α. for small α.2013-2649 conditions but the functional formed by Eq. n −1 ⎛ mi ⎞ ⎡ ⎛ mSn + mPn + mPL + m f ⎞ ⎛ mSn + (mPn ) fd + mPL ⎞ ⎤ ΔVtot = ∑ I spi g 0 ln ⎜ i 0 i ⎟ + I spn g 0 ⎢ ln ⎜ n ⎟⎟ + ln ⎜⎜ ⎟⎟ ⎥ − ΔVlosses ⎝ m0 − mP ⎠ ⎢⎣ ⎜⎝ mS + (mP ) fd + mPL + m f mSn + mPL n i =1 ⎠ ⎝ ⎠ ⎥⎦ (2) n m0i = ∑ ( mSj + mPj ) + mPL + m f j =i tf q∞ (t )CD (t ) ΔVlosses = ΔVg + ΔVD = ΔVg + S ∫ dt (3) t0 mv (t ) In this study we seek to minimize not the drag force itself nor even the drag force at a finite set of discrete Downloaded by STANFORD UNIVERSITY on July 28.

These include both the contribution of aerodynamic forces and the inertial forces for a given load case defined by an axial acceleration.α = 0) dt . (2) requiring consideration is the mass of the fairing. 2013 | http://arc.φ) its mass is given by 2π π / 2 mf = ∫ ∫σ 0 0 s (θ . Published by the American Institute of Aeronautics and Astronautics. In summary. Av ≈ 2π Rv d v . D. acting as a nonlinear constraint. With this distribution.org | DOI: 10. these were combined to form an equivalent static line load at the vehicle-fairing interface.2013-2649 was considered (shear is neglected). a reference stress for the vehicle primary structure material(s) (σref) is required for comparison to the line loads. If the faring’s surface is specified as r(θ. Employing a one-dimensional structural model for the vehicle. The total load applied to the vehicle was reduced to an equivalent axial force and bending moment at the interface. circular cross-section was assumed for the vehicle and the second moment of area and cross-sectional area estimated via the thin-walled formulation. A representative structural cross-sectional area (Av) and second moment of area (Iv) for the vehicle primary structure are required parameters. C A (α ) = C A (0)cos(α ) = CD cos(α ) . a characteristic wall thickness (dv) replaces the parameter Iv. .φ ) 2 sin(θ ) dφ dθ . Returning to the context of the optimization problem.2514/6. S = π Rv2 .aiaa. I v ≈ π Rv3 d v . Here. (9) σ ref ⎝ Av I v ⎠ σ ref d v ⎝ 2 ⎠ 2π Rv d vσ ref Lateral inertial loads. have been neglected. With this substitution. and angle of attack (α). (7) FA M CA ≡ . dynamic pressure (q∞). the surface density distribution (mass per unit area) of the fairing’s skin is an input parameter. In similar fashion to the drag losses above. which are primarily a function of vehicle structural dynamics. Cload is taken to be the loads applied to the vehicle from the fairing’s aerodynamic loading only and not a representation of the total static or dynamic loading on the vehicle (which is considerably more complex). (9) and simplified through judicious choices of the normalizing area (S) and length (L) for aerodynamic coefficients. Here. (6) t0 mv (t ) f C. Inc. φ )rf (θ . CM ≡ . averaged over the surface area. was designed independent of the launch vehicle on which it would fly. (7). defined in Eq. (10) 5 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno. the fairing’s mass can be evaluated by integrating over the surface area in a spherical coordinate system. (8) q∞ S q∞ SRv 1 ⎛ FA + m f ax MRv ⎞ q∞ Rv ⎛ 1 ⎞ m f ax Cload f ≡ ⎜ + ⎟= ⎜ C A + CM ⎟ + . Structural Load Constraint Since the fairing. Without a detailed structural design the internal structure of the must remain constant with respect to changes in the design variables (though it can be an a priori function of position) and the mass of the fairing is simply a function of the fairing’s surface area – a major simplification. This density can include both structural and non-structural mass. an additional scalar parameter. the drag loss due to the fairing is given by tf q∞ (t ) ΔVD f = S ∫ C A (t . Assuming a thin-walled primary structure. ax. only the axial stress state Downloaded by STANFORD UNIVERSITY on July 28. a thin-walled. (7). Eqs.. This essentially requires the fairing’s internal structure to be overdesigned and neglects any additional gains in performance possible through an aero-structural optimization. The maximum of these total interface forces must be constrained to ensure the vehicle is not overloaded. Mass Properties The final aspect of Eq. Eqs. with permission. limiting loads at the vehicle-fairing interface must be established a priori. (5) Finally. for the purposes of this problem. Cload must be ≤ its maximum allowable value for the vehicle. Finally.

1. The payload volume and dynamic envelope (RPL and Lcyl) were assumed fixed by satellite needs resulting in a volume constraint on the fairing’s shape. ri} (tensor product spline surface with control points {θi. r(θ. two approaches were considered: 1) A cubic spline interpolant of m control points {θi. This fairing includes a flair-down or “boat tail” region aft of the payload compartment allowing for payload Downloaded by STANFORD UNIVERSITY on July 28. R RPL Payload Volume Rv Lbt Lcyl Lfor Figure 1. III. 2 and subsequent section for detailed discussion of geometry). Any parameterization that can produce the profile of the fairing. (11). though the optimization method applies to any general geometry of the form r(θ. While analytical sensitivities are possible for sufficiently simple r(θ.2013-2649 diameters which are greater than the vehicle’s diameter. 2) A set of Hicks-Henne bump functions at m control points {θi. Geometry & Grid Deformation 17. 2013 | http://arc. π /2 m f = 2π ∫σ 0 s (θ )rf (θ ) 2 sin(θ )dθ . only the forward region is parameterized and considered for optimization. The curved portion of the fairing is typically a spline or otherwise-parameterized curve tangent at either end to the cylindrical section of the fairing and a normal to the axis of rotation at the nose. ri} for general surfaces) and dr /dθ = 0 boundaries to provide the proper tangency. (10) or Eq.φ). a complex step method was used here for general application to any parameterization. yielding the sensitivities ∂mf/∂xi. Published by the American Institute of Aeronautics and Astronautics. An axisymmetric geometry is shown in Fig.φ) the sensitivities of mf to the design parameters can be evaluated directly from Eq. For an axisymmetric fairing only the distribution in θ is required.aiaa.φ).org | DOI: 10. . Cross-sectional view of nominal fairing design. Inc. with permission. (11) If the geometry is parameterized in r(θ. Due to the payload constraints mentioned above.φ) or r(θ) for axisymmetric geometries. φi. Δri} which are added to the baseline surface. can be accommodated within the current framework. The general parameterization of the geometry is depicted in Fig 2.. 6 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno. 1 and is used as the sample case in this study.24 A nominal EELV-sized payload fairing (of similar shape and scale that of a ULA Delta II23) is shown in Fig. Dotted line shows payload volume envelope. Here.2514/6. where θ is the polar angle from the nose and φ is the azimuth (see Fig.

