Non-intrusive aerodynamic shape optimisation with a discrete

empirical interpolation method

Simão Marques∗ and Lucas Kob†

Department of Mechanical Engineering Sciences, The University of Surrey, Guildford, England, GU2 7XH U.K.

Trevor Robinson‡
School of Mechanical and Aerospace Engineering, Queen’s University Belfast, Belfast, Northern Ireland, BT9 5AH, U.K.

Weigang Yao§
Faculty of computing, Engineering and Media, De Montfort University, Leicester, LE1,9BH, U.K.

Liang Sunn¶
School of Computing, Engineering and Intelligent Systems, Ulster University, Northern Ireland, BT48 7JL, U.K.

This work presents a strategy to build reduced order models suitable for aerodynamic shape
optimization, resulting in a multi-fidelity optimisation framework. A reduced-order model based
on a Discrete Empirical Interpolation (DEIM) method is proposed to replace computational
fluid dynamics solvers, for fast, nonlinear, aerodynamic modeling. The DEIM builds a set
of interpolation points to reconstruct the flow fields using a set of basis obtained by Proper
Orthogonal Decomposition of a matrix of snapshots. The aerodynamic reduced order model is
completed by introducing a nonlinear mapping function between surface deformation and the
DEIM interpolation points. The resulting multi-fidelity optimisation problem is managed by a
trust-region algorithm. The design space is initially restricted; as the optimisation trajectory
evolves, new samples enrich the ROM. The proposed methodology is evaluated using a series
of transonic and subsonic test cases, including a rectuangular wing and the Onera M6 wing.
Preliminary results demonstrate a factor of 2-4 reduction in full-order model evaluation with
respect to adjoint methods for problems with a moderate number of design variables.

I. Nomenclature

A = Snapshot matrix
𝑎 = cylinder diameter
𝐶𝐷 = lift coefficient
𝐶𝐿 = drag coefficient
𝐶𝑝 = pressure coefficient
c = DEIM coefficients
CE = equality constraint
CI = inequality constraint
F = objective function
𝑙𝑏 = design parameter lower bound values
𝑀∞ = free-stream Mach number
R = vector of fluid equations residuals
℘ = DEIM interpolation indices matrix
𝑠 = optimisation step
S = eigenvalues of snapshot matrix
∗ Senior,Lecturer, [email protected], MAIAA.
† Post-Doctoral Fellow, [email protected]
‡ Reader, [email protected], MAIAA
§ Senior Lecturer, [email protected]
¶ Lecturer, [email protected]

1
T𝑀 = reduced basis matrix
T = left singular vectors
𝑢𝑏 = design parameter upper bound values
V = right singular vectors
w = conserved flow variables
U = velocity vector
u = vector containing surface preesures
Greek Symbols =
𝛼 = step length for Newtown method, angle of attack
Δ = trust region radius
𝜇 = design parameters