With the outer surface of the fairing parameterized. (15) When using spline interpolation.25.2514/6. Inc. we seek only to eliminate invalid geometries. the minimum r must be found in each segment of the spline. Defining the Hicks-Henne independent variable. u ∈ [0→1]. Hence. θ (14) u (θ ) = + 0.5. 2013 | http://arc. This new set of partially-modified Hicks-Henne functions provides a means to smoothly deform the baseline surface to an arbitrary level of detail. Even if all the control points are taken to have rj > 0 the interpolated values between them (especially for higher-order polynomials basis functions and a small number of control points) can still become negative.ri} Downloaded by STANFORD UNIVERSITY on July 28. a modification is required at the nose. Allowing geometries with inherently poor performance may impede converge of the optimization but will not result in errors. (12) π the general Hicks-Henne function in this application takes the form pi fi (r ) = Δri ⎡⎣sin(π u ni ) ⎤⎦ . with permission. First we seek to eliminate any surface that crosses the axis of rotation: min[r (θ . . log10 (ui ) The amplitude (Δri). enforcement of this constraint requires some detail.ri} φ r θ {θi.org | DOI: 10.2013-2649 Figure 2.. resulting in a more complex statement of the constraint. by construction.0): unose = 0. In order to allow movement of the nose point when using Hicks-Henne functions14. 7 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno. Hicks-Henne functions.aiaa. To circumvent this at the nose the definition is modified slightly to capture only half of the function (0.φi. Published by the American Institute of Aeronautics and Astronautics. π With the modification above the nose can be smoothly stretched independently.5 < u < 1. go to zero at u = 0 and u = 1 and hence no function could be placed at the nose (u = 0). General parameterization of the fairing nose geometry and control points. {θi. peak location (ui). log10 (0.φ )] > 0 . as θ u (θ ) ≡ 2 . care must be taken in the form of constraints to ensure valid geometries. and width (pi) are all design parameters.5) (13) ni = .5. Here.

2013 | http://arc. (18) ⎝ ∂r (θ .φ .φ ) ⎞ π tan −1 ⎜ ⎟≤ . The parrameterizations discussed abo ove only define points on thee fairing surfacce. (8). This comppletes the parameteriization of the geometry g and itts correspondin ng constraints.2514/6. The new w surface cooordinates act ass a displacemen nt boundary co ondition and thhe equilibrium position of thee volume nodess is solved for. Inc. m-1. 8 Americcan Institute off Aeronautics aand Astronautiics Copyright © 2013 by Michael Colonno. rj . functions must ⎡ m ⎤ min ⎢ r0 (θ . Heence. Here. Notee that the integraand of the drag vvelocity incremen nt is biased towaard higher Mach number range oof flight due to thhe decreasing m mass. froom a practical matter.. Figure 3. with permission. in optimization o ap pplications where a potentiallly-large numbeer of designs w will be evaluateed in an automaated fashion.φ ) / ∂θ ⎠ 2 Downloaded by STANFORD UNIVERSITY on July 28. Collectively. min[r (θ .2013-2649 Surfaces with w reward-faccing portions arre permitted th hough will likelly perform pooorly. we prevent an ny profiles which form an an ngle with the axxis of rotation ggreater than noormal. ….org | DOI: 10. In other words. In order to m modify the aerodynam mic model (detaails below in a subsequent secction) the voluume mesh mustt be deformed iin a way that ddoes not producce cells which are a inverted or of very poor quality. (16) Hicks-Hennne functions arre similar: the new profile formed by the suum of the baselline and all thee Hicks-Henne m have positive radius eveerywhere. (17) ⎣ i =1 ⎦ Finally y. Seelected data from m nominal ascentt trajectory. tthese constrainnts will be referrred to “geometricc constraints” throughout t the remainder of this t work. cells whiich cause either divergence in n the flow solvver or locally innaccurate resullts can disrupt aan optimizatioon in ways thatt are not alwayys easy to detecct. rj +1 )] > 0 for j = 1. similar to that t of Ref.aiaa. q This is of particular im mportance.φ ) + ∑ fi (θ . . the fairing surfacee should be norrmal to the axis of rotation att the nose only and have an anngle of incidennce everywhere else between n normal and zeero: ⎛ r (θ .φ ) ⎥ > 0 . we us e a spring anallogy deformatiion algorithm. which treats the t edges of ceells as linear “ssprings” whichh resist deformaation. Published by the American Institute of Aeronautics and Astronautics.

with permission. Nominal Trajectoory A nomiinal trajectory to 185 km circcular orbit was computed corrresponding to tthe baseline veehicle (summarrized below) and d fairing yieldin ng q∞(t). tthese equationss are considereed in a domain. IV.org | DOI: 10. S SU2 features aan unstructuredd flow djoint formulatiion for surface sensitivities too aerodynamicc properties. Norm mal vectors too the boundarry surfaces are directed outt of the domaain by convention n. m(t). Thhis section contains a solver and continuous adj summary ofo the governing flow equaations and corrresponding tim me-accurate continuous adjooint formulatioon for aerodynammic analysis and d design. The atmospheeric portion of tthe trajectory uused in this stuudy is plotted abbove in Fig.aiaa. Ω. including a tiime-accurate aadjoint formulation n allowing forr the evaluatio on of the drag loss functionaal in Eq. We aree interested in time-accurate fluid behaviorr around aeroddynamic bodiess in arbitrary m motion for situuations where visccous effects cann be consideredd negligible. Γ∞. Downloaded by STANFORD UNIVERSITY on July 28. Published by the American Institute of Aeronautics and Astronautics. where G G ⎧ρ ⎫ ⎧ ρ (v − uΩ ) ⎫ G ⎪ G⎪ G ⎪⎪ G G G ⎪⎪ U = ⎨ ρ v ⎬ .22 was uttilized for flow and adjointt solutions15. 2013 | http://arc. and the transient t free sttream conditioons required forr CFD analysiss.2013-2649 A. In our particuular problem. (20) ⎪ρ E ⎪ ⎪ ρ E (vG − uG ) + pvG ⎪ ⎩ ⎭ ⎪⎩ Ω ⎪⎭ 9 Americcan Institute off Aeronautics aand Astronautiics Copyright © 2013 by Michael Colonno. Figure 4. The T governing fflow equationss in the limit of vanishing visscosity are the com mpressible Euller equations. . M(t). Cubic C spline intterpolation was used between n the trajectoryy data point whhere necessary to generate freee stream con nditions. 3. Inc. The su urface S will also a be referreed to as the ddesign surface and it is conssidered continuoussly differentiabble (C1). and these conservation equaations can be eexpressed in aarbitrary Lagraangian- Eulerian diifferential form m as G ∂U G + ∇ ⋅ Falle = 0 in Ω ∂t G G G ( v − uΩ ) ⋅ nS = 0 on S (19) G WΓ+∞ (U ) = W∞ on Γ ∞.. 4.2514/6. Descrip ption of the Phhysical Probleem Ideal flluids are goverrned by the Eu uler equations. as seen in Fig. Governing Flow and Adjointt Problems foor Unsteady Aerodynamiics The SUU2 suite of to ools20. Ω bounded by a disconnected d boundary wh hich is dividedd into a far-fielld component. V. Falee = ⎨ ρ v ⊗ (v − u Ω ) + Ip ⎬ . Schematic of volume and surfface domains forr the continuouss adjoint formulaation with surfacce shape perturbaation. S. and a soliid wall boundary. (2). A Standard Atmosphere A mo odel was used.

x . uΩ is the local velocity for a domain in motion. (24) where S has been deformed to a new surface S’ by applying an infinitesimal profile deformation. Surface shape deformations will result in variations of the pressure distribution along the surface.2013-2649 B. E is the G total energy per unit mass. . in the local G G normal direction. 2013 | http://arc. w}T is the flow velocity.sin β ) / C∞ . In order to close the system of equations after assuming a perfect gas. the fluid states at the boundaries are updated depending on the sign of the eigenvalues. Surface Sensitivities via a Time-Accurate Continuous Adjoint Approach The objective of this section is to describe the way in which we quantify the influence of geometric modifications on the pressure distribution at a solid surface in the flow domain. The boundary conditions take into account any boundary velocity due to control volume motion. This leads directly to a gradient-based optimization framework. The shape deformations applied to S will be infinitesimal in nature and can be described mathematically by G G G G S ' = { x + δ S ( x )nS ( x ).) The second line of Eqs.. at the surface to a desired quantity of interest. δS. at a point. example candidates are G CD: d = ( cos α cos β .cos β ) / C∞ .cos α − L sin α cos β ⎟ . the goal is to compute the variation. we will concentrate on functionals defined as time-varying quantities on the solid surface S. J. p. − sin β sin α .2514/6.cos α . constant vector which can be chosen to relate the pressure. S. (19) represents the flow tangency condition at a solid wall. of Eq. S.cos α . (26) G 1 ⎛ CL C C ⎞ CL/CD: d= ⎜ − sin α − cos α cos β . (21) ⎣ 2 ⎦ and the stagnation enthalpy is given by p H =E+ . Therefore. nS . The final line represents a characteristic-based boundary condition at the far-field where. with respect to changes in the shape of the boundary. (22) ρ Downloaded by STANFORD UNIVERSITY on July 28. − L sin β . 4. and p is the static pressure.org | DOI: 10. G G ρ is the fluid density. For aerodynamic applications.aiaa. x ∈ S } . G CL: d = ( − sin α . G CSF: d = ( − sin β . A typical shape optimization problem seeks the minimization of a certain cost function. Published by the American Institute of Aeronautics and Astronautics. Inc. so we will focus on pressure-based functionals with the form G G jS = d ⋅ ( pnS ). (In the present application. to S S (23) where jS is a time-dependent scalar function defined at each point on S. on the surface.sin α cos β . the pressure is determined from ⎡ 1 G G ⎤ p = (γ − 1) ρ ⎢ E − (v ⋅ v ) ⎥ . (23) caused by arbitrary but small (and multiple) deformations of S and to use this information to drive our geometric changes in order to find an optimal shape for the design surface. and it is an arbitrary. tf J= ∫ ∫ j dsdt . as shown above in Fig. v. or change. v = {u. uΩ is zero. C∞CD ⎝ CD CD CD ⎠ 10 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno. with permission. (25) G The vector d is the force projection vector. Therefore. in general.0 ) / C∞ .