II. Introduction
he pressure to reduce design cycles or further explore design spaces keeps demanding more efficient analysis tools.
T Typical engineering problems, such as aerodynamic shape optimisation, require sophisticated analysis tools to solve
parametric Partial Differential Equations (PDEs). This is a challenging process due to the high computational cost
associated with having to interrogate a large and complex model multiple times. The advent of adjoint methods to
compute gradients at a cost independent of the number of design variable opened up the perspective of affordable design
whilst retaining a high-fidelity model [? ? ? ? ? ? ]. The solution of the the CFD adjoint equations, typically requires
the same effort as what is required to solve the original flow equations. Therefore, the use of high-fidelity models in
design optimisation is feasible, but not necessarily trivial or affordable.
Several strategies are available to accelerate the time to solution required for a given analysis, in particular for
problems such as optimisation or uncertainty quantification where it is necessary to interrogate a large and complex
model multiple times. One common approach is to build a surrogate models and devise a multi-fidelity strategy to
compute the quantity of interest efficiently. Surrogates are typically based on interpolation methods like kriging or
neural networks it is also possible to resort to lower fidelity models that remain representative of some relevant physical
phenomenon, to act as an effective surrogate The obvious penalty of such strategies is on how to guarantee the method
converges to a local optimum common to the high-fidelity model; in addition users face challenges in associating models
and respective parameterizations that use very different representations of the same problem.
A slightly different alternative is to build a reduced-order model (ROM) by examining the governing equations
and performing some type of model reduction, suitable for the physics of the problem. Here, the common approach is
to assume the parametric behaviour of the high-fidelity model can be approximated by a small number of modes or
basis, typically obtained by methods such as Proper Orthogonal Decomposition (POD) [? ], Balance POD relationship
between residual norm of the ROM and its error and proposed a [? ? ], Proper Generalized Decomposition [? ] -
extensive reviews on the wider subject of model reduction can be found in [? ? ? ? ? ? ]. The exploitation of ROMs
for aerodynamic shape optimisation was investigated by LeGresley and Alonso [? ] for inverse aerofoil design, where
a POD ROM was used to approximate the gradient by finite-differencing. Investigators have also been successful in
deriving efficient ROMs for shape optimisation problems described by elliptical PDEs such as those found in marine
or biological applications [? ? ? ]. The problem of aerofoil shape optimisation subject to compressible flows was
revisited by Zahr and Farhat, the authors assumed a monotonic nonlinear trust-region optimisation method that updates
the reduced basis as the optimisation progresses, when the residual fails to reach a required threshold [? ]. The authors
exploit a Least-Squares Petrov-Galerkin projection [? ] to reduce the state equations and respective sensitivities, hence,
each snapshot involved concatenating samples of the fluid state variables and sensitivities with respect to the design
variables. Yao et al. employed a similar strategy, but for gradient based optimisation for aerofoils in compressible and
transonic flows, reducing the time to solution by 30 − 50% with respect to an adjoint based optimisation [? ].
One particular challenge with the aforementioned ROM strategies, is that they are intrusive, i.e. they require the
manipulation of original CFD solver, which may not be available nor desirable. In reference [? ], the authors investigate
the ability of DEIM [? ] of reconstructing unsteady flowfields for dynamic aeroelastic instabilities prediction. Inspired
by these findings, this paper proposes a strategy suitable for aerodynamic shape optimisation that relies on a nonlinear
mapping between the surface DEIM points and flow quantities of interest, allowing the DEIM ROM to reconstruct
surface flow fields for new shapes. A trust-region model management ROM is able to drive the aerodynamic shape to
new local minima, at a reduced cost with respect to adjoint based optimisation, while treating the high-fidelity model as
a black box.

2
 The remainder of the paper will provide details of the method implemented. This is followed by two initial test
cases. The abstract finishes with a conclusion and plans for the complete paper.

III. Non Intrusive Reduce-Order Modelling

A constrained aerodynamic shape optimisation can be represented by
minimize F (w(𝜇), 𝜇)
𝜇∈𝒟

subject to R(w(𝜇), 𝜇) = 0, (1)

CE (w(𝜇), 𝜇) = 0,
CI (w(𝜇), 𝜇) ≤ 0.
In Eq.1 the objective function F depends on the flow variables w(𝜌, 𝜌U, 𝜌𝐸), and parameters 𝜇. For the purpose of
optimisation problems in this work, the flow quantities of interest can be restricted to the pressure field u ∈ R 𝑁 , where
𝑁 is the number of mesh points. The DEIM aims to approximate u(𝜇) as a nonlinear function of 𝜇, by projecting it on
the subspace spanned by the basis T 𝑀 :

u(𝜇) ≈ T 𝑀 c(𝜇), (2)

where c(𝜇) represents a set of coefficients and the matrix T 𝑀 is obtained by POD of flow snapshots. The task now is
how to select the 𝑀 distinct rows or interpolation indices that allow obtaining a unique solution of c(𝜇). This can be
achieved by following the algorithm in [? ]; the interpolation indices are coded into matrix ℘ which results in the DEIM
approximation:
  −1
u(𝜇) ≈ T 𝑀 ℘𝑇 T 𝑀 ℘𝑇 û(𝜇). (3)
For cases such as those presented here, the objective function is obtained through integrating the fluid forces over
surfaces, hence the size of the ROM can be further reduced by letting û(𝜇) represent only the values of the surface
pressure corresponding to the non-zeros indices in ℘, i.e. the 𝑀 elements required to recover the full order vector u(𝜇).

A. Proper Orthogonal Decomposition

The basis T 𝑀 can be built by taking the POD modes of the flow field. This requires obtaining a set of snapshots and
assemble the matrix A, where each column represents the fluid solution vector for one particular set of parameter values
𝜇:
 
A = u1 , u2 , . . . , u𝑛𝑠 , (4)
where 𝑛𝑠 is the number of samples. Since 𝑛𝑠 is much smaller than the dimension of u, the POD modes can be
obtained by performing singular value decomposition (SVD) to A𝑇 A, where A𝑇 A ∈ R𝑛𝑠 ×𝑛𝑠 :

A𝑇 A = TSV𝑇 (5)
where S is a diagonal matrix containing the eigenvalues and T, V are the left and right singular vectors, respectively.
The POD basis, T 𝑀 ∈ R𝑛𝑠 × 𝑀 , are obtained with proper truncation according to the energy rank given by the amplitude
of the first 𝑀 eigenvalues:

T 𝑀 = AVS−1/2
𝑀 (6)

B. Interpolation points estimation

To employ the DEIM in lieu of a high-fidelity model, it is necessary to obtain û(𝜇) for any shape produced
by the optimiser. Hence, in this work, a nonlinear mapping between the design parameters 𝜇 and the flow values
at the interpolation indices needs to be built. Note that 𝜇 and û are vectors of distinct sizes, which results in a
multi-input/multi-output problem. Any suitable interpolation method can be used here; following from reference [? ],
an exact radial basis function (RBF) neural network using the MatLab function newrbe is built using the respective
input/output sample pairs (𝜇𝑖 , û𝑖 ), 𝑖 = 1, ..., 𝑛𝑠 . This interpolation method ensures the function values interpolated are
exact at the sample points.

3
 IV. Trust-Region Model Management
Obtaining a set of global basis is usually not achievable or would require and excessive number of snapshots,
rendering the ROM approach unattractive. An alternative is to augment the snapshots matrix as the optimisation
progresses as demonstrated in [? ? ]. Trust-region methods are provable convergent and able to manage the exchanges
between different fidelity models. At each major iteration 𝑘, an optimisation subproblem is defined on the trust-region
centred at 𝜇 𝑘 and radius Δ 𝑘 :
minimize F̂ (u(𝜇 𝑘 ), 𝜇)
𝑠∈𝔅 𝑘

subject to ĈE (u(𝜇 𝑘 ), 𝜇 𝑘 + 𝑠) = 0,

ĈI (u(𝜇 𝑘 ), 𝜇 𝑘 + 𝑠) ≤ 0, (7)
𝑙𝑏 ≤ (𝜇 𝑘 + 𝑠) ≤ 𝑢𝑏,
||𝑠|| ∞ ≤ Δ 𝑘 ,
where the “ ˆ ” symbol indicates quantities computed using the Eq.3; 𝑠 is the optimisation step size and 𝔅 𝑘 = {𝜇 ∈
R𝑛 𝑝 : ||𝜇 − 𝜇 𝑘 || ≤ Δ 𝑘 }. The optimisation subproblem is solved using MatLab’s fmincon with the gradients obtained by
finite differences using the DEIM ROM.