Furthermore. the constrained optimization problem in Eq. ∂t along with the linearized form of the boundary condition at the surface. (19) can be considered a problem of optimal control whereby the behavior of the governing flow equation system is controlled by the surface shape with deformations of the surface acting as the control input. T d ( pnS )dsdt (29) t0 S t0 Ω where we have introduced the adjoint variables.. where C∞ = ρ∞V∞2 A / 2 . . we take the first variation of Eq. Mathematically. with permission.2514/6. and A is a reference area. ρ∞ is the free stream density. G G G ∂U G ∂U G G G R (U ) = + ∇ ⋅ Fale = + ∇ ⋅ F − ∇ ⋅ (U ⊗ uΩ ) = 0. (31) can be expanded by including the linearized version of the governing equations with respect to small perturbations of the design surface.2013-2649 G Such that R(U ) = 0 Following the adjoint approach to optimal design.2 based on differential geometry formulas. 2013 | http://arc. (28) can be transformed into an unconstrained optimization problem by adding the inner product of an unsteady adjoint variable vector.aiaa. (27) ∂t ∂t where the terms involving the control volume motion have been separated from the traditional Euler fluxes. et. 11 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno. The third term of Eq. V∞ is the free stream velocity. Inc. Published by the American Institute of Aeronautics and Astronautics. any variations of the flow variables due to surface deformations are constrained to satisfy the system of governing flow equations. The minimization of Eqn. which operate as Lagrange multipliers and are defined as ⎧ψ ρ ⎫ ⎪ψ ⎪ ⎪ ρ u ⎪ ⎧ ψGρ ⎫ ⎪ ⎪ ⎪ ⎪ Ψ = ⎨ψ ρ v ⎬ = ⎨ ϕ ⎬. and the governing equations integrated over the domain (space and time) to form the Lagrangian: tf tf G G G J= ∫∫ ⋅ + ∫ ∫ Ψ R(U )d Ωdt . (30) ⎪ψ ⎪ ⎪ψ ⎪ ⎪ ρw ⎪ ⎩ ρE ⎭ ⎪⎩ψ ρ E ⎪⎭ To find the gradient information needed to minimize the objective function. the constrained optimization problem can be formulated as follows: tf G G Minimize J = ∫ ∫ d ⋅ ( pnS )dsdt t0 S (28) Downloaded by STANFORD UNIVERSITY on July 28. (31) are found by using a result from previous work by Palacios. ψ. (31) t0 S t0 S t0 Ω It is important to note that the first two terms of Eq. al. and this is a key feature differentiating the current formulation from other adjoint approaches. (29) with respect to perturbations of the surface shape: tf tf tf G G G G δ J = ∫ ∫ (d ⋅ ∇p )δ Sdsdt + ∫ ∫ (d ⋅ nS )δ pdsdt + ∫ ∫ Ψ T δ R (U )d Ωdt. G ∂ G G G G δ R (U ) = (δ U ) + ∇ ⋅ δ F − ∇ ⋅ δ (U ⊗ uΩ ) ∂t G G G ∂ G ⎛ ∂F G ⎞ ⎡ ∂ (U ⊗ uΩ ) G ⎤ = (δ U ) + ∇ ⋅ ⎜ G δ U ⎟ − ∇ ⋅ ⎢ G δU ⎥ ∂t ⎝ ∂U ⎠ ⎣ ∂U ⎦ (32) ( ) ∂ G G G G = (δ U ) + ∇ ⋅ A − IuΩ δ U .org | DOI: 10.

2013-2649 tf tf ∂ G G tf ∂Ψ T G ∫ ∫Ψ (δ U )d Ωdt = ∫ ⎡⎣ Ψ T δ U ⎤⎦ d Ω − ∫ ∫ δ Ud Ωdt T (35) t0 Ω ∂t Ω t 0 t0 Ω ∂t A zero-value initial condition for the adjoint variables can be imposed.. (37) Now. allowing for a zero final condition. We now perform manipulations to remove this dependence. (35). and (38) yields an intermediate expression for the variation of the cost function. t f ) = 0. the dynamic pressure (and hence the aerodynamic forces) will begin and end very close to zero. (31) to produce tf tf tf tf G G G ∂ G G G G δ J = ∫ ∫ ( d ⋅ ∇p )δ Sdsdt + ∫ ∫ (d ⋅ nS )δ pdsdt + ∫ ∫ Ψ T (δ U ) d Ωdt + ∫ ∫ Ψ T ∇ ⋅ ( A − IuΩ )δ Ud Ωdt (34) t0 S t0 S t0 Ω ∂t t0 Ω By removing any dependence on variations of the flow variables. t0 Ω With the appropriate choice of characteristic-based boundary conditions. the variation of the objective function for multiple surface deformations can be found without the need for multiple flow solutions which results in a computationally efficient method for aerodynamic design involving many design variables. Referring to the trajectory above. G G G G G G G G δ v ⋅ nS = −(v − uΩ ) ⋅ δ nS − ∂ n (v − uΩ ) ⋅ nS δ S (33) G G where A is the Jacobian of F using conservative variables. Published by the American Institute of Aeronautics and Astronautics. and assuming an unsteady flow. with permission. (36). Combining and rearranging the results from Eqs. t0 ) = 0. .2514/6. (34) yields tf tf tf G G G G G G G ⎡ Ψ T ( A − IuG )δ U ⎤d Ωdt − ∇Ψ T ⋅( A − IuG )δ Ud Ωdt . integrating the fourth term of Eq. (34). tf tf tf G G G G G G G δ J = ∫ ∫ ( d ⋅ ∇p )δ Sdsdt + ∫ ∫ (d ⋅ nS )δ pdsdt + ∫ ∫ Ψ T ( A − IuΩ ) ⋅ nS δ Udsdt t0 S t0 S t0 S tf (40) ⎡ ∂ΨT G G ⎤ G −∫ ∫ ⎢ + ∇Ψ T ⋅ ( A − IuΩ ) ⎥δ Ud Ωdt. (34) by parts gives Downloaded by STANFORD UNIVERSITY on July 28. (33) can now be introduced into Eq. Eq. (37). 2013 | http://arc. (36) G Ψ ( x . After changing the order of integration. ∫t Ω∫ Ψ ∇ ⋅ − δ Ω = ∫t Ω∫ ⎣ ∇ ⋅ ∫t Ω∫ T ( A Iu ) Ud dt (38) Ω Ω ⎦ Ω 0 0 0 and applying the divergence theorem to the first term on the right hand side of Eq. gives tf tf G G G G ⎡ ΨT ( A − IuGΩ )δ U ⎤ d Ωdt = ΨT ( A − IuGΩ ) ⋅ nGS δ Udsdt ∫t Ω∫ ⎣ ∇ ⋅ ⎦ ∫t ∫S 0 0 tf G G G G +∫ ∫Ψ T ( A − IuΩ ) ⋅ nSδ Udsdt t0 Γ ∞ (39) tf G G G G − ∫ ∫ ΨT ( A − IuΩ ) ⋅ nS δ Ud Ωdt. G Ψ ( x . δp. t0 Ω ⎣ ∂t ⎦ 12 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno. the first term on the right hand side of Eq. (38). (35) can be eliminated with the temporal conditions below (the cost function does not depend on tf).org | DOI: 10. the integral over the far-field boundary can be forced to vanish. assuming a smooth solution.aiaa. Inc. integrating the third term of Eq.

org | DOI: 10. (41) ∂t or. The final term of Eq. C.aiaa. if its integrand is zero at every point in the domain. the domain integral will vanish provided that the adjoint equations are satisfied as ∂ΨT G G + ∇ΨT ⋅ ( A − IuΩ ) = 0 in Ω. . et. after taking the transpose. a large enough number of surface nodes is used such that using the locations of surface nodes directly as independent variables in an optimization results in an optimization problem of impractically high dimensionality. the wall boundary condition.. and the variation of the objective function becomes tf tf G G G G ∂J δ J = ∫ ∫ ⎡⎣ (d ⋅ ∇p ) + (∇ ⋅ v )υ + (v − uΩ ) ⋅ ∇(υ ) ⎤⎦ δ Sdsdt = ∫ ∫ δ Sds. G G G (v − uΩ ) ⋅ nS = 0 . The surface sensitivity provides a measure of the variation of the objective function with respect to infinitesimal variations of the surface shape in the direction of the local surface normal. (40) can be evaluated given our knowledge of A . This value is computed at each surface node of the numerical grid at each physical time step with negligible computational cost. (45) t0 S t0 ∂ S S where ∂J G G G G = d ⋅ ∇p + (∇ ⋅ v )υ + (v − uΩ ) ⋅ ∇(υ ) (46) ∂S is what we call the surface sensitivity. ∂t (44) G G G G G G nS ⋅ ϕ = d ⋅ nS − ψ ρ E (v ⋅ nS ) on S. (43) to vanish can be written as ∂Ψ G G + ( A − IuΩ )T ⋅ ∇Ψ = 0 in Ω. the surface integral in the third term on G G Downloaded by STANFORD UNIVERSITY on July 28. Typically. the terms within the brackets constitute the set of partial differential equations which are commonly referred to as the adjoint equations. Note that the final expression for the variation involves only a surface integral at each physical time step and has no dependence on the volume mesh. By leveraging previous derivation of Economon. Published by the American Institute of Aeronautics and Astronautics. Therefore. (42) ∂t The accompanying boundary condition will be given below. it can be shown that evaluating the surface integral and rearranging the variation of the functional gives tf tf G G G G G G G G G G δ J = ∫ ∫ ⎡⎣(d ⋅ ∇p ) + (∇ ⋅ v )υ + (v − uΩ ) ⋅ ∇(υ ) ⎤⎦ δ Sdsdt + ∫ ∫ ⎡⎣ d ⋅ nS − nS ⋅ ϕ − ψ ρ E (v ⋅ nS ) ⎤⎦ δ pdsdt . (43) t0 S t0 S G G where υ = ρψ ρ + ρ v ⋅ ϕ + ρ Hψ ρ E . Therefore. Inc. al9 with some slight modifications and including time integration. Hence. (33). Palacios. and the linearized wall boundary condition in Eq.2514/6. uΩ . ∂Ψ G G + ( A − IuΩ ) ⋅ ∇Ψ = 0 in Ω. Application to Spherically-Parameterized Geometries The surface sensitivity discussed in the previous section provides the sensitivity of the chosen functional to locally-normal displacements of the surface of interest. (40) can also be made to vanish. the adjoint equations with the admissible adjoint boundary condition that eliminates the dependence on the fluid flow variations (δp) by forcing the second term on the right hand side of Eq. When set equal to zero. with permission. Furthermore.2013-2649 the right hand side of Eq. a parameterization of the geometry is used to 13 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno. 2013 | http://arc.

summed over the surface (N nodes. …. ⎝ r 2 + 1 ⎠ −1 ⎛ dr ⎞ r ≡ r ⎜ ⎟ . the unit normal vectors are given by Eqs. sin(θ ) + r cos(θ ) nx = − . m}. and ∂z/∂xi define the surface parameterization. The sensitivity of any geometric parameter to the functional can be expressed as the sum of the dot product of the local normal vector. The geometry is parameterized by m control points. The unit normal vector can be evaluated either from the surface mesh or analytically from the parameterization depending on the level of geometric complexity. Here. Differentiating Eqs. reduce the dimensionality via a mapping of the parameters of interest to the positions of the surface nodes. (51) ∂ri Δri 14 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno.aiaa.) SU2 provides the values of dJ/dS while the values of ∂x/∂xi. Inc. xi. a complex step method was used to perturb each control point and evaluate drj/dri: ∂rj Im[ rj ( ri + Δrii )] = . (50). Several such mappings are available in the SU2 design suite and in this work we add a spherical parameterization convenient to rocket shapes. In order to achieve the reduction in dimensionality an interpolation scheme is required to determine rj from the set of control points { ri }. (47) ∂xi j =1 ∂xi ⎝ ∂S ⎠ j j =1 ⎝ ∂S ⎠ j ⎝ ∂xi ∂xi ∂xi ⎠ (Note that x in Eq.2514/6. nˆ . the Cartesian coordinates are expressed as x = r cos(θ ) y = r sin(θ ) cos(φ ) . 2013 | http://arc. 2 with the axis positive forward through the nose. with permission. indexed by j): N ∂r ∂J ⎛ ∂J ⎞ N ⎛ ∂J ⎞ ⎛ ∂x j ∂y j ∂z j ⎞ =∑ ⋅ ⎜ nˆ ⎟ = ∑ ⎜ ⎟ ⎜ j nx + ny + nz ⎟ . Any appropriate numerical method can be used to find drj/dri depending on the operator and implementation. For a local spherical coordinate shown above in Fig.. For an axisymmetric geometry. (47) is used to represent both the Cartesian coordinate as well as a specific design parameter.org | DOI: 10. . ri {i = 1. r 2 + 1 ⎛ cos(θ ) − r sin(θ ) ⎞ ny = ⎜ ⎟ cos(φ ). provided an interpolation operator ∂rj/∂ri is available. (47) yields ∂J N ⎛ ∂J ⎞ ∂rj = ∑⎜ ⎟ ∂ri j =1 ⎝ ∂n ⎠ j ∂ri ( cos(θ )nx + cos(θ ) cos(φ )n y + cos(θ )sin(φ )nz ) j . with the change in position of the corresponding node due to a change in the parameter. ∂y/∂xi. ⎝ r 2 + 1 ⎠ (50) ⎛ cos(θ ) − r sin(θ ) ⎞ nz = ⎜ ⎟ sin(φ ). allowing for dJ/dri to be evaluated via drj/dri.2013-2649 spherical geometry. (48) and substituting into Eq. (48) z = r sin(θ )sin(φ ) where r is a function of θ and φ. (49) Equation (50) can be used to find the sensitivity of the functional J to changes in ri. Here the surface normal vectors are evaluated analytically from Downloaded by STANFORD UNIVERSITY on July 28. ⎝ dθ ⎠ This closes the geometric system. Published by the American Institute of Aeronautics and Astronautics. given the location and local unit normal vector of each surface node from the adjoint surface sensitivities dJ/dS.