V. Results
Preliminary results of the performance of the proposed method from two test cases are reported here: the first
problem is adapted from the AIAA Aerodynamic Design Optimization Discussion Group case 3∗ , a constrained drag
minimization problem for a rectangular wing using twist as a design variable at 𝑀∞ = 0.5 in inviscid flow; the second
problem employes the widely used ONERA M6 wing at transonic conditions benchmark, also in inviscid flow. Both
geometries are parameterised using a lattice of control points and a mesh morphing method based on radial basis
functions (RBF)[? ]; the RBF mesh morphing method allows manipulating not only the wing section shapes, but also
twist, sweep, span, dihedral and is used during the optimisation subproblem iterations. The open-source CFD solver
𝑆𝑈 2 [? ] is used to obtain all snapshots and high-fidelity results.

A. Twist Optimization
The first test case requires the optimisation and a NACA 0012 rectangular wing using twist as sole variable. The
problem is expressed as:

minimize 𝐶𝐷
𝜇∈𝒟
(8)
subject to 𝐶 𝐿 > 0.375
where 𝜇 defines the twist angle at the perpendicular planes to the span shown in Fig. 1. The RBF mesh morphing
requires the set up of control points, that are responsible for deforming the surface mesh, each intersection shown in
Fig. 1, represents one control point. However, to control the wing twist, a virtual axis is introduced, that rotates all
control points on each perpendicular plane to the spanwise direction, hence Fig. 1 represents a problem with nine design
variables, since the root is fixed.
The ROM based optimisation used three samples to build the initial ROM, these included a reference solution, i.e.
the original wing, plus two samples obtained by randomizing the nine design variables within ±0.5◦ . A total of 13
high-fidelity function evaluations were necessary to converge the problem. The convergence history is shown in Fig. 2,
which includes the initial three samples. The final optimum shape was computed with 𝑆𝑈 2 , the comparison between the
initial and final shapes flowfields is illustrated in Fig. 3, showing the impact of the new twist distribution on the Mach n.,
and a reduction on the peak towards the wing tip for the optimised shape.
An equivalent problem using the Free-Form Deformation box parameterization and adjoint solver available within
𝑆𝑈 2 was also computed, the comparison between the methods is shown in Fig. 4. Both optimisation show consistent
results, converging to slightly different optima. The adjoint solution required 37 function evaluations. Considering
that the ROM evaluations consume a total of 10% of the total computation effort for the ROM based optimisation, this
shows a speed-up factor of 3-4 with respect to the adjoint strategy.
∗ https://drive.google.com/file/d/18I_yV7hmStazeLI5KbZbbXe7MDdPQjuF/view

4
 Fig. 1 RBF Mesh morphing lattice.

Fig. 2 Optimisation convergence history.

B. Transonic Optimization
The second test case involves the ONERA M6 wing at 𝑀∞ = 0.83 and angle of attack 𝛼 = 3.08◦ . The constrained
optimisation problem is defined in Eq. 9 and is also included in the test cases suite from 𝑆𝑈 2 .

minimize 𝐶𝐷
𝜇∈𝒟
(9)
subject to 𝐶 𝐿 > 0.28

The parameterisation used, controls the local surface shape via 24 control points, i.e. design variables, using the
aforementioned RBF mesh morphing algorithm. A second paramererization with 36 control points is also investigated.
A representative lattice is shown in Fig. 5. The optimisation problem was initialized with four random samples plus the
nominal wing, and terminated after 15 iterations. The comparison between the initial and optimised flowfield, Fig. 6,
shows how the optimiser is able to eradicate the shock, leading to a significant reduction in drag. The optimisation
convergence history is shown in Fig. 7 and compared with a 𝑆𝑈 2 adjoint based optimisation using 160 FFD control
points. Fig. 7 it includes the initial five samples used to set up the ROM, then each iteration is equivalent to one FOM
call, where as the each 𝑆𝑈 2 may include calls to the adjoint solver. The total cost and performance of both strategies are
summarized in table 1, showing that the ROM based optimisation gives a 3% higher 𝐶𝐷 value, with a 30% reduction in
computational cost with respect to 𝑆𝑈 2 .