5. VII. Hence. for a given x. ∂J/∂x. The SQP algorithm requires the value of the objective.000 cells. ∂c/∂x. 2013 | http://arc. Inc. however. and corresponding gradients. in addition to the nonlinear constraint function(s). VI..2514/6. It is noted. J. the overall procedure is unchanged if viscous flow is considered. that the mesh deformation method discussed above can be considerably less robust when boundary layer cells are involved. The baseline mesh. with permission. A hemispherical farfield was placed approximately 30 vehicle radii away nodes were bunched near the nose and near any corners. and corresponding gradients. is shown below in Fig. Downloaded by STANFORD UNIVERSITY on July 28. Published by the American Institute of Aeronautics and Astronautics. Inviscid CFD mesh of baseline fairing and portion of the launch vehicle showing detail near the nose of corners of the fairing profile.org | DOI: 10.16 can be used and a sequential quadratic programming (SQP) algorithm was chosen for use here to make efficient use of the gradient information provided by the adjoint sensitivities. Inviscid flow physics were chosen in order to reduce computational cost. The procedure described above allows for a straightforward implementation of well-established optimization algorithms. Optimization Procedure This section summarizes the optimization procedure used.aiaa. A 180° model of the fairing and a portion of the launch vehicle was created using quadrilateral surface elements and tetrahedral volume elements with a layer of pyramid elements between. totaling approximately 800. standard gradient- based optimization algorithms10. The specific formulation of 15 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno.13. c. The locations of the spline control points or locations and amplitude of the Hicks-Henne function are continuous variables with simple bounds and it is expected that the objective function(s) and nonlinear constraints (both of which are simple functions of the aerodynamic quantities CA and CM) will be smooth and continuous in the design space. Aerodynamic Model This section summarizes the computational fluid dynamics (CFD) model used to evaluate the aerodynamic properties of the launch via SU2. .2013-2649 Figure 5.

The baseline medium-lift launch vehicle parameters used for the sample results (and to generate the trajectory data shown above) are summarized in Table 1. Initial points were “corners” of the design space. x0. x0 = {6. This is discussed in more detail in the following section. Project the gradients from adjoint surface data to the design variables x. It is useful to summarize the complete process of each in the context of SU2: Objective Evaluation: 1. The fairing nose was parameterized with four spline control points. VIII.0 m}. 5. It is noted that. ∂c/∂x Of the items above. Perform an adjoint solution of functional CA on the deformed mesh (requires the flow solution data) 5. Perform a time-accurate flow solution with the deformed mesh over the trajectory 3. this optimization problem requires CA(x) and mf(x) for objective evaluation and CA(x). Published by the American Institute of Aeronautics and Astronautics.000 kg Reference Stress = 275 MPa (Al 6061-T6 yield) Reference Skin Thickness = 2. at each time step → ∂CA(t)/∂x 6. Fixed θ was chosen only to reduce the problem dimensionality and hence computational expense.0 m. Deform CFD mesh from baseline.org | DOI: 10. Baseline Vehicle This section describes the results of a specific case used to demonstrate the method in addition to discussing future directions of research.aiaa. due to the wide range of flow conditions and static CFD mesh (with respect to the flow features) very conservative solver settings were required throughout the entire time-accurate solution. Perform time-accurate flow solution on deformed mesh over trajectory 3. CM(t) Downloaded by STANFORD UNIVERSITY on July 28. the flow and adjoint solutions required the greatest computational expense with adjoint solutions somewhat more expensive than flow solutions.85 m Fairing Cylinder Radius = 2. ∂CM(t)/∂x 7. and mf(x) for constraint evaluation. ∂J/∂x Constraint Evaluation: 1. to current value. Perform adjoint solution of functional CA on deformed mesh (requires flow solution data) 5. Perform adjoint solution of functional CM on deformed mesh (requires flow solution data) 6. π/16. to current value. with permission.60 m 16 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno.2514/6. Process flow solution data (surface integration) at each time step → CA(t). The SQP algorithm will efficiently find a local minimum. x 2. Project gradients from adjoint surface data to design variables x at each time step → ∂CA(t)/∂x. 4. x 2.54 mm (0. Volume mesh deformation was a small but significant expense. or locations were at least two geometric constraints intersected. ∂CA(t)/∂x. Launch Vehicle Parameters First stage structural mass = 13. ∂CM(t)/∂x in time over the trajectory data → c. 3.000 kg Frist stage Isp = 300 s Second stage structural mass = 3. Integrate CA(t) and ∂CA(t)/∂x in time over the trajectory data → J. 2013 | http://arc. Results & Discussion A. Process the flow solution data (surface integration) at each time step → CA(t) 4.0 m.600 kg Second stage propellant mass = 45. x0. Recall that r(π/2) is fixed at Rcyl and both ends of the spline are constrained by ∂r/∂θ = 0.000 kg Propellant Mass at Fairing Deployment = 43. CM(x).2013-2649 4. at fixed θ = {0. Integrate CA(t). In order to improve the chances of capturing the global minimum several different values of x0 were used as initial points in addition to the baseline shown above. Inc.000 kg Second stage Isp = 350 s Payload Mass = 8. π/4}.. CM(t). .10”) Reference Vehicle Radius = 1. π/8. Deform CFD mesh from baseline.0 m. It is noted that this procedure is heuristic and does not guarantee global minimization but has been effective method to improve coverage of the design space in previous studies6.000 kg First stage propellant mass = 240.

466 km/s Fairing Mass = 2.0986 Table 2. respectively. 5. Fairing Surface Mass Density = 11 kg/m2 (general carbon / Al composite) Limit Cload = 0. CA = 0) = 9. 7-10.0 m. Summary of launch vehicle parameters required for fairing optimization.381 kg Cload = 0. The baseline vehicle and fairing yielded the performance summarized in Table 2. θ = {0.aiaa. Published by the American Institute of Aeronautics and Astronautics. Flow and adjoint contour plots corresponding to the baseline design are shown in Fig.org | DOI: 10.0 m. Downloaded by STANFORD UNIVERSITY on July 28. 3.2013-2649 Figure 6. 17 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno.0 m.596 km/s Table 1. Summary of launch vehicle parameters required for fairing optimization. Inc. 4. x0 = {6. . 2013 | http://arc. with permission. 6 and Figs.2514/6. π/8.0 m}. π/16. π/4} Total Vehicle ΔV = 9.0°..100 Ideal Vehicle ΔV (mf = 0. Mach contours for the baseline design under maximum q∞ conditions and α = 5.

2514/6.2013-2649 ure 7. 2013 | http://arc. ..0° for functtional CA.Downloaded by STANFORD UNIVERSITY on July 28. Ψρ contou Figu urs for the baseliine design underr maximum q∞ cconditions and α = 5. Figuure 8.aiaa. Inc. 18 Americcan Institute off Aeronautics aand Astronautiics Copyright © 2013 by Michael Colonno.0° for S contours for the funcctional CA. Shape seensitivity (∂J/∂S) t baseline dessign under maxiimum q∞ condittions and α = 55.org | DOI: 10. with permission. Published by the American Institute of Aeronautics and Astronautics.

Shape sensitivity (∂J/∂S) contours for the baseline design under maximum q∞ conditions and α = 5.aiaa. the optimization procedure is a balance between the conflicting effects of geometry on mass (surface area) and drag. The drag (more specifically CA) is shared by both the objective function and the structural 19 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno. 2013 | http://arc. Figure 10. Inc. Minimizing the drag alone results in a longer profile with higher L/R but also larger surface area.Downloaded by STANFORD UNIVERSITY on July 28.org | DOI: 10. . Optimum Design Qualitatively.. Ψρ contours for the baseline design under maximum q∞ conditions and α = 5.0° for functional CM. Published by the American Institute of Aeronautics and Astronautics. Minimizing the mass of the fairing will tend to shorten it and move the profile inward toward the axis to reduce the surface area.2013-2649 Figure 9. with permission.2514/6. B.0° for functional CM.

the combination of grid deformation CFD solutions (flow and adjoint) are currently the least robust. In addition. a more robust grid deformation algorithm would increase the range of x over which designs could be accurately evaluated.08655 Indented profile Downloaded by STANFORD UNIVERSITY on July 28.501 km/s Fairing Mass = 2. At high M∞ the more-pointed nose acts like a “sting” while the rest of the surface experiences flow behind an oblique shock. This reduces the sensitivity of the rest of the surface geometry and allows r(θ) to move toward the axis and save mass. the spring-analogy solver. Published by the American Institute of Aeronautics and Astronautics. it is noted that that value of the constraint. with permission. Cross-sectional view of an optimum design (blue) vs. This improvement is also expected in SU2 and will be utilized in future studies. flow or adjoint solutions which do not converge properly). modifying the flow and adjoint solution algorithm to allow for settings which are a function of the free stream conditions would accelerate convergence performance while safeguarding conditions at which conservative settings are required. In addition. C. especially near the nose at high M∞. π/4} Total Vehicle ΔV = 9. 2. 11. θ = {0. while generally very robust.2871 m. Improvements & Future Directions Of the many steps listed above required to evaluate a given design and its corresponding gradients. The optimal profile is shown below in Fig. an intermediate x* was re-meshed to reset the grid deformation range. Based on this intuition and the form of the surface sensitivity contours with respect to J we should expect an optimum shape that reduces r(θ) everywhere possible until the drag penalty incurred becomes too large. Second. 20 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno. may be able to tolerate x at which results are unavailable or of poor accuracy (at the likely expense of more function evaluations).org | DOI: 10. The extreme nature of the flight regime results in a high sensitivity of convergence to grid quality. π/8. constraint.307 kg Cload = 0. The length of the fairing (r at the nose) may increase or decrease depending on the trade between mass and drag. such as particle swarm optimization. Finally. the use a more tolerant optimization algorithm. at some points in the overall procedure above. Note the “pinched” shape of the fairing which reduces the fairing’s mass while maintaining comparable aerodynamic performance at higher M∞. First. gradient-based algorithms like SQP are generally not robust against objective or constraint functions that cannot be evaluated or yield inaccurate results (practically.. Optimum Design x* = {6. 3. . 2013 | http://arc. π/16. 4. This procedure can hence be improved in a number of ways.2514/6. actually decreased from its nominal value in the optimum design. the bending moment is emphasized. eventually creates cells which are more skewed than those of the original mesh as x moves away from x0.5). As of this writing. an elasticity- based grid deformation capability is nearing completion in the SU2 suite and will be utilized in future studies. (9). though in the formulation of the constraint. Cload.4071 m. Eq. the baseline design (black). Inc. Were viscous flow physics used.2013-2649 Payload Volume Vehicle Elongated nose Figure 11. Hence. This underscores the need to simulate over the entire atmospheric portion of the ascent trajectory as the optimum profile has superior performance for much of the flight after maximum q∞ (M∞ > 1.aiaa.7479 m}. this issue would be exacerbated. It is worth noting that the profile above a ctually causes a performance reduction (larger CA but lower mass) when analyzed only at the maximum q∞ condition.5578 m.