5
 Fig. 3 Comparison between initial (left) and optimised (right) flows.

(a) (b)

(c) (d)

Fig. 4 Comparison between ROM based and adjoint based optimisation optima – surface pressure coefficient

6
 Fig. 5 Comparison between initial (left) and optimised (right) flows.

Fig. 6 Comparison between initial (left) and optimised (right) upper surface pressure coefficient.

VI. Conclusions and Results for the final paper

An optimisation strategy suitable for aerodynamics shape optimisation using a non-intrusive DEIM ROM has been
demonstrated. Furthermore, results indicate the ROM’s potential to speed-up finding optimum solutions with respect to
current state-of-the-art approaches used for shape optimisation.
The final paper will give full details of the methodology and algorithms implemented, together with additional
results detailing how to set up the ROM based problems, the ROM’s capabilities will be exercise its performance
assessed relative to the current state-of-the-art methods.

Acknowledgments
The funding provided for this research from the Engineering and Physical Sciences Research Council (Grant No.
EP/P025692/1) is gratefully acknowledged.

𝑆𝑈 2 Adjoint Trust-Region ROM

Initial 𝐶𝐷 0.0127 0.0127
Final 𝐶𝐷 0.0095 0.0098
FOM Calls 33 20
Table 1 Performance comparison between adjoint based and ROM based optimisations

7
 Fig. 7 Optimisation convergence history.

References
[] Pironneau, O., Optimal Shape Design for Elliptic Systems, Springer series in computational physics, Springer-Verlag, 1984.
URL https://books.google.co.uk/books?id=6JqyAAAAIAAJ.

[] Jameson, A., “Aerodynamic design via control theory,” Journal of scientific computing, Vol. 3, No. 3, 1988, pp. 233–260.

[] Jameson, A., Martinelli, L., and Pierce, N., “Optimum aerodynamic design using the Navier–Stokes equations,” Theoretical
and computational fluid dynamics, Vol. 10, No. 1-4, 1998, pp. 213–237.

[] Elliott, J., and Peraire, J., “Practical three-dimensional aerodynamic design and optimization using unstructured meshes,” AIAA
journal, Vol. 35, No. 9, 1997, pp. 1479–1485.

[] Nielsen, E. J., and Anderson, W. K., “Aerodynamic design optimization on unstructured meshes using the Navier-Stokes
equations,” AIAA journal, Vol. 37, No. 11, 1999, pp. 1411–1419.

[] Giles, M. B., and Pierce, N. A., “An Introduction to the Adjoint Approach to Design,” Flow, Turbulence and Combustion,
Vol. 65, No. 3, 2000, pp. 393–415.

[] Sirovich, L., “Turbulence and the dynamics of coherent structures. I: Coherent Structures. II: Symmetries and Transformations.
III: Dynamics and Scaling,” Quarterly of Applied Mathematics, Vol. 45, No. 3, 1987, pp. 583–590.

[] Willcox, K., and Peraire, J., “Balanced model reduction via the proper orthogonal decomposition,” AIAA journal, Vol. 40,
No. 11, 2002, pp. 2323–2330.

[] Rowley, C. W., “Model Reduction for Fluids, using Balanced Proper Orthogonal Decomposition,” International Journal of
Bifurcation and Chaos, Vol. 15, No. 3, 2005, pp. 997–1013.

[] Ammar, A., Mokdad, B., Chinesta, F., and Keunings, R., “A new family of solvers for some classes of multidimensional partial
differential equations encountered in kinetic theory modeling of complex fluids,” Journal of Non-Newtonian Fluid Mechanics,
Vol. 139, No. 3, 2006, pp. 153 – 176.