631-646. F.. 41 No. Ph. AIAA Journal. No.. M.. September 2002. 16 Jones. J. R. and Vanderplaats. R. including the fairing’s effect on the trajectory near maximum q∞ and the sensitivity of the launch vehicle’s structural mass to the loads applied by the fairing. Colonno. “Continuous Adjoint Approach for the Spalart-Allmaras Model in Aerodynamic Optimization". and Welch. 20 Palacios. M. H. 21 American Institute of Aeronautics and Astronautics Copyright © 2013 by Michael Colonno. 80. Department of Aeronautics and Astronautics. 17 Knight. 1988. New Orleans.stanford. doi: 10. D. 2013. Yang. 22.. September. Vol. accepted for publication. “Stanford University Unstructured (SU2): An Open Source Integrated Computational Environment for Multiphysics Simulation and Design. January 2003. D. J. CA. pp.. S. K. J. are likely comparable in effect to those studied here.. Vol. 2004. M. 15 Jameson. A. R. “Efficient Global Optimization of Expensive Black-Box Functions”. "Aero-structural Wing Planform Optimization. N. Vol. J. Lepsch. D. M. "A Globally Convergent Method for Nonlinear Programming. Vol. M. . J. Journal of Aircraft. Palacios.2514/6. N. J. A. R. Murray.D Dissertation. January 1998. M. NV. F. 47. 13 Han. Liu. C.. 297. A. Inc. and Newman. W.. G. December 1998.. and Obayashi. 1977..org | DOI: 10. “Delta II payload Planners Guide”.. A. 2012.. 13. L. 19 Martins. 9th AIAA / ISSMO Multidisciplinary Analysis and Optimization Conference. Economon. 1. Powell.. Alonso. “Collaborative Approach to Launch Vehicle Design”. No. R. Alonso. 10 Gill. A Multidisciplinary Optimization Environment for Rocket Vehicle Systems. 18 Leoviriyakit. J.. “Optimal Shape Design for Open Rotor Blades”. 9 Economon T. O... A. M.. E.A32301. and Saunders. Jameson.. (Ed. C. Ph. NASA. Journal of Spacecraft and Rockets. 3. 2006. 6 Colonno. W.). A.J. D. 36(1):51-74. 2008. H. C.2013-2649 5 Brown." SIAM Review. 50. M. References 1 Blades. “Study on the Use of High-Fidelity Methods in Aeroelastic Optimization”. “Evaluation of Multidisciplinary Optimization (MDO) Techniques Applied to a Reusable Launch Vehicle”... 12 Guruswamy. I. and Olds. P.. 4. 1 August 1995.. J. 2002. March 2012. Murman. 22 Stanford University Unstructured (SU2): http://su2. R. “Aerodynamic Design via Control Theory". Duraisamy. 25 Wu.-Y... E.. “Constrained Multipoint Aerodynamic Shape Optimization Using an Adjoint Formulation and Parallel Computers: Part I & II". Journal of Scientific Computing. Downloaded by STANFORD UNIVERSITY on July 28.. 30th AIAA Applied Aerodynamics Conference. J. R.. 43rd AIAA Aerospace Sciences Meeting and Exhibit. Alonso. and Saunders. July-August 1997. D.. AIAA 2003-4095.... July 1974.. and Zuazua. Published by the American Institute of Aeronautics and Astronautics. A. S. December. P. Stanley.. 14 Hicks..” 51st AIAA Aerospace Sciences Meeting. 2004. D. the optimization problem as formulated here is merely a subset of a larger MDO problem as discussed above.” Computers & Structures.. January 2005... pp. F. E.. R. and Farhat. Jan. Moore.. 21 Reuther. T. J. Journal of Aircraft. R. Journal of Optimization Theory and Applications. and Kroo. M. 2013. LA. Vol.. 4 Braun.. M.. and Tsai.. Stanford. R." 42nd Aerospace Science Meetings & Exhibit.. Hicken. A. G. p." AIAA Journal of Spacecraft and Rockets.. Campos. Victoria. 2002. 34 No. Lukaczyk. S. Journal of Global Optimization. Some of the neglected dependencies. G... P.. Lonkar. TX.. 11 Guruswamy.. T. 455-492.. 3. 2013 | http://arc.. D. AFPSPC 002-93-I. Castro. pp. S. K. NASA TM X-3092. "Computational-Fluid-Dynamics-Based Design Optimization of a Large Asymmetric Payload Fairing. J. “The Optimum Launch Vehicle Fairing: an MDO Approach”. 305-316. “Comparison of Two Multidisciplinary Optimization Strategies for Launch-Vehicle Design”. and Kroo. R. A. 2007. Oct.. Grapevine. “Use of High-Fidelity Methods in Multidisciplinary Optimization – A Preliminary Survey”. thesis. Stanford University. “A Three-dimensional Torsional Spring Analogy Method for Unstructured Dynamic Meshes. 2005. 1999. F. pp. 23 United Launch Alliance (ULA). 7 Colonno. J. “Evolved Expendable Launch Vehicle Standard Interface Specification. Vol. No. 3. Stanford University. Schonlau. 12th AIAA Multidisciplinary Analysis and Optimization Conference.1.. Reno. and Obayashi.. I. 3:233-260. L."SNOPT: An SQP Algorithm for Large-Scale Constrained Optimization. Journal of Spacecraft and Rockets." Vol. J. T. and Jameson. June. F. May-June 1995. Taylor. Copeland. Jan.-M. with permission.. P.2514/1. 8 Degand. Palacios. W. Bringing the trajectory calculation and a structural analysis – even a relatively simple or low-fidelity one – into the optimization loop would improve the completeness of the results without adding substantially to the overall computational cost.D. 99-131.” Version 4. 2 Bueno-Orovio.. and Alonso. J. Aranake. Rimlinger. F. A. S. Finally. A. C. “Operational Requirements Document for the Evolved Expendable Launch Vehicle”. Vol.. "A Coupled-Adjoint Method for High-Fidelity Aero-Structural Optimization". 4. “An Assessment of Airfoil Design by Numerical Optimization”. 3 Braun.aiaa. III. E.edu. 24 US Air Force Space Command. A. 32 No.. Comparison of Three Geometric Representations of Airfoils for Aerodynamic Optimization. BC.