[] Lucia, D. J., Beran, P. S., and Silva, W. A., “Reduced-order modeling: new approaches for computational physics,” Progress in
Aerospace Sciences, Vol. 40, No. 1, 2004, pp. 51 – 117. https://doi.org/10.1016/j.paerosci.2003.12.001.

8
[] Noack, B. R., Morzynski, M., and Tadmor, G., Reduced-order modelling for flow control, Vol. 528, Springer Science & Business
Media, 2011.

[] Lassila, T., Manzoni, A., Quarteroni, A., and Rozza, G., “Model order reduction in fluid dynamics: challenges and perspectives,”
Reduced Order Methods for modeling and computational reduction, Springer, 2014, pp. 235–273.

[] Benner, P., Gugercin, S., and Willcox, K., “A Survey of Projection-Based Model Reduction Methods for Parametric Dynamical
Systems,” SIAM Review, Vol. 57, No. 4, 2015, pp. 483–531.

[] Amsallem, D., Farhat, C., and Haasdonk, B., “Special Issue on Model Reduction,” International Journal for Numerical Methods
in Engineering, Vol. 102, No. 5, 2015, pp. 931–932. https://doi.org/10.1002/nme.4889.

[] Yondo, R., Andrés, E., and Valero, E., “A review on design of experiments and surrogate models in aircraft real-time and
many-query aerodynamic analyses,” Progress in Aerospace Sciences, Vol. 96, 2018, pp. 23–61.

[] LeGresley, P., and Alonso, J., “Airfoil design optimization using reduced order models based on proper orthogonal decomposition,”
Fluids 2000 conference and exhibit, 2000, p. 2545.

[] Demo, N., Tezzele, M., Mola, A., and Rozza, G., “A complete data-driven framework for the efficient solution of parametric
shape design and optimisation in naval engineering problems,” arXiv preprint arXiv:1905.05982, 2019.

[] Manzoni, A., Quarteroni, A., and Rozza, G., “Shape optimization for viscous flows by reduced basis methods and free−form
deformation,” International Journal for Numerical Methods in Fluids, Vol. 70, No. 5, 2012, pp. 646–670.

[] Quarteroni, A., and Rozza, G., “Optimal control and shape optimization of aorto-coronaric bypass anastomoses,” Mathematical
Models and Methods in Applied Sciences, Vol. 13, No. 12, 2003, pp. 1801–1823.

[] Zahr, M., and Charbel, F., “Progressive construction of a parametric reduced−order model for PDE−constrained optimization,”
International Journal for Numerical Methods in Engineering, Vol. 102, No. 5, 2015, pp. 1111–1135.

[] Kevin, C., Charbel, B., and Charbel, F., “Efficient non-linear model reduction via a least-squares Petrov-Galerkin projection and
compressive tensor approximations,” International Journal for Numerical Methods in Engineering, Vol. 86, No. 2, 2011, pp.
155–181.

[] Yao, W., Marques, S., Robinson, T. T., and Sun, L., “A Reduced-order Model for Aerodynamic Shape Optimization,” AIAA
Scitech 2019 Forum, 2019, p. 0975.

[] Yao, W., and Marques, S., “Nonlinear aerodynamic and aeroelastic model reduction using a discrete empirical interpolation
method,” AIAA Journal, Vol. 55, No. 2, 2017, pp. 624–637.

[] Chaturantabut, S., and Sorensen, D. C., “Nonlinear model reduction via discrete empirical interpolation,” SIAM Journal on
Scientific Computing, Vol. 32, No. 5, 2010, pp. 2737–2764.

[] Lapuh, R., “Mesh Morphing Technique used with Open-Source CFD Toolbox in Multidisciplinary Design Optimisation,” ,
2018.

[] Economon, T. D., Palacios, F., Copeland, S. R., Lukaczyk, T. W., and Alonso, J. J., “SU2: An Open-Source Suite for
Multiphysics Simulation and Design,” AIAA Journal, Vol. 54, No. 3, 2016, pp. 828–846. https://doi.org/10.2514/1.J053813